[{"data":1,"prerenderedAt":2015},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p":3,"blog-posts-sidebar-en":1451},{"id":4,"title":5,"body":6,"categories":1432,"date":1437,"description":1438,"extension":1439,"heading":5,"image":1440,"meta":1441,"navigation":1442,"originalUrl":1443,"path":1444,"seo":1445,"slug":1446,"socialImage":1447,"stem":1448,"updated":1449,"video":1449,"__hash__":1450},"blog\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p.md","Deepseek V4.1 Flash on a single MI350P",{"type":7,"value":8,"toc":1411},"minimark",[9,13,22,38,43,128,139,150,156,161,164,167,170,181,184,188,191,194,198,204,268,271,275,278,281,286,342,360,364,370,380,448,454,459,475,487,491,494,585,591,594,598,601,647,650,653,657,664,667,672,801,804,807,811,826,829,834,958,961,965,979,982,987,1109,1112,1116,1135,1150,1157,1161,1164,1167,1172,1243,1246,1249,1264,1275,1279,1290,1299,1303,1324,1336,1343,1353,1369,1382,1385,1393,1398,1402,1405],[10,11,12],"p",{},"We wanted to see how much we could squeeze out of one GPU. More projects are finding ways to run huge models by keeping some weights in system RAM or on disk, instead of needing enough GPU memory for everything. We wanted to explore that on CDNA4.",[10,14,15,16,21],{},"We had already done something similar with ",[17,18,20],"a",{"href":19},"\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700","Qwen3.8 Flash Next on two Radeon AI PRO R9700s",", keeping less-used experts in system RAM. This time, we picked a completely different, much heavier model: DeepSeek V4.1 Flash. The hardware limit was one MI350P, backed by a substantial CPU host, not a rack of GPUs.",[10,23,24,25,29,30,33,34,37],{},"And we're pretty happy with the results. ",[26,27,28],"strong",{},"DeepSeek V4.1 Flash generates 77.9 tokens per second for one request on that single MI350P with Paiton."," At 64 concurrent requests, LocalMaxxing records ",[26,31,32],{},"709.6 tokens\u002Fs in total",". A separate batch-serving profile reaches ",[26,35,36],{},"1,107.6 output tokens\u002Fs across 512 requests with long answers",".",[10,39,40],{},[26,41,42],{},"In this article",[44,45,46,53,59,65,71,82,98,104,110,116,122],"ul",{},[47,48,49],"li",{},[17,50,52],{"href":51},"#keeping-useful-work-on-the-gpu","Hardware and quantization",[47,54,55],{},[17,56,58],{"href":57},"#two-profiles-two-different-uses","Serving profiles and cache capacity",[47,60,61],{},[17,62,64],{"href":63},"#decode-speed-without-the-prompt-processing-work","Decode speed and prefill",[47,66,67],{},[17,68,70],{"href":69},"#how-this-compares-with-an-rtx-pro-6000","RTX PRO 6000 comparison",[47,72,73,77,78],{},[17,74,76],{"href":75},"#more-single-machine-results","RTX 5090, RTX 3090 and Mac Studio",", plus ",[17,79,81],{"href":80},"#what-changes-with-more-gpus","multi-GPU setups",[47,83,84,88,89,93,94],{},[17,85,87],{"href":86},"#what-happens-when-requests-arrive-together","Concurrent requests",", ",[17,90,92],{"href":91},"#long-inputs-are-the-harder-serving-workload","long inputs"," and ",[17,95,97],{"href":96},"#long-answers-improve-total-throughput","long answers",[47,99,100],{},[17,101,103],{"href":102},"#power-the-limit-is-not-the-consumption","Power consumption",[47,105,106],{},[17,107,109],{"href":108},"#how-close-are-the-predictions-to-the-original","Accuracy and retrieval checks",[47,111,112],{},[17,113,115],{"href":114},"#what-happened-with-stock-engines","Stock vLLM and SGLang attempts",[47,117,118],{},[17,119,121],{"href":120},"#inspect-the-measurements","Weights, source data and recording",[47,123,124],{},[17,125,127],{"href":126},"#building-a-token-factory-on-amd","Discuss your token factory",[10,129,130,131,134,135,138],{},"That last number is for batch processing, not responsive chat. Each request generates about 2.3 tokens\u002Fs after a median wait of three minutes for its first token. All MI350P measurements use a ",[26,132,133],{},"450 W GPU power limit"," and a host that holds ",[26,136,137],{},"about 290 GB of model data",". Its CPUs compute the experts that do not fit on the GPU. “One GPU” does not mean the GPU does all the work.",[10,140,141,142,145,146,149],{},"The model is DeepSeek-V4.1-Flash-1017. Its original checkpoint occupies ",[26,143,144],{},"475 GiB",", far beyond this card's ",[26,147,148],{},"144 GB of HBM",". The result comes from changing how the model is stored and executed, rather than fitting the entire original model onto the card.",[10,151,152,155],{},[26,153,154],{},"Paiton is our framework for optimizing AI on AMD hardware."," These measurements use its proprietary native CDNA4 runtime, which is not included with the public weights. Downloading the quantization does not provide a working stock-vLLM deployment or everything needed to reproduce these results.",[157,158,160],"h2",{"id":159},"keeping-useful-work-on-the-gpu","Keeping useful work on the GPU",[10,162,163],{},"The original checkpoint contains about 269 GiB of routed experts, 189 GiB of engram lookup tables and 17 GiB of other components. We compress selectively rather than changing every weight.",[10,165,166],{},"We keep the most-used experts on the GPU. The 3,000 hottest experts retain their original 4-bit values; the next 2,812 use our 3-bit format, at about 3.3 bits per weight including scales. Another 9,548 experts remain in host memory in their original 4-bit format. A set of 1,024 GPU slots follows the experts used by current requests.",[10,168,169],{},"During generation, the host CPUs compute cold experts in place. Only their inputs and outputs cross PCIe, not entire expert weights at each step. For long prompts, the runtime streams cold experts to the GPU instead, where the larger batch makes that worthwhile. The engram tables stay in host memory in a compressed 4-bit row format.",[10,171,172,173,176,177,180],{},"The tested machine has ",[26,174,175],{},"two AMD EPYC 9654 CPUs",", with the MI350P connected over PCIe Gen5 x16. Released host profile B uses approximately ",[26,178,179],{},"287.5 GB for the model",", including host experts, compressed engram tables and pinned buffers. That is a model allocation, not a minimum total system-RAM specification: the OS, loading and other processes need headroom too.",[10,182,183],{},"Separate engine diagnostics point to CPU computation and host-memory bandwidth as constraints under heavier decode load. Enough RAM alone does not guarantee these speeds. We have not measured another host or a second MI350P.",[157,185,187],{"id":186},"two-profiles-two-different-uses","Two profiles, two different uses",[10,189,190],{},"The interactive profile reserves 64 request slots, each configured for up to 70,000 tokens. It uses the model's own speculative drafter for up to 12 running requests: the drafter proposes two tokens, which the main model checks together. We did not record draft acceptance rates for these release runs.",[10,192,193],{},"The token-factory profile disables drafting and shortens the per-request limit to 12,288 tokens, leaving room for 512 slots. It favors total completed work over the speed of an individual answer.",[195,196],"ds41-figure",{"kind":197},"cache",[10,199,200],{},[201,202,203],"em",{},"Runtime-reported physical KV-cache capacity. Slot budgets are configuration limits, not measured simultaneous full-length conversations.",[205,206,207,230],"table",{},[208,209,210],"thead",{},[211,212,213,217,221,224,227],"tr",{},[214,215,216],"th",{},"Profile",[214,218,220],{"align":219},"right","Physical cache, tokens",[214,222,223],{"align":219},"Request slots",[214,225,226],{"align":219},"Limit per request",[214,228,229],{"align":219},"Configured slot budget",[231,232,233,251],"tbody",{},[211,234,235,239,242,245,248],{},[236,237,238],"td",{},"Interactive",[236,240,241],{"align":219},"4,484,095",[236,243,244],{"align":219},"64",[236,246,247],{"align":219},"70,000",[236,249,250],{"align":219},"4,480,000",[211,252,253,256,259,262,265],{},[236,254,255],{},"Token factory",[236,257,258],{"align":219},"6,324,224",[236,260,261],{"align":219},"512",[236,263,264],{"align":219},"12,288",[236,266,267],{"align":219},"6,291,456",[10,269,270],{},"The KV cache holds working state for processed tokens. Neither total is a context window for a single request. We also did not test 64 simultaneous 70K conversations or 512 simultaneous maximum-length requests. Allocating their slots is different from measuring that workload.",[157,272,274],{"id":273},"decode-speed-without-the-prompt-processing-work","Decode speed, without the prompt-processing work",[10,276,277],{},"LocalMaxxing measures sustained generation from an empty