[{"data":1,"prerenderedAt":2013},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p":3,"blog-posts-sidebar-en":1468},{"id":4,"title":5,"body":6,"categories":1448,"date":1453,"description":1454,"extension":1455,"heading":1456,"image":1457,"meta":1458,"navigation":1459,"originalUrl":1460,"path":1461,"seo":1462,"slug":1463,"socialImage":1464,"stem":1465,"updated":1466,"video":1466,"__hash__":1467},"blog\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p.md","Qwen3.8 Flash Next on MI350P: twice the speed | Paiton",{"type":7,"value":8,"toc":1426},"minimark",[9,25,32,35,55,61,66,88,103,125,228,238,241,244,314,318,327,331,336,378,385,403,409,413,419,433,438,441,452,459,466,469,476,480,488,491,499,663,666,680,684,687,690,695,767,782,786,789,792,797,858,869,876,883,887,900,903,912,932,943,947,958,961,966,1042,1057,1060,1064,1071,1074,1079,1134,1141,1144,1148,1163,1179,1197,1200,1204,1211,1220,1224,1231,1241,1279,1283,1299,1324,1334,1338,1344,1350,1363,1377,1391,1394,1404,1409,1413,1416,1422],[10,11,12,16,17,20,21,24],"p",{},[13,14,15],"strong",{},"On one CDNA4-based AMD Instinct MI350P, Paiton generates text at 273.1 tokens per second, compared with 122.1 for our tested stock vLLM + AITER setup."," That is ",[13,18,19],{},"2.24× the speed"," for a single request. With 64 requests running together, total output reaches ",[13,22,23],{},"2,508.6 tokens\u002Fs",".",[10,26,27,28,31],{},"There is more room for conversation history, too: ",[13,29,30],{},"about 2.2 million tokens of GPU KV-cache capacity",", 22% more than the tested vLLM setup. That capacity is shared across requests, not a 2.2-million-token context window for one conversation.",[10,33,34],{},"The practical benefit is less waiting while text appears and more output from the same GPU. The headline comparison measures generation from an empty context, not the complete time from prompt to answer. It compares our A5 weights running on Paiton's native CDNA4 runtime against AMD Quark MXFP4 on vLLM, so it is a full-stack comparison rather than an engine-only test.",[10,36,37,38,41,42,49,50,24],{},"Our 4-bit quantization also stays close to the original model: ",[13,39,40],{},"93.78% next-token agreement with BF16"," in the separate weight evaluation. That measures agreement between predictions, not the percentage of correct answers. The ",[43,44,48],"a",{"href":45,"rel":46},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-CDNA4",[47],"nofollow","weights and quantization recipe are available"," under the upstream ",[43,51,54],{"href":52,"rel":53},"https:\u002F\u002Fhuggingface.co\u002FQwen\u002FQwen3.8-Flash-Next\u002Fblob\u002Fmain\u002FLICENSE",[47],"Qwen Community License 1.0",[10,56,57,60],{},[13,58,59],{},"These are results from Paiton, our AI optimization framework, using its native CDNA4 runtime."," The quantization targets the CDNA4 family, including MI350P, MI350X and MI355X, with a compatible runtime and sufficient GPU and host memory. The measurements below are from one MI350P only. The A5 format currently requires the proprietary Paiton runtime, which is not distributed with the weights; stock vLLM, Transformers and llama.cpp cannot load it directly.",[62,63,65],"h2",{"id":64},"what-paiton-delivers-on-one-mi350p","What Paiton delivers on one MI350P",[10,67,68,69,72,73,76,77,80,81,83,84,87],{},"With the A5 weights and Paiton's native CDNA4 runtime, one ",[13,70,71],{},"144 GB AMD Instinct MI350P at a 450 W power limit"," generates ",[13,74,75],{},"273.1 tokens\u002Fs for a single request"," in LocalMaxxing. That is ",[13,78,79],{},"2.24×"," our tested stock vLLM + AITER setup using AMD's MXFP4 checkpoint. At 64 concurrent requests, aggregate decode reaches ",[13,82,23],{},", with ",[13,85,86],{},"2.19 million KV-cache tokens"," available on the GPU.",[10,89,90,91,94,95,98,99,102],{},"These are ",[13,92,93],{},"our measurements",", not independent certification. They compare complete stacks: ",[13,96,97],{},"Paiton's A5 weights and native runtime"," against ",[13,100,101],{},"AMD Quark MXFP4 weights on vLLM or SGLang",". The weights and execution paths differ, so this is not