context, without the prompt-processing work included in a serving test.",[195,279],{"kind":280},"decode",[10,282,283],{},[201,284,285],{},"Aggregate output tokens\u002Fs in LocalMaxxing. One MI350P at a configured 450 W limit; higher is better. These are not end-to-end request rates.",[205,287,288,314],{},[208,289,290],{},[211,291,292,294,297,300,303,306,309,312],{},[214,293,87],{},[214,295,296],{"align":219},"1",[214,298,299],{"align":219},"2",[214,301,302],{"align":219},"4",[214,304,305],{"align":219},"8",[214,307,308],{"align":219},"16",[214,310,311],{"align":219},"32",[214,313,244],{"align":219},[231,315,316],{},[211,317,318,321,324,327,330,333,336,339],{},[236,319,320],{},"Output tokens\u002Fs",[236,322,323],{"align":219},"77.9",[236,325,326],{"align":219},"112.7",[236,328,329],{"align":219},"166.9",[236,331,332],{"align":219},"241.4",[236,334,335],{"align":219},"377.9",[236,337,338],{"align":219},"518.4",[236,340,341],{"align":219},"709.6",[10,343,344,345,348,349,352,353,93,356,359],{},"Prompt processing, or prefill, reaches ",[26,346,347],{},"4,366 input tokens\u002Fs"," for an 8,119-token prompt and ",[26,350,351],{},"4,227"," for a 63,742-token prompt. The measured first-token times are ",[26,354,355],{},"1.86 seconds",[26,357,358],{},"15.08 seconds",", respectively. These are single-prompt tests, not the latency under simultaneous arrivals.",[157,361,363],{"id":362},"how-this-compares-with-an-rtx-pro-6000","How this compares with an RTX PRO 6000",[10,365,366,369],{},[26,367,368],{},"We keep the original 4-bit values for most routed experts, rather than starting from a 2-bit Q2_K checkpoint."," Our custom 3-bit tier compresses a selected group of GPU experts. That distinction matters as much as the tokens-per-second figure: a faster answer is only useful if the model still gives you the answer you need.",[10,371,372,373,379],{},"The public ",[17,374,378],{"href":375,"rel":376},"https:\u002F\u002Fgithub.com\u002Finferencekey\u002Fdeepseek-one-gpu\u002Fblob\u002F41fa317b1f04dc87075c224c915b7e8c29735bb0\u002Fdocs\u002FRESULTS.md",[377],"nofollow","deepseek-one-gpu results"," provide a useful single-card reference. They use an RTX PRO 6000 Blackwell with CPU expert offload and a Q2_K checkpoint. Here are the separately reported decode results, not a matched GPU benchmark:",[205,381,382,394],{},[208,383,384],{},[211,385,386,388,391],{},[214,387],{},[214,389,390],{},"Paiton \u002F 1 × MI350P",[214,392,393],{},"deepseek-one-gpu \u002F 1 × RTX PRO 6000",[231,395,396,407,418,428,438],{},[211,397,398,401,404],{},[236,399,400],{},"GPU memory",[236,402,403],{},"144 GB",[236,405,406],{},"96 GB",[211,408,409,412,415],{},[236,410,411],{},"Expert weights",[236,413,414],{},"Original FP4 for most experts; custom 3-bit GPU tier",[236,416,417],{},"Q2_K checkpoint, described as 2-bit by its authors",[211,419,420,423,425],{},[236,421,422],{},"One request, output tok\u002Fs",[236,424,323],{},[236,426,427],{},"63.11",[211,429,430,433,435],{},[236,431,432],{},"Four requests, total output tok\u002Fs",[236,434,329],{},[236,436,437],{},"127.58",[211,439,440,443,445],{},[236,441,442],{},"Eight requests, total output tok\u002Fs",[236,444,332],{},[236,446,447],{},"152.53",[10,449,450,453],{},[26,451,452],{},"Paiton reports higher decode rates here while retaining more expert-weight precision."," The measurements differ: ours uses LocalMaxxing's empty-context, 30-second cells; the RTX results average three 15-second windows after reading approximately 16K tokens per request. They also use different CPUs, memory capacities and checkpoints. The RTX one- and four-request results use an EPYC 9575F, while its eight-request result uses an EPYC 9655; ours uses two EPYC 9654s. These numbers do not isolate the GPU or quantify a controlled Paiton speedup.",[455,456,458],"h3",{"id":457},"why-we-kept-more-precision","Why we kept more precision",[10,460,461,462,88,465,93,468,471,472,37],{},"Our served checks measure approximately ",[26,463,464],{},"0.031 chat KL",[26,466,467],{},"0.042–0.044 long-context KL",[26,469,470],{},"94–95% next-token agreement"," against the original model. The retrieval checks also recover every tested passphrase. We publish the sample sizes and limitations in the ",[17,473,474],{"href":108},"accuracy section below",[10,476,477,478,483,484],{},"The RTX project's ",[17,479,482],{"href":480,"rel":481},"https:\u002F\u002Fgithub.com\u002Finferencekey\u002Fdeepseek-one-gpu\u002Fblob\u002F41fa317b1f04dc87075c224c915b7e8c29735bb0\u002FREADME.md#quality",[377],"quality report"," measures downstream tasks instead, including GSM8K and HumanEval, against the DeepSeek API. We have not run those same tests for this release, and it does not provide a matching KL measurement. ",[26,485,486],{},"Our selective quantization preserves more original expert-weight precision; a numerical claim that it is more accurate than Q2_K still needs a shared evaluation.",[455,488,490],{"id":489},"more-single-machine-results","More single-machine results",[10,492,493],{},"The same model family also runs on smaller NVIDIA cards and Apple Silicon. These reports make different tradeoffs between expert precision, GPU memory, host RAM and SSD access. Rates below are the authors' measurements, not reruns on our test bench.",[205,495,496,512],{},[208,497,498],{},[211,499,500,503,506,509],{},[214,501,502],{},"Setup and source",[214,504,505],{},"Weights and memory",[214,507,508],{"align":219},"Reported single-request output",[214,510,511],{},"Test context",[231,513,514,531,549,567],{},[211,515,516,520,523,528],{},[236,517,518],{},[26,519,390],{},[236,521,522],{},"Most experts retain original FP4; selected GPU experts use 3-bit. 144 GB GPU memory, about 290 GB of host model data",[236,524,525],{"align":219},[26,526,527],{},"77.9 tok\u002Fs",[236,529,530],{},"LocalMaxxing, empty starting context, 30-second cells; two EPYC 9654 CPUs compute cold experts",[211,532,533,540,543,546],{},[236,534,535],{},[17,536,539],{"href":537,"rel":538},"https:\u002F\u002Fdev.to\u002Fseppegadeyne\u002Frun-deepseek-v41-flash-748b-locally-with-project-maya-on-one-rtx-5090-4i4a",[377],"RTX 5090 \u002F Project Maya",[236,541,542],{},"Maya-S 2–3-bit experts; 32 GB VRAM, 128 GB RAM, NVMe offload",[236,544,545],{"align":219},"26.3–26.7 tok\u002Fs",[236,547,548],{},"Author's October 29 update, Maya 1.1.27, drafting off; Ryzen 9 9950X3D",[211,550,551,558,561,564],{},[236,552,553],{},[17,554,557],{"href":555,"rel":556},"https:\u002F\u002Fgithub.com\u002Fsybil-solutions\u002Fdsv41-flash-offload\u002Fblob\u002F08802b6fe00101830d75733d2a7c2dd386e121e3\u002FREADME.md",[377],"RTX 3090 \u002F dsv41-flash-offload",[236,559,560],{},"EXL3 3-bit experts; 24 GB VRAM, measured host with 503 GiB DDR4; Engram on NVMe",[236,562,563],{"align":219},"19.3 tok\u002Fs",[236,565,566],{},"Published D141 default; EPYC 7443P. Also reports 26.6 aggregate tok\u002Fs at four streams",[211,568,569,576,579,582],{},[236,570,571],{},[17,572,575],{"href":573,"rel":574},"https:\u002F\u002Fhuggingface.co\u002Frapid-mlx\u002FDeepSeek-V4.1-Flash-1017-MLX\u002Fblob\u002Ff385ddc5cd3bf11582c0c38632223d9338af672b\u002FREADME.md",[377],"Mac Studio M3 Ultra \u002F Rapid-MLX",[236,577,578],{},"Mixed 2–3-bit routed experts; 256 GB unified memory; 4-bit Engram on SSD",[236,580,581],{"align":219},"34.3–42.7 tok\u002Fs",[236,583,584],{},"Default depth-two drafting, roughly 0.7–1K-token prompts and 256 output tokens; up to 49.0 tok\u002Fs on code with depth-four drafting",[10,586,587,590],{},[26,588,589],{},"Our approach keeps more of the original expert weights intact instead of quantizing all routed experts to 2–3 bits."," The smaller-card setups use much less GPU memory, and the RTX 5090 host also has substantially less RAM. Their lower rates are not evidence of an isolated GPU or software disadvantage.",[10,592,593],{},"Quality numbers need the same care. Rapid-MLX reports 0.260 KL and 85.1% top-1 agreement on its 19,456-position held-out set, using full-vocabulary KL. Our roughly 0.031 chat KL and 95% agreement come from 1,767 positions with top-64-bucket KL. Those different tests do not establish an accuracy ratio. The RTX 3090 project's 0.0068 KL checks an execution path against the same quantized pack, not quantization loss against the original model.",[455,595,597],{"id":596},"what-changes-with-more-gpus","What changes with more GPUs?",[10,599,600],{},"More cards can provide substantially higher speeds, but they solve a different hardware-budget problem. These are total setup results, not per-GPU figures.",[205,602,603,615],{},[208,604,605],{},[211,606,607,609,612],{},[214,608,502],{},[214,610,611],{},"Model placement",[214,613,614],{},"Separately reported result",[231,616,617,632],{},[211,618,619,626,629],{},[236,620,621],{},[17,622,625],{"href":623,"rel":624},"https:\u002F\u002Fhuggingface.co\u002Fdiffbot\u002FDeepSeek-V4.1-Flash-EXL3-3bpw-2x-RTX-PRO-6000\u002Fblob\u002F451f7dcb8ecb0502a963f3314e16b34c5a8ce5ae\u002FREADME.md",[377],"2 × RTX PRO 6000 \u002F Diffbot",[236,627,628],{},"192 GB combined VRAM; calibrated EXL3 3-bit routed experts, one third in pinned host RAM",[236,630,631],{},"One-stream agentic replay: 224.1 tok\u002Fs with GPU expert computation, or 261.8 with the optional CPU runtime",[211,633,634,641,644],{},[236,635,636],{},[17,637,640],{"href":638,"rel":639},"https:\u002F\u002Flambda.ai\u002Finference-models\u002Fdeepseek-ai\u002Fdeepseek-v4.1-flash",[377],"4 × B200 \u002F Lambda",[236,642,643],{},"Official checkpoint, vLLM TP4, five speculative tokens",[236,645,646],{},"Approximately 8K input \u002F 1K output: 374.8 aggregate output tok\u002Fs at one request; 3,713.9 at 64 and 4,858.7 at 256",[10,648,649],{},"Diffbot's host-resident experts mean its two-card result is not an all-VRAM deployment. Lambda's numbers include prompt processing across 512 prompts per concurrency point, unlike our LocalMaxxing decode matrix. Neither is a matched comparison with our single-card run.",[10,651,652],{},"For a latency-sensitive service, the extra GPUs may be worth it. Our result addresses a different question: what can one MI350P do with a capable CPU host, while preserving most original expert-weight precision?",[157,654,656],{"id":655},"what-happens-when-requests-arrive-together","What happens when requests arrive together?",[10,658,659,660,663],{},"The serving sweep includes both input processing and generation. Here ",[26,661,662],{},"1K means 1,024 tokens",". Requests use distinct source-text windows, greedy generation and fixed output lengths, with EOS ignored. These are concurrent synthetic requests, not a measurement of the same number of people using an assistant.",[195,665],{"kind":666},"serving-short",[10,668,669],{},[201,670,671],{},"1K input \u002F 1K output. Total throughput and median decode per request measure different things. First-token times include prompt processing.",[205,673,674,692],{},[208,675,676],{},[211,677,678,680,683,686,689],{},[214,679,216],{},[214,681,682],{"align":219},"Requests",[214,684,685],{"align":219},"Total output, tok\u002Fs",[214,687,688],{"align":219},"Decode per request, tok\u002Fs",[214,690,691],{"align":219},"First token, median \u002F sample p99",[231,693,694,709,724,739,754,770,786],{},[211,695,696,698,700,703,706],{},[236,697,238],{},[236,699,296],{"align":219},[236,701,702],{"align":219},"90.7",[236,704,705],{"align":219},"95.4",[236,707,708],{"align":219},"0.69 \u002F 0.77 s",[211,710,711,713,715,718,721],{},[236,712,238],{},[236,714,302],{"align":219},[236,716,717],{"align":219},"154.3",[236,719,720],{"align":219},"42.4",[236,722,723],{"align":219},"0.80 \u002F 2.87 s",[211,725,726,728,730,733,736],{},[236,727,238],{},[236,729,308],{"align":219},[236,731,732],{"align":219},"203.8",[236,734,735],{"align":219},"14.4",[236,737,738],{"align":219},"8.91 \u002F 11.42 s",[211,740,741,743,745,748,751],{},[236,742,238],{},[236,744,244],{"align":219},[236,746,747],{"align":219},"297.2",[236,749,750],{"align":219},"5.2",[236,752,753],{"align":219},"28.25 \u002F 44.64 s",[211,755,756,758,761,764,767],{},[236,757,255],{},[236,759,760],{"align":219},"256",[236,762,763],{"align":219},"542.8",[236,765,766],{"align":219},"2.6",[236,768,769],{"align":219},"91.56 \u002F 178.51 s",[211,771,772,774,777,780,783],{},[236,773,255],{},[236,775,776],{"align":219},"384",[236,778,779],{"align":219},"620.8",[236,781,782],{"align":219},"2.1",[236,784,785],{"align":219},"131.76 \u002F 249.57 s",[211,787,788,790,792,795,798],{},[236,789,255],{},[236,791,261],{"align":219},[236,793,794],{"align":219},"682.5",[236,796,797],{"align":219},"1.7",[236,799,800],{"align":219},"167.90 \u002F 331.24 s",[10,802,803],{},"One request starts answering in under a second. At 16, median first-token time is nearly nine seconds; at 64 it is 28 seconds. The card produces more overall, but each answer slows down. At 512 factory requests, the wait approaches three minutes.",[10,805,806],{},"Per-request decode is derived from median time per output token after the first. It is not total throughput divided by concurrency. Sample p99 describes these recorded requests, not a population percentile or a production service-level guarantee. Variable arrivals, prompt lengths, tool calls and scheduling need their own capacity test.",[157,808,810],{"id":809},"long-inputs-are-the-harder-serving-workload","Long inputs are the harder serving workload",[10,812,813,814,817,818,821,822,825],{},"An 8K input with a 1K answer spends more of its time processing prompts. At 64 requests, the interactive median wait reaches ",[26,815,816],{},"70.45 seconds",". At 512 factory requests it is ",[26,819,820],{},"586.39 seconds",", nearly ",[26,823,824],{},"9.8 minutes",", with a sample p99 of about 19.9 minutes.",[195,827],{"kind":828},"serving-long-input",[10,830,831],{},[201,832,833],{},"8K input \u002F 1K output. The longer prompt raises first-token latency substantially, particularly in the batch profile.",[205,835,836,851],{},[208,837,838],{},[211,839,840,842,844,846,848],{},[214,841,216],{},[214,843,682],{"align":219},[214,845,685],{"align":219},[214,847,688],{"align":219},[214,849,850],{"align":219},"Median first token",[231,852,853,868,883,898,913,928,943],{},[211,854,855,857,859,862,865],{},[236,856,238],{},[236,858,296],{"align":219},[236,860,861],{"align":219},"79.5",[236,863,864],{"align":219},"95.2",[236,866,867],{"align":219},"1.83 s",[211,869,870,872,874,877,880],{},[236,871,238],{},[236,873,302],{"align":219},[236,875,876],{"align":219},"129.8",[236,878,879],{"align":219},"39.9",[236,881,882],{"align":219},"3.87 s",[211,884,885,887,889,892,895],{},[236,886,238],{},[236,888,308],{"align":219},[236,890,891],{"align":219},"151.1",[236,893,894],{"align":219},"12.1",[236,896,897],{"align":219},"17.09 s",[211,899,900,902,904,907,910],{},[236,901,238],{},[236,903,244],{"align":219},[236,905,906],{"align":219},"201.6",[236,908,909],{"align":219},"4.0",[236,911,912],{"align":219},"70.45 s",[211,914,915,917,919,922,925],{},[236,916,255],{},[236,918,760],{"align":219},[236,920,921],{"align":219},"299.1",[236,923,924],{"align":219},"1.8",[236,926,927],{"align":219},"285.42 s",[211,929,930,932,934,937,940],{},[236,931,255],{},[236,933,776],{"align":219},[236,935,936],{"align":219},"317.8",[236,938,939],{"align":219},"1.3",[236,941,942],{"align":219},"434.14 s",[211,944,945,947,949,952,955],{},[236,946,255],{},[236,948,261],{"align":219},[236,950,951],{"align":219},"328.9",[236,953,954],{"align":219},"1.1",[236,956,957],{"align":219},"586.39 s",[10,959,960],{},"Single-prompt prefill cannot predict latency for a queue of long documents. This batch profile does not suit someone waiting for an immediate reply.",[157,962,964],{"id":963},"long-answers-improve-total-throughput","Long answers improve total throughput",[10,966,967,968,971,972,975,976,37],{},"With 1K input and 8K output, prompt processing is spread across much more generated text. The factory reaches ",[26,969,970],{},"1,107.6 tokens\u002Fs at 512 requests",", compared with 682.5 for short answers. Individual generation still runs at only ",[26,973,974],{},"2.3 tokens\u002Fs",", after a median first-token wait of ",[26,977,978],{},"180.84 seconds",[195,980],{"kind":981},"serving-long-output",[10,983,984],{},[201,985,986],{},"1K input \u002F 8K output. More total output does not make each batch request interactive.",[205,988,989,1003],{},[208,990,991],{},[211,992,993,995,997,999,1001],{},[214,994,216],{},[214,996,682],{"align":219},[214,998,685],{"align":219},[214,1000,688],{"align":219},[214,1002,850],{"align":219},[231,1004,1005,1020,1035,1050,1065,1080,1094],{},[211,1006,1007,1009,1011,1014,1017],{},[236,1008,238],{},[236,1010,296],{"align":219},[236,1012,1013],{"align":219},"109.0",[236,1015,1016],{"align":219},"120.0",[236,1018,1019],{"align":219},"0.72 s",[211,1021,1022,1024,1026,1029,1032],{},[236,1023,238],{},[236,1025,302],{"align":219},[236,1027,1028],{"align":219},"199.3",[236,1030,1031],{"align":219},"52.3",[236,1033,1034],{"align":219},"2.99 s",[211,1036,1037,1039,1041,1044,1047],{},[236,1038,238],{},[236,1040,308],{"align":219},[236,1042,1043],{"align":219},"243.6",[236,1045,1046],{"align":219},"15.4",[236,1048,1049],{"align":219},"11.02 