an engine-only comparison or the speed you get just by downloading our quant.",[10,104,105,106,109,110,113,114,117,118,121,122,24],{},"The separate serving test adds 1,024 input tokens followed by exactly 1,024 output tokens. With one request, Paiton starts answering in a median ",[13,107,108],{},"98 milliseconds",". With 16 requests in flight, that becomes ",[13,111,112],{},"525 ms",", compared with ",[13,115,116],{},"1,782 ms"," for the tested stock vLLM setup. At 64, Paiton's median is ",[13,119,120],{},"861 ms",", against ",[13,123,124],{},"2,811 ms",[126,127,128,148],"table",{},[129,130,131],"thead",{},[132,133,134,138,142,145],"tr",{},[135,136,137],"th",{},"Concurrent requests",[135,139,141],{"align":140},"right","Paiton total output, tok\u002Fs",[135,143,144],{"align":140},"Decode per request, tok\u002Fs",[135,146,147],{"align":140},"First token, median \u002F sample p99",[149,150,151,166,180,194,208],"tbody",{},[132,152,153,157,160,163],{},[154,155,156],"td",{},"1",[154,158,159],{"align":140},"331",[154,161,162],{"align":140},"337.6",[154,164,165],{"align":140},"98 \u002F 101 ms",[132,167,168,171,174,177],{},[154,169,170],{},"4",[154,172,173],{"align":140},"746",[154,175,176],{"align":140},"193.8",[154,178,179],{"align":140},"150 \u002F 332 ms",[132,181,182,185,188,191],{},[154,183,184],{},"16",[154,186,187],{"align":140},"1,046",[154,189,190],{"align":140},"67.6",[154,192,193],{"align":140},"525 \u002F 1,300 ms",[132,195,196,199,202,205],{},[154,197,198],{},"32",[154,200,201],{"align":140},"1,444",[154,203,204],{"align":140},"46.4",[154,206,207],{"align":140},"632 \u002F 2,572 ms",[132,209,210,213,218,223],{},[154,211,212],{},"64",[154,214,215],{"align":140},[13,216,217],{},"1,790",[154,219,220],{"align":140},[13,221,222],{},"28.9",[154,224,225],{"align":140},[13,226,227],{},"861 \u002F 5,530 ms",[10,229,230],{},[231,232,233,234,24],"em",{},"1K input \u002F 1K output serving sweep. Per-request decode is the reciprocal of median time per output token after the first; it is not total throughput divided by concurrency. Sample p99 describes these recorded requests, not a production guarantee. ",[43,235,237],{"href":236},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fresults.json","Chart data",[10,239,240],{},"As load rises, each answer arrives more slowly, but the card does more work in total. Per-request speed affects how quickly an assistant replies; aggregate output determines how many requests can share the hardware.",[10,242,243],{},"These are fixed-length requests using source-text windows, not a trial with 64 people chatting. Real traffic brings variable prompts, pauses, tool calls and arrival rates. The table is a measured starting point for capacity planning, not a substitute for testing that traffic.",[245,246,247,251,262,307],"details",{},[248,249,250],"summary",{},"An illustrative card count for 1,000 simultaneously active users",[10,252,253,254,257,258,261],{},"Suppose the target is median decode of at least ",[13,255,256],{},"20 English words\u002Fs"," and ",[13,259,260],{},"sample-p99 first-token time within 10 seconds",". Using an illustrative 0.75 words\u002Ftoken conversion, the highest tested 1K \u002F 1K concurrency meeting both thresholds gives:",[126,263,264,277],{},[129,265,266],{},[132,267,268,271,274],{},[135,269,270],{},"Setup",[135,272,273],{"align":140},"Requests per card under this rule",[135,275,276],{"align":140},"Cards for 1,000",[149,278,279,288,297],{},[132,280,281,284,286],{},[154,282,283],{},"Paiton",[154,285,212],{"align":140},[154,287,184],{"align":140},[132,289,290,293,295],{},[154,291,292],{},"vLLM + AITER",[154,294,198],{"align":140},[154,296,198],{"align":140},[132,298,299,302,304],{},[154,300,301],{},"SGLang MTP, 16 slots",[154,303,184],{"align":140},[154,305,306],{"align":140},"63",[10,308,309,310,313],{},"This is arithmetic from a short synthetic test, ",[13,311,312],{},"not a measured 1,000-user deployment or capacity guarantee",". It excludes burst headroom, high availability and whole-system costs. The English-word conversion is approximate, not a property of every prompt or language.",[62,315,317],{"id":316},"how-much-of-the-original-model-does-a5-retain","How much of the original model does A5 retain?",[10,319,320,321,326],{},"We checked the quantized weights against the original BF16 checkpoint, emulating the arithmetic used during serving. The ",[43,322,325],{"href":323,"rel":324},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W4A8-Paiton-CDNA4\u002Fblob\u002Fe949b1c368f99dbfaaa91c7e7d1e109a72b7d97c\u002FREADME.md",[47],"published model card"," reports the full method and results.",[328,329],"mi350p-figure",{"kind":330},"weight-accuracy",[10,332,333],{},[231,334,335],{},"Held-out weight evaluation against BF16. Full-vocabulary KL is lower-is-closer; top-1 measures next-token agreement, not task accuracy. The INT4-g128 comparison uses our RDNA4 recipe in the same evaluation tool. This is separate from the 130-position engine check later in the article.",[126,337,338,348],{},[129,339,340],{},[132,341,342,345],{},[135,343,344],{},"Weight-quality check",[135,346,347],{"align":140},"A5 result",[149,349,350,360,370],{},[132,351,352,355],{},[154,353,354],{},"Same most-likely next token as BF16",[154,356,357],{"align":140},[13,358,359],{},"93.78%",[132,361,362,365],{},[154,363,364],{},"Full-vocabulary KL divergence, lower is closer",[154,366,367],{"align":140},[13,368,369],{},"0.0265",[132,371,372,375],{},[154,373,374],{},"95% interval for KL",[154,376,377],{"align":140},"0.0239 to 0.0292",[10,379,380,381,384],{},"These scores cover text and assistant positions from a held-out ",[13,382,383],{},"294,912-token evaluation set",". Top-1 agreement means both models select the same next token at that position. It does not mean 93.78% of answers are correct. KL compares their full next-token probability distributions; zero means identical. Two correct BF16 implementations score about 0.008 in this method.",[10,386,387,388,391,392,257,395,398,399,402],{},"At longer contexts, negative log-likelihood stays ",[13,389,390],{},"within +0.2% of BF16"," across the 8K to 32K document buckets. Needle tests reproduce the reference answers for ",[13,393,394],{},"8 of 8 prompts at 131K",[13,396,397],{},"4 of 4 at 258K",". A continuation after a 254K-token user turn is ",[13,400,401],{},"1.9% higher in negative log-likelihood"," than BF16.",[10,404,405,406,24],{},"These are weight-quality and retrieval checks, not proof of production serving at those lengths. Math, coding, knowledge, multilingual and tool-calling task quality were ",[13,407,408],{},"not benchmarked for this release",[62,410,412],{"id":411},"what-is-paiton","What is Paiton?",[10,414,415,418],{},[13,416,417],{},"Paiton is ElioVP's framework for making AI models run more efficiently on AMD GPUs."," It brings together model compilation, hardware-specific kernels, runtime optimization and integration with serving software such as vLLM. The native runtime used in this benchmark is one part of that framework, not the whole product.",[10,420,421,422,427,428,432],{},"AMD described this approach in ",[43,423,426],{"href":424,"rel":425},"https:\u002F\u002Fwww.amd.com\u002Fen\u002Fblogs\u002F2025\u002Fintroduction-to-paiton-by-eliovp.html",[47],"Introduction to Paiton by ElioVP",", published in October 2025. Our ",[43,429,431],{"href":430},"\u002Fproducts\u002Fpaiton","Paiton product page"," covers the current offering, including language models, image generation and video.",[434,435,437],"h3",{"id":436},"what-we-changed-for-this-cdna4-release","What we changed for this CDNA4 release",[10,439,440],{},"The release combines our quantization, native CDNA4 kernels and a vLLM server.",[10,442,443,444,447,448,24],{},"First, ",[13,445,446],{},"our own A5 quantization"," of Qwen3.8-Flash-Next. The routed experts, which hold most of the model's weights, use native FP4 codes with one scale per 128 weights. The main non-expert weights use 8-bit formats. The recipe uses roughly four million calibration tokens and is documented in the ",[43,449,451],{"href":45,"rel":450},[47],"public model card",[10,453,454,455,458],{},"Second, ",[13,456,457],{},"Paiton's native CDNA4 runtime",". It executes