s",[211,1051,1052,1054,1056,1059,1062],{},[236,1053,238],{},[236,1055,244],{"align":219},[236,1057,1058],{"align":219},"370.5",[236,1060,1061],{"align":219},"5.9",[236,1063,1064],{"align":219},"29.09 s",[211,1066,1067,1069,1071,1074,1077],{},[236,1068,255],{},[236,1070,760],{"align":219},[236,1072,1073],{"align":219},"813.2",[236,1075,1076],{"align":219},"3.3",[236,1078,1079],{"align":219},"92.90 s",[211,1081,1082,1084,1086,1089,1091],{},[236,1083,255],{},[236,1085,776],{"align":219},[236,1087,1088],{"align":219},"953.4",[236,1090,766],{"align":219},[236,1092,1093],{"align":219},"138.96 s",[211,1095,1096,1098,1100,1103,1106],{},[236,1097,255],{},[236,1099,261],{"align":219},[236,1101,1102],{"align":219},"1,107.6",[236,1104,1105],{"align":219},"2.3",[236,1107,1108],{"align":219},"180.84 s",[10,1110,1111],{},"Offline summaries, evaluations and data generation can tolerate that tradeoff. An interactive coding assistant generally cannot. The useful operating point depends on the output length and acceptable wait, not just the largest throughput number.",[157,1113,1115],{"id":1114},"power-the-limit-is-not-the-consumption","Power: the limit is not the consumption",[10,1117,1118,1119,1122,1123,1126,1127,1130,1131,1134],{},"The ",[26,1120,1121],{},"450 W setting is a GPU power limit",", not a claim that the GPU constantly draws 450 W. At 512 factory requests, mean GPU power is ",[26,1124,1125],{},"212.5 W"," for 1K \u002F 1K and ",[26,1128,1129],{},"214.7 W"," for 1K \u002F 8K. Combining those readings with output gives approximately ",[26,1132,1133],{},"0.086 and 0.054 Wh per 1,000 output tokens",", respectively.",[10,1136,1137,1138,1141,1142,1145,1146,1149],{},"The long-input factory cell draws more, ",[26,1139,1140],{},"273.4 W"," on average at 512 requests, while producing less output. Interactive 1K \u002F 1K cells average ",[26,1143,1144],{},"251.8 W"," at one request and ",[26,1147,1148],{},"207.6 W"," at 64. More batching can improve GPU energy per token even when individual responses slow down.",[10,1151,1152,1153,1156],{},"These figures cover ",[26,1154,1155],{},"GPU board energy only",". They exclude the two CPUs computing offloaded experts, system RAM, storage, power-supply losses and cooling. Without whole-server power measurements, they cannot establish total energy cost or a financial return.",[157,1158,1160],{"id":1159},"how-close-are-the-predictions-to-the-original","How close are the predictions to the original?",[10,1162,1163],{},"We compare teacher-forced next-token distributions against the original model on two MI355X GPUs. KL divergence measures the difference between distributions; zero means identical. The diagnostic groups the reference's top 64 tokens and remaining probability mass. Top-1 agreement asks whether both setups pick the same most-likely next token.",[195,1165],{"kind":1166},"accuracy",[10,1168,1169],{},[201,1170,1171],{},"Served next-token checks at eight concurrent requests, except the explicitly labelled 512-request chat check. These are not downstream answer-accuracy scores.",[205,1173,1174,1193],{},[208,1175,1176],{},[211,1177,1178,1181,1184,1187,1190],{},[214,1179,1180],{},"Profile \u002F check",[214,1182,1183],{"align":219},"Chat KL",[214,1185,1186],{"align":219},"Chat top-1",[214,1188,1189],{"align":219},"Long-context KL",[214,1191,1192],{"align":219},"Long-context top-1",[231,1194,1195,1212,1228],{},[211,1196,1197,1200,1203,1206,1209],{},[236,1198,1199],{},"Interactive, c8",[236,1201,1202],{"align":219},"0.0307",[236,1204,1205],{"align":219},"95.0%",[236,1207,1208],{"align":219},"0.0438",[236,1210,1211],{"align":219},"94.2%",[211,1213,1214,1217,1220,1223,1226],{},[236,1215,1216],{},"Token factory, c8",[236,1218,1219],{"align":219},"0.0311",[236,1221,1222],{"align":219},"94.5%",[236,1224,1225],{"align":219},"0.0418",[236,1227,1211],{"align":219},[211,1229,1230,1233,1236,1238,1241],{},[236,1231,1232],{},"Token factory, c512",[236,1234,1235],{"align":219},"0.0306",[236,1237,1222],{"align":219},[236,1239,1240],{"align":219},"Not measured",[236,1242,1240],{"align":219},[10,1244,1245],{},"Chat covers 1,767 scored positions. The long-context check covers 1,000 positions across 40 conversations. The profile names describe server configurations, not evidence that every accuracy test ran at 64 or 512 requests. A 95% token agreement does not mean 95% of answers are correct. We did not benchmark math, coding, knowledge, multilingual or tool-use tasks for this release.",[10,1247,1248],{},"Repeated runs also differ slightly because of arithmetic and discrete expert routing. Published interactive noise checks give KL of 0.0039 for chat and 0.0033 for long context, but used an older 16K prefill chunk rather than the released 8K chunk. Factory repeats give 0.0006 and 0.0086. Those help interpret the diagnostic; they are not a matching release-build control for every row.",[10,1250,1251,1252,1255,1256,1259,1260,1263],{},"There is an important limit to the long-context result. Our internal gate is ",[26,1253,1254],{},"0.045 KL",". The measured 1,000-position served checks fall below it. On a larger 171,200-position set, the model card reports ",[26,1257,1258],{},"0.0444 from an arithmetic emulator"," and projects about ",[26,1261,1262],{},"0.0456 when served",", slightly above the gate. That is a projection, not a full-set measurement of the running server. Raw full-set emulator results are not in the public proof.",[10,1265,1266,1267,1270,1271,1274],{},"Retrieval checks found all ",[26,1268,1269],{},"12 hidden passphrases"," in interactive prompts up to 68,160 tokens, and all ",[26,1272,1273],{},"six"," factory passphrases up to 11,033 tokens. That supports this narrow retrieval test, not general long-document accuracy.",[157,1276,1278],{"id":1277},"what-happened-with-stock-engines","What happened with stock engines?",[10,1280,1281,1282,1285,1286,1289],{},"Our published stock-attempt summary covers the original checkpoint on the same MI350P host. ",[26,1283,1284],{},"vLLM 0.33.0"," eventually answers tiny prompts on its reference MoE path at about ",[26,1287,1288],{},"0.2 tokens\u002Fs",", after 35 to 47 minutes of loading. A 1,024-token prompt then runs out of GPU memory. The tested fast MoE paths also fail to fit; SGLang 0.5.23 and the tested ROCm preview run out of memory while loading.",[10,1291,1292,1293,1298],{},"These attempts did not produce a usable stock-serving baseline for this card. They are not proof that every possible configuration fails. The ",[17,1294,1297],{"href":1295,"rel":1296},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FDeepSeek-V4.1-Flash-Paiton-CDNA4\u002Ftree\u002F2dd8058eab24484c9e2a54884244d68c825ccd7c\u002Fproof\u002Fstock-attempts",[377],"public stock evidence"," contains a summary, OOM excerpts and a small request record, not complete commands, full container digests and memory traces for every attempt.",[157,1300,1302],{"id":1301},"inspect-the-measurements","Inspect the measurements",[10,1304,1118,1305,1310,1311,1315,1316,1319,1320,1323],{},[17,1306,1309],{"href":1307,"rel":1308},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FDeepSeek-V4.1-Flash-Paiton-CDNA4\u002Ftree\u002F2dd8058eab24484c9e2a54884244d68c825ccd7c",[377],"public weights and model card"," are pinned to revision ",[1312,1313,1314],"code",{},"2dd8058e",". The derivative files use the MIT license, as does the original model. This is a ",[26,1317,1318],{},"135.8 GiB patch",", not a standalone model: the runtime also needs the original ",[26,1321,1322],{},"475 GiB checkpoint",". Stock engines cannot directly load our 3-bit expert format and placement map.",[10,1325,1118,1326,1331,1332,1335],{},[17,1327,1330],{"href":1328,"rel":1329},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FDeepSeek-V4.1-Flash-Paiton-CDNA4\u002Ftree\u002F2dd8058eab24484c9e2a54884244d68c825ccd7c\u002Fproof",[377],"proof directory"," includes LocalMaxxing matrices, per-request serving records, power logs, accuracy checks and the original terminal recording. We checked ",[26,1333,1334],{},"68 proof files against its SHA-256 manifest",". That confirms file integrity, not independent validation of the experiment. Public logs support inspection, but the proprietary runtime and missing evaluation artifacts prevent complete independent reproduction.",[10,1337,1338],{},[1339,1340],"img",{"alt":1341,"src":1342},"Original LocalMaxxing result tables from the Paiton run on one MI350P, including 77.9 tokens\u002Fs at one request and 709.6 at 64.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Finteractive-lmx-tables.png",[10,1344,1345],{},[201,1346,1347,1348,1352],{},"Unaltered benchmark screenshot. The ",[17,1349,1351],{"href":1350},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Flmx-live.gif","original animated terminal recording"," shows a separate live run, not the selected headline matrix.",[10,1354,1355,1356,1359,1360,1363,1364,1368],{},"LocalMaxxing uses ",[1312,1357,1358],{},"lmx"," 0.1.54 and ",[1312,1361,1362],{},"llm-inference-bench"," 0.7.7, with 30-second decode cells. We ran three complete matrices on the warm server and selected the whole run with median single-request decode. The serving sweep uses a separate streaming client and two warm-up rounds. ",[17,1365,1367],{"href":1366},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Fresults.json","Chart data and source paths"," keep the tests separate.",[10,1370,1371,1372,1376,1377,1381],{},"Our ",[17,1373,1375],{"href":1374},"\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p","previous MI350P article"," covers Qwen3.8 Flash Next with much less host computation. The ",[17,1378,1380],{"href":1379},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","Radeon RDNA4 work"," explores a different model and memory budget. Both follow the same Paiton approach: choose the storage format and execution path together, then measure the result on the target hardware.",[10,1383,1384],{},"For DeepSeek, this release makes a large model usable with one MI350P and a substantial CPU host. It does not replace the latency of an all-GPU deployment. The next useful checks are whole-host energy, task quality and a measured two-card configuration.",[10,1386,1387,1388,1392],{},"Learn more about ",[17,1389,1391],{"href":1390},"\u002Fproducts\u002Fpaiton","Paiton"," or inspect the linked evidence before choosing a serving profile.",[10,1394,1395],{},[201,1396,1397],{},"Cover: AI-generated editorial illustration, not a product photograph or benchmark capture.",[157,1399,1401],{"id":1400},"building-a-token-factory-on-amd","Building a token factory on AMD?",[10,1403,1404],{},"If you are planning batch inference or already serving models on AMD hardware, we can help work through throughput, latency and memory constraints with your workload.",[10,1406,1407,37],{},[17,1408,1410],{"href":1409},"\u002Fcontact","Let's talk about your token factory",{"title":1412,"searchDepth":1413,"depth":1413,"links":1414},"",2,[1415,1416,1417,1418,1424,1425,1426,1427,1428,1429,1430,1431],{"id":159,"depth":1413,"text":160},{"id":186,"depth":1413,"text":187},{"id":273,"depth":1413,"text":274},{"id":362,"depth":1413,"text":363,"children":1419},[1420,1422,1423],{"id":457,"depth":1421,"text":458},3,{"id":489,"depth":1421,"text":490},{"id":596,"depth":1421,"text":597},{"id":655,"depth":1413,"text":656},{"id":809,"depth":1413,"text":810},{"id":963,"depth":1413,"text":964},{"id":1114,"depth":1413,"text":1115},{"id":1159,"depth":1413,"text":1160},{"id":1277,"depth":1413,"text":1278},{"id":1301,"depth":1413,"text":1302},{"id":1400,"depth":1413,"text":1401},[1391,1433,1434,1435,1436],"AMD Instinct","AI Inference","DeepSeek","Quantization","2026-11-03T12:00:00Z","DeepSeek V4.1 Flash on one 144 GB MI350P: 77.9 tokens\u002Fs for one request and 1,107.6 in a 512-request batch, with Paiton and CPU expert offload.","md","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Fhero.webp",{},true,"https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p","\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p",{"title":5,"description":1438},"paiton-deepseek-v41-flash-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Fsocial.webp","blog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p",null,"_SGz7KwJJ60lsd1q-eLGvnGX6ZrkoOwgzMygT0G6hXI",[1452,1464,1466,1475,1484,1499,1508,1516,1524,1534,1547,1557,1569,1578,1592,1601,1613,1645,1656,1678,1696,1715,1733,1750,1767,1778,1794,1809,1821,1830,1838,1853,1865,1876,1887,1897,1910,1920,1933,1944,1954,1965,1974,1986,1997,2006],{"path":1453,"title":1454,"description":1455,"date":1456,"slug":1457,"image":1458,"originalUrl":1459,"categories":1460},"\u002Fblog\u002Fpaiton-idol-radeon-ai-pro-r9700","The Radeon AI PRO R9700 does 3D too: Idol in 16.4 seconds","Idol turns an image into a 3D Gaussian splat in 16.44 seconds on one Radeon AI PRO R9700 with Paiton, versus 39.22 seconds for stock on the same GPU.","2026-11-06T09:00:00Z","paiton-idol-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-idol\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-idol-radeon-ai-pro-r9700",[1391,1461,1462,1463],"AMD Radeon","Local AI","3D Generation",{"path":1444,"title":5,"description":1438,"date":1437,"slug":1446,"image":1440,"originalUrl":1443,"categories":1465},[1391,1433,1434,1435,1436],{"path":1374,"title":1467,"description":1468,"date":1469,"slug":1470,"image":1471,"originalUrl":1472,"categories":1473},"Qwen3.8 Flash Next on MI350P: twice the speed | Paiton","Qwen3.8 Flash Next on one AMD MI350P: 2.24× the tested vLLM setup's single-request decode speed and about 2.2 million shared KV-cache tokens with Paiton.","2026-10-30T12:00:00Z","paiton-qwen38-flash-next-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p",[1391,1433,1434,1474,1436],"Qwen",{"path":1476,"title":1477,"description":1478,"date":1479,"slug":1480,"image":1481,"originalUrl":1482,"categories":1483},"\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp","Qwen3.8 Flash Next: 3-bit near 16-bit accuracy, with MTP prefix caching","Paiton's new Qwen3.8 3-bit weights approach 16-bit task scores on two R9700 cards. Native MTP prefix caching resumes a 200K prompt in 0.68 seconds. Results, limits and setup.","2026-10-24T18:45:00Z","paiton-qwen38-3bit-accuracy-prefix-caching-mtp","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-3bit-mtp\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp",[1391,1461,1462,1474,1436],{"path":1485,"title":1486,"description":1487,"date":1488,"slug":1489,"image":1490,"originalUrl":1491,"categories":1492},"\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers","Nvidia confirms the modular DC direction","NVIDIA DSX POD brings Vera Rubin AI factories to modular data centers. ElioVP explains site readiness, cooling and the thinking behind ModFlex.","2026-10-21T13:00:00Z","nvidia-dsx-pod-vera-rubin-modular-data-centers","\u002Fasset\u002Fimages\u002Fblog\u002Fnvidia-dsx-pod-modflex\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fnvidia-dsx-pod-vera-rubin-modular-data-centers",[1493,1494,1495,1496,1497,1498],"Modular DC","AI Infrastructure","ModFlex","NVIDIA","Vera Rubin","Liquid Cooling",{"path":1500,"title":1501,"description":1502,"date":1503,"slug":1504,"image":1505,"originalUrl":1506,"categories":1507},"\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","Qwen3.8 Flash Next in 4-bit on 2 × Radeon R9700: tested against 16-bit","Paiton runs Qwen3.8 Flash Next in 4-bit on two R9700 cards: 96.8% GSM8K, 96.3% HumanEval, 189.8 tokens\u002Fs and an optional RAM prefix cache. Accuracy, speed and setup.","2026-10-12T15:00:00Z","paiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-4bit-radeon-ai-pro-r9700",[1391,1461,1462,1474,1436],{"path":19,"title":1509,"description":1510,"date":1511,"slug":1512,"image":1513,"originalUrl":1514,"categories":1515},"Qwen3.8 Flash Next on 2 × Radeon AI PRO R9700: 216 tokens\u002Fs, 200K context","Paiton runs Qwen3.8 Flash Next with 3-bit experts on two R9700 cards: 216.3 tokens\u002Fs, 565.1 aggregate tokens\u002Fs and a separate 200K mode. Benchmarks, quality and setup.","2026-10-10T10:30:00Z","paiton-qwen38-flash-next-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-flash-next\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700",[1391,1461,1462,1474,1436],{"path":1379,"title":1517,"description":1518,"date":1519,"slug":1520,"image":1521,"originalUrl":1522,"categories":1523},"Qwen3.8 27B on 1 × Radeon AI PRO R9700: 3-bit weights, 20% faster decode","Qwen3.8 27B on one R9700: 19.9% faster weighted decode, a new 4-bit cache, 200K context and optional vision. Benchmarks and current Paiton setup.","2026-09-27T09:00:00Z","paiton-qwen38-w3a4-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-w3a4\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700",[1391,1461,1462,1474,1436],{"path":1525,"title":1526,"description":1527,"date":1528,"slug":1529,"image":1530,"originalUrl":1531,"categories":1532},"\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700","Qwen-Image 2.1 on 1 × Radeon AI PRO R9700: 2048×2048 images in 103 seconds","Generate 2048×2048 Qwen-Image 2.1 images locally with 1 × Radeon AI PRO R9700. The released Paiton v1.0.2 container