the forward pass and sampling with kernels designed for this hardware, including fused paths for decode, prefill and expert computation. Asynchronous scheduling overlaps host preparation with GPU execution.",[10,460,461,462,465],{},"Third, ",[13,463,464],{},"the regular vLLM serving interface",". The Paiton plugin connects the runtime to vLLM's OpenAI-compatible server. Existing clients can keep using that API.",[10,467,468],{},"Speculative decoding helps at low concurrency. A fine-tuned MTP drafter proposes three tokens, and the main model checks them together. Above four running requests, this release turns drafting off because its verification overhead outweighs the saving. That tradeoff becomes visible in the results below.",[10,470,471,472,475],{},"“One GPU” does not mean every tensor lives in GPU memory. Approximately ",[13,473,474],{},"48 GiB of n-gram embeddings live in host RAM",", with the required rows gathered during execution. This, alongside the smaller expert weights, leaves more GPU memory for serving state.",[62,477,479],{"id":478},"faster-decode-across-the-tested-range","Faster decode across the tested range",[10,481,482,483,487],{},"LocalMaxxing's ",[484,485,486],"code",{},"lmx"," speed test measures generation from an empty context, separately from prompt processing. Paiton leads every tested concurrency level in this benchmark, from one request to 64.",[328,489],{"kind":490},"decode",[10,492,493],{},[231,494,495,496,24],{},"Aggregate output tokens\u002Fs; higher is better. One MI350P at a configured 450 W limit. Paiton uses A5; the other stacks use AMD Quark MXFP4. ",[43,497,498],{"href":236},"Results and provenance",[126,500,501,519],{},[129,502,503],{},[132,504,505,507,509,511,514,516],{},[135,506,137],{},[135,508,283],{"align":140},[135,510,292],{"align":140},[135,512,513],{"align":140},"SGLang MTP, 8 slots",[135,515,301],{"align":140},[135,517,518],{"align":140},"SGLang, no MTP",[149,520,521,542,563,583,604,624,644],{},[132,522,523,525,530,533,536,539],{},[154,524,156],{},[154,526,527],{"align":140},[13,528,529],{},"273.1",[154,531,532],{"align":140},"122.1",[154,534,535],{"align":140},"118.6",[154,537,538],{"align":140},"119.6",[154,540,541],{"align":140},"Not run",[132,543,544,547,552,555,558,561],{},[154,545,546],{},"2",[154,548,549],{"align":140},[13,550,551],{},"453.4",[154,553,554],{"align":140},"195.4",[154,556,557],{"align":140},"223.1",[154,559,560],{"align":140},"219.0",[154,562,541],{"align":140},[132,564,565,567,572,575,578,581],{},[154,566,170],{},[154,568,569],{"align":140},[13,570,571],{},"676.5",[154,573,574],{"align":140},"384.7",[154,576,577],{"align":140},"397.0",[154,579,580],{"align":140},"403.5",[154,582,541],{"align":140},[132,584,585,588,593,596,599,602],{},[154,586,587],{},"8",[154,589,590],{"align":140},[13,591,592],{},"839.6",[154,594,595],{"align":140},"602.8",[154,597,598],{"align":140},"639.2",[154,600,601],{"align":140},"629.9",[154,603,541],{"align":140},[132,605,606,608,613,616,619,622],{},[154,607,184],{},[154,609,610],{"align":140},[13,611,612],{},"1,265.8",[154,614,615],{"align":140},"1,029.0",[154,617,618],{"align":140},"634.4",[154,620,621],{"align":140},"829.3",[154,623,541],{"align":140},[132,625,626,628,633,636,639,641],{},[154,627,198],{},[154,629,630],{"align":140},[13,631,632],{},"1,886.7",[154,634,635],{"align":140},"1,544.1",[154,637,638],{"align":140},"Not measured",[154,640,638],{"align":140},[154,642,643],{"align":140},"1,088.6",[132,645,646,648,653,656,658,660],{},[154,647,212],{},[154,649,650],{"align":140},[13,651,652],{},"2,508.6",[154,654,655],{"align":140},"2,185.2",[154,657,638],{"align":140},[154,659,638],{"align":140},[154,661,662],{"align":140},"1,093.0",[10,664,665],{},"SGLang uses version 0.5.22 with AITER; vLLM uses 0.32.0 with AITER. The SGLang MTP profiles did not accommodate a configuration for 32 or more running requests in our setup, so those measurements use a separately labelled non-MTP profile. These are tested configurations, not universal limits for either engine.",[10,667,668,669,671,672,675,676,679],{},"At one request, the advantage over the