measured 103.29 seconds per warm request, through PNG delivery.","2026-09-23T09:00:00Z","paiton-qwen-image-21-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen-image-21\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen-image-21-radeon-ai-pro-r9700",[1391,1461,1462,1533,1474],"Image Generation",{"path":1535,"title":1536,"description":1537,"date":1538,"slug":1539,"image":1540,"originalUrl":1541,"categories":1542},"\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700","Qwen3.8: 400.7 tok\u002Fs on R9700 | Paiton","Qwen3.8 on one R9700: 400.7 aggregate tok\u002Fs with ROCm 10 and vLLM 0.29, plus public 200K\u002F220K chat profiles. Benchmarks, limits and launch commands.","2026-09-16T07:30:00Z","paiton-qwen38-mxfp4-dflash2-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-qwen38-mxfp4\u002Fupdate-2026-09-19\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700",[1391,1461,1543,1544,1545,1546],"vLLM","Qwen3.8","Inference Optimization","DFlash2",{"path":1548,"title":1549,"description":1550,"date":1551,"slug":1552,"image":1553,"originalUrl":1554,"categories":1555},"\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700","Qwen3.8 GGUF in vLLM: Faster Responses on One Radeon","Run the original NEO CODER MAX GGUF in vLLM with Paiton on an R9700. Explore measured latency gains, image input and local deployment.","2026-09-14T07:30:00Z","paiton-qwen38-neo-gguf-vllm-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-neo-gguf\u002F00-hero-neo-gguf-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-neo-gguf-vllm-r9700",[1391,1461,1462,1556,1543],"GGUF",{"path":1558,"title":1559,"description":1560,"date":1561,"slug":1562,"image":1563,"originalUrl":1564,"categories":1565},"\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700","MiniMax H3 on Radeon: 15-Second Video With Native Sound","Paiton generates a 15-second MiniMax H3 video with stereo audio on one Radeon AI PRO R9700 in 5m 33s, with 16.7% lower latency than matched stock.","2026-09-09T07:30:00Z","paiton-minimax-h3-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-minimax-h3\u002F00-featured-minimax-h3-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-minimax-h3-radeon-ai-pro-r9700",[1391,1461,1462,1566,1567,1568],"Video Generation","MiniMax H3","ComfyUI",{"path":1570,"title":1571,"description":1572,"date":1573,"slug":1574,"image":1575,"originalUrl":1570,"categories":1576},"\u002Fblog\u002Fpaiton-flux2-klein-radeon-ai-pro-r9700","Local FLUX.2 klein on Radeon AI PRO R9700: Faster Image Generation with Less VRAM","Paiton generates 1024×1024 FLUX.2 klein images in 1.054 seconds on an R9700: 16.2% lower warm latency and 33.4% lower peak Torch allocation.","2026-09-07T09:00:00","paiton-flux2-klein-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-flux2-klein\u002Ffox-paiton.webp",[1391,1461,1462,1533,1577,1568],"FLUX",{"path":1579,"title":1580,"description":1581,"date":1582,"slug":1583,"image":1584,"originalUrl":1585,"categories":1586},"\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700","Ornith 1.5 at 44.6 tok\u002Fs on One Radeon AI PRO R9700","Paiton serves Ornith 1.5 35B A3B at 44.63 output tok\u002Fs on one Radeon AI PRO R9700, 27% faster than tuned stock vLLM, with 21.3% lower modeled cost.","2026-09-05T09:00:00","paiton-ornith15-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-ornith15\u002F00-featured-ornith15-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-ornith15-radeon-ai-pro-r9700",[1391,1587,1461,1434,1588,1589,1545,1590,1543,1591],"Artificial Intelligence","GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1593,"title":1594,"description":1595,"date":1596,"slug":1597,"image":1598,"originalUrl":1599,"categories":1600},"\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700","Paiton: 21% More Qwen3.8 Throughput on Radeon AI PRO R9700","Paiton serves AMD’s Qwen3.8 27B at 39.77 output tokens\u002Fs on one Radeon AI PRO R9700, delivering 21% more throughput and 17.4% lower modeled cost.","2026-09-04T09:00:00","paiton-qwen38-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-r9700\u002F00-featured-paiton-r9700.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-radeon-ai-pro-r9700",[1391,1587,1461,1434,1588,1589,1545,1590,1543,1591],{"path":1602,"title":1603,"description":1604,"date":1605,"slug":1606,"image":1607,"originalUrl":1449,"categories":1608},"\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt","AI Data Center Power Requirements: The GPU\u002FMW Illusion","Why do AI data-center proposals quote different GPU capacities? See how PUE, peak loads, storage, networking and cooling determine deployable compute.","2026-07-27T23:52:00","ai-data-center-power-requirements-gpu-per-megawatt","\u002Fasset\u002Fimages\u002Fblog\u002Fai-data-center-power-requirements-gpu-per-megawatt\u002Fgpu-per-megawatt-illusion.webp",[1609,1494,1610,1495,1611,1612],"All","Data Centers","HPC","AMD Helios",{"path":1614,"title":1615,"description":1616,"date":1617,"slug":1618,"image":1619,"originalUrl":1620,"categories":1621},"\u002Fblog\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","Wan2.2 Video Generation: Paiton on AMD MI355X","Compare Wan2.2-T2V-A14B video generation on AMD MI355X with Paiton and NVIDIA B200 using Diffusers, and explore our diffusion optimization approach.","2026-06-10T14:04:04","paiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonwan2.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-returns-to-its-diffusion-roots-optimizing-wan2-2-t2v-a14b-on-amd-mi355x\u002F",[1609,1587,1391,1622,1623,1624,1625,1626,1627,1628,1629,1630,1631,1632,1633,1634,1635,1636,1637,1638,1391,1639,1640,1641,1642,1643,1644],"14B","AMD","B200","Benchmarks","Blackwell","Compute","Diffusion","Eliovp","Generative AI","GPU","Hardware","Inference","Instinct","MI355x","NVidia","On-Premise","Optimization","Sovereign AI","T2V","Text-to-Video","Tuning","Video-Generation","Wan2.2",{"path":1646,"title":1647,"description":1648,"date":1649,"slug":1650,"image":1651,"originalUrl":1652,"categories":1653},"\u002Fblog\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","ElioVP in De Tijd: Chip Optimization and Data Centers","Read about De Tijd's coverage of ElioVP, from its origins in chip optimization to its work on modular data centers and high-density cooling.","2026-02-10T20:48:12","from-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fphysicalnewspaper.webp","https:\u002F\u002Feliovp.com\u002Ffrom-the-attic-to-the-front-page-eliovp-recognized-as-a-pioneer-in-chip-optimization-data-center-infrastructure\u002F",[1609,1587,1493,1654,1623,1655,1493,1636],"Uncategorized","De Tijd",{"path":1657,"title":1658,"description":1659,"date":1660,"slug":1661,"image":1662,"originalUrl":1663,"categories":1664},"\u002Fblog\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","AI Privacy: A Strategic Priority for Benelux Businesses","Explore generative AI privacy risks, trust, data retention and governance, and why Benelux businesses need a strategic approach to secure AI.","2026-01-29T13:51:11","privacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fheaderimage.webp","https:\u002F\u002Feliovp.com\u002Fprivacy-is-geen-it-probleem-meer-het-is-een-strategische-prioriteit\u002F",[1609,1587,1665,1654,1666,1667,1668,1669,1670,1671,1672,1673,1629,1674,1630,1675,1676,1677],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1679,"title":1680,"description":1681,"date":1682,"slug":1683,"image":1684,"originalUrl":1685,"categories":1686},"\u002Fblog\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","Why We Do Not Use itsme: Privacy and Data Sovereignty","Why ElioVP does not use itsme: our assessment of identity metadata, cloud dependence, data sovereignty and authentication risks.","2025-11-27T09:32:14","itsme-bij-ons-is-het-its-not-me-en-dit-is-waarom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontimage.webp","https:\u002F\u002Feliovp.com\u002Fitsme-bij-ons-is-het-its-not-me-en-dit-is-waarom\u002F",[1609,1687,1688,1689,1671,1690,1691,1692,1674,1693,1694,1695,1676],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1697,"title":1698,"description":1699,"date":1700,"slug":1701,"image":1702,"originalUrl":1703,"categories":1704},"\u002Fblog\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025","Field Report. The Reality of Building Agentic AI in 2025","Lessons from building on-premise AI agents in 2025 cover workflow design, observability, model training, hallucinations and GPU memory limits.","2025-11-25T14:03:39","field-report-the-reality-of-building-agentic-ai-in-2025","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffieldreport.webp","https:\u002F\u002Feliovp.com\u002Ffield-report-the-reality-of-building-agentic-ai-in-2025\u002F",[1609,1587,1705,1665,1706,1707,1708,1709,1710,1711,1712,1713,1639,1714],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1716,"title":1717,"description":1718,"date":1719,"slug":1720,"image":1721,"originalUrl":1722,"categories":1723},"\u002Fblog\u002Fthe-synthetic-unicorn-bubble","The Synthetic Unicorn Bubble","An analysis of AI neocloud investment risks, examining circular financing, infrastructure claims, contract terms and due diligence.","2025-11-24T19:22:28","the-synthetic-unicorn-bubble","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsyntheticunicorn.webp","https:\u002F\u002Feliovp.com\u002Fthe-synthetic-unicorn-bubble\u002F",[1609,1587,1665,1494,1724,1725,1726,1727,1728,1729,1730,1731,1732],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1734,"title":1735,"description":1736,"date":1737,"slug":1738,"image":1739,"originalUrl":1740,"categories":1741},"\u002Fblog\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","NVIDIA GB300 NVL72: A Four-Month Modular Data Center Plan","Explore a modular data center design for NVIDIA GB300 NVL72, covering redundant power, hybrid cooling and a four-month deployment plan.","2025-11-20T14:10:19","building-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fsuperpodmodflexfrontimage.webp","https:\u002F\u002Feliovp.com\u002Fbuilding-the-engine-for-the-ai-race-the-4-month-path-to-nvidia-gb300-nvl72-power\u002F",[1609,1493,1654,1742,1494,1743,1744,1498,1745,1746,1747,1748,1749],"150kW Rack","DLC","High Density","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1751,"title":1752,"description":1753,"date":1754,"slug":1755,"image":1756,"originalUrl":1757,"categories":1758},"\u002Fblog\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential","Why “CUDA” Translation Won’t Unlock AMD’s Real Potential","Why CUDA compatibility is not the same as AMD performance: explore ROCm, HIP, kernel tuning and the case for hardware-specific optimization.","2025-11-12T14:48:37","why-cuda-translation-wont-unlock-amds-real-potential","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fchatgpt-image-nov-11-2025-09_16_10-pm-1.webp","https:\u002F\u002Feliovp.com\u002Fwhy-cuda-translation-wont-unlock-amds-real-potential\u002F",[1609,1587,1391,1654,1759,1587,1760,1761,1762,1763,1764,1765,1391,1766],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1768,"title":1769,"description":1770,"date":1771,"slug":1772,"image":1773,"originalUrl":1774,"categories":1775},"\u002Fblog\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference","Paiton: The Simplest Way to Supercharge AI Inference","Learn how Paiton integrates with existing inference stacks, with AMD MI300X benchmark results and performance-per-dollar comparisons.","2025-11-11T10:31:22","paiton-the-simplest-way-to-supercharge-ai-inference","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton-powaaah.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-the-simplest-way-to-supercharge-ai-inference\u002F",[1609,1587,1391,1434,1433,1759,1591,1776,1545,1765,1391,1777,1543],"High Throughput","SGLang",{"path":1779,"title":1780,"description":1781,"date":1782,"slug":1783,"image":1784,"originalUrl":1785,"categories":1786},"\u002Fblog\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","Paiton MoE Benchmarks: MI300X vs H200 and B200","Compare Qwen3-30B-A3B MoE inference with Paiton on MI300X against H200 and B200, including throughput and cost per million tokens.","2025-09-26T13:36:18","stop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhulkvshulkpaitonwins.webp","https:\u002F\u002Feliovp.com\u002Fstop-overpaying-paiton-mi300x-moe-beats-h200-b200-on-1m-tokens\u002F",[1609,1587,1391,1787,1759,1788,1545,1789,1790,1791,1792,1391,1793],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":1795,"title":1796,"description":1797,"date":1798,"slug":1799,"image":1800,"originalUrl":1801,"categories":1802},"\u002Fblog\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","Local Agentic AI: From Inbox to Action","Local-first AI agents turn email, documents and images into tickets, reports and actions, using models tailored to your data and systems.","2025-09-16T13:09:00","agentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ffrontfotoblog.webp","https:\u002F\u002Feliovp.com\u002Fagentic-ai-but-make-it-local-from-inbox-to-insight-to-action-en\u002F",[1609,1587,1705,1654,1706,1803,1804,1805,1806,1711,1713,1639,1807,1808],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":1810,"title":1811,"description":1812,"date":1813,"slug":1814,"image":1815,"originalUrl":1816,"categories":1817},"\u002Fblog\u002Fmi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach","MI300X FP8 Benchmarks: GPU Partitioning with Paiton","Explore Llama 3.1 8B FP8 benchmarks on partitioned MI300X GPUs with Paiton, comparing throughput and latency against NVIDIA H200 and B200.","2025-07-31T13:32:57","mi300x-fp8-data-parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fimage-2-1.webp","https:\u002F\u002Feliovp.com\u002Fmi300x-fp8-data%e2%80%91parallel-benchmarks-8-64-gpus-h200-left-behind-b200-within-reach\u002F",[1609,1587,1391,1654,1818,1623,1624,1819,1820,1635,1636,1391,1543],"AI","H200","MI300X",{"path":1822,"title":1823,"description":1824,"date":1825,"slug":1826,"image":1827,"originalUrl":1828,"categories":1829},"\u002Fblog\u002Fapplicable-ai-for-businesses","Applicable AI for Businesses","Explore ElioVP's approach to local AI for business workflows, including custom model training and automated damage detection for logistics.","2025-07-09T21:35:30","applicable-ai-for-businesses","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fscherm_afbeelding-2025-07-09-om-23.30.17.webp","https:\u002F\u002Feliovp.com\u002Fapplicable-ai-for-businesses\u002F",[1609,1587,1705,1706,1803,1804,1805,1806,1711,1713,1639,1807,1808],{"path":1831,"title":1832,"description":1833,"date":1834,"slug":1835,"image":1412,"originalUrl":1836,"categories":1837},"\u002Fblog\u002Fintroducing-paitons-free-evaluation-models","Introducing Paiton’s Free Evaluation Models","Test Paiton with free evaluation models for AMD GPUs. Compare text, vision and image generation performance using your own workloads.","2025-07-07T11:26:13","introducing-paitons-free-evaluation-models","https:\u002F\u002Feliovp.com\u002Fintroducing-paitons-free-evaluation-models\u002F",[1609,1587,1391],{"path":1839,"title":1840,"description":1841,"date":1842,"slug":1843,"image":1844,"originalUrl":1845,"categories":1846},"\u002Fblog\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","Llama 3.1 405B: Faster Startup with Paiton on MI300X","See Paiton benchmarks for Llama 3.1 405B on eight AMD MI300X GPUs, covering model startup, tensor parallelism, throughput and latency.","2025-06-12T20:15:23","paiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fservingscreenshot.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-dramatically-faster-startup-and-performance-for-llama-3-1-405b\u002F",[1609,1587,1391,1654,1434,1759,1847,1761,1848,1849,1850,1391,1851,1852],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":1854,"title":1855,"description":1856,"date":1857,"slug":1858,"image":1859,"originalUrl":1860,"categories":1861},"\u002Fblog\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x","Paiton FP8 Beats NVIDIA’s H200 on AMD’s MI300X","Compare Paiton on AMD MI300X with NVIDIA H200 for Llama 3.1 70B FP8, including throughput, first-token delay and latency across batch sizes.","2025-06-08T19:12:40","paiton-fp8-beats-nvidias-h200-on-amds-mi300x","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fblognewfp8.webp","https:\u002F\u002Feliovp.com\u002Fpaiton-fp8-beats-nvidias-h200-on-amds-mi300x\u002F",[1609,1587,1391,1654,1759,1862,1710,1630,1588,1589,1590,1849,1863,1864],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":1866,"title":1867,"description":1868,"date":1869,"slug":1870,"image":1871,"originalUrl":1872,"categories":1873},"\u002Fblog\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","MI300X vs H200 vs RX 7900 XTX vs Tenstorrent n300s with vLLM","Compare MI300X, H200, RX 7900 XTX and Tenstorrent n300s on Llama 3 8B with vLLM, including throughput, modeled token costs and hardware limits.","2025-05-09T14:03:58","mi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisontenstor.webp","https:\u002F\u002Feliovp.com\u002Fmi300x-vs-h200-vs-rx-7900-xtx-vs-tenstorrent-n300s-with-vllm\u002F",[1609,1587,1391,1705,1654,1623,1820,1636,1874,1875],"RX7900XTX","tenstorrent",{"path":1877,"title":1878,"description":1879,"date":1880,"slug":1881,"image":1882,"originalUrl":1883,"categories":1884},"\u002Fblog\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","ClusterP&L: Financial Modeling for GPU Clusters","Explore how ClusterP&L models GPU cluster costs, profitability and investment scenarios, with ROI metrics, risk simulations and exportable reports.","2025-05-03T10:52:22","clusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fcomparisonscenarios.webp","https:\u002F\u002Feliovp.com\u002Fclusterpl-empowering-gpu-cluster-investors-with-real-world-financial-insights\u002F",[1609,1587,1493,1705,1624,1819,1885,1636,1886],"MI325x","pnl calculator",{"path":1888,"title":1889,"description":1890,"date":1891,"slug":1892,"image":1893,"originalUrl":1894,"categories":1895},"\u002Fblog\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","AMD MI300X vs. NVIDIA H200: Qwen3-32B with Paiton","Compare Qwen3-32B benchmarks on Paiton-optimized AMD MI300X and NVIDIA H200, covering throughput, latency and hardware costs.","2025-05-02T21:10:30","cranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002F3ac59a73-2466-4422-b7e5-ef2e4a8ca58e.webp","https:\u002F\u002Feliovp.com\u002Fcranking-out-faster-tokens-for-fewer-dollars-amd-mi300x-vs-nvidia-h200\u002F",[1609,1587,1391,1818,1623,1819,1896,1636,1391,1543],"MI300",{"path":1898,"title":1899,"description":1900,"date":1901,"slug":1902,"image":1903,"originalUrl":1904,"categories":1905},"\u002Fblog\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","Modular Data Centers for NVIDIA NVL: 1 to 2 MW","Explore modular data center designs for NVIDIA NVL systems, covering power capacity, liquid cooling, redundancy and deployment planning.","2025-05-02T14:09:59","power-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Feliovp_critical-1mw-pod_rev-2_transparent.webp","https:\u002F\u002Feliovp.com\u002Fpower-meets-precision-high-density-modular-data-center-for-nvidia-nvl-deployments-1-2-mw\u002F",[1609,1493,1906,1494,1744,1611,1498,1745,1907,1908,1748,1909],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":1911,"title":1912,"description":1913,"date":1914,"slug":1915,"image":1916,"originalUrl":1917,"categories":1918},"\u002Fblog\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","Examining AI agents in the medical field: AI that speaks DICOM","Explore a local AI agent demo for DICOM workflows, from patient and study retrieval to a comparison of vision models using anonymized medical images.","2025-04-11T14:45:48","examining-ai-agents-in-the-medical-field-ai-that-speaks-dicom","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fhealthcareblog-1.webp","https:\u002F\u002Feliovp.com\u002Fexamining-ai-agents-in-the-medical-field-ai-that-speaks-dicom\u002F",[1609,1587,1705,1654,1818,1623,1919],"Healthcare",{"path":1921,"title":1922,"description":1923,"date":1924,"slug":1925,"image":1926,"originalUrl":1927,"categories":1928},"\u002Fblog\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","U.S. Tariffs and AI Supply Chain Resilience: April 2025","Read ElioVP's April 2025 perspective on U.S. tariffs and supply chain resilience for AI servers, HPC systems and modular data centers.","2025-04-04T10:01:27","eliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftariffsshipping.webp","https:\u002F\u002Feliovp.com\u002Feliovp-bv-your-trusted-partner-for-supply-chain-resilience-amidst-new-u-s-tariffs\u002F",[1609,1665,1818,1623,1929,1930,1931,1932],"import","Taiwan","Tariffs","Trump",{"path":1934,"title":1935,"description":1936,"date":1937,"slug":1938,"image":1939,"originalUrl":1940,"categories":1941},"\u002Fblog\u002Fwhy-ai-agents-are-the-future","Why AI Agents Are the Future","Explore AI agents for ERP, CRM, finance and customer support, with practical use cases and a path from workflow assessment to pilot and deployment.","2025-03-23T22:06:59","why-ai-agents-are-the-future","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ferp2.jpeg","https:\u002F\u002Feliovp.com\u002Fwhy-ai-agents-are-the-future\u002F",[1609,1587,1705,1818,1942,1943],"AI Agents","ERP",{"path":1945,"title":1946,"description":1947,"date":1948,"slug":1949,"image":1950,"originalUrl":1951,"categories":1952},"\u002Fblog\u002Fthe-rise-of-open-source-ai-model-optimization","The Rise of Open-Source AI Model Optimization","Explore open-source AI optimization trends, from quantization and mixture-of-experts models to hardware-aware tuning, RAG and edge deployment.","2025-03-22T20:59:23","the-rise-of-open-source-ai-model-optimization","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Friseofopensource.jpeg","https:\u002F\u002Feliovp.com\u002Fthe-rise-of-open-source-ai-model-optimization\u002F",[1609,1587,1665,1953,1623,1631,1636],"AI news",{"path":1955,"title":1956,"description":1957,"date":1958,"slug":1959,"image":1960,"originalUrl":1961,"categories":1962},"\u002Fblog\u002Fintroducing-our-benchmarking-tool-powered-by-dstack","Introducing Our Benchmarking Tool: Powered by dstack","Explore our dstack-powered tool for reproducible vLLM benchmarks, automated parameter sweeps and performance reports across local and cloud GPUs.","2025-03-20T14:21:59","introducing-our-benchmarking-tool-powered-by-dstack","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fbenchmarktool.jpeg","https:\u002F\u002Feliovp.com\u002Fintroducing-our-benchmarking-tool-powered-by-dstack\u002F",[1609,1587,1391,1818,1623,1963,1964,1820,1391],"benchmark","LLM",{"path":1966,"title":1967,"description":1968,"date":1969,"slug":1970,"image":1971,"originalUrl":1972,"categories":1973},"\u002Fblog\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","Optimizing QwQ-32B (by Qwen): AMD MI300X vs. NVIDIA H200","Compare QwQ-32B throughput and latency on AMD MI300X with Paiton and NVIDIA H200, from small batches to higher concurrency.","2025-03-19T21:41:44","optimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaiton4.jpeg","https:\u002F\u002Feliovp.com\u002Foptimizing-qwq-32b-by-qwen-amd-mi300x-vs-nvidia-h200\u002F",[1609,1587,1391],{"path":1975,"title":1976,"description":1977,"date":1978,"slug":1979,"image":1980,"originalUrl":1981,"categories":1982},"\u002Fblog\u002Feliovp-featured-on-amd-tech-talk-podcast","Eliovp Featured on AMD “Tech Talk” Podcast","Listen to Elio Van Puyvelde and Jim Greene on AMD's Tech Talk podcast, discussing ElioVP's origins and its AI hardware and software services.","2025-03-19T07:53:39","eliovp-featured-on-amd-tech-talk-podcast","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Ftechtalkjimgreene.jpeg","https:\u002F\u002Feliovp.com\u002Feliovp-featured-on-amd-tech-talk-podcast\u002F",[1609,1623,1983,1984,1985],"Jim Greene","Podcast","Tech Talk",{"path":1987,"title":1988,"description":1989,"date":1990,"slug":1991,"image":1992,"originalUrl":1993,"categories":1994},"\u002Fblog\u002Ffurther-optimizing-amd-powered-inference-with-paiton","Further Optimizing AMD-Powered Inference with Paiton","Explore Paiton's DeepSeek R1 Distill Llama 8B benchmarks on AMD MI300X, focusing on throughput and first-token latency at smaller batch sizes.","2025-03-13T06:18:30","further-optimizing-amd-powered-inference-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost3.webp","https:\u002F\u002Feliovp.com\u002Ffurther-optimizing-amd-powered-inference-with-paiton\u002F",[1609,1587,1391,1623,1995,1996,1819,1820,1885,1391,1543],"Deepseek","H100",{"path":1998,"title":1999,"description":2000,"date":2001,"slug":2002,"image":2003,"originalUrl":2004,"categories":2005},"\u002Fblog\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","Paiton Benchmarks: DeepSeek R1 Distill Llama 3.1 8B","Compare stock and Paiton-optimized DeepSeek R1 Distill Llama 3.1 8B on AMD MI300X, with throughput and latency benchmarks across batch sizes.","2025-01-31T09:11:02","a-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost2.webp","https:\u002F\u002Feliovp.com\u002Fa-first-look-at-paiton-in-action-deepseek-r1-distill-llama-3-1-8b\u002F",[1609,1587,1391,1623,1995,1996,1819,1820,1885,1391,1543],{"path":2007,"title":2008,"description":2009,"date":2010,"slug":2011,"image":2012,"originalUrl":2013,"categories":2014},"\u002Fblog\u002Fai-model-optimization-with-paiton","AI Model Optimization with Paiton","Learn how Paiton uses model compilation, custom kernels and kernel fusion to optimize AI inference on AMD GPUs.","2025-01-30T19:53:25","ai-model-optimization-with-paiton","\u002Fasset\u002Fimages\u002Fblog\u002Fimported\u002Fpaitonpost1.webp","https:\u002F\u002Feliovp.com\u002Fai-model-optimization-with-paiton\u002F",[1609,1587,1391,1623,1996,1819,1820,1885,1391,1543],1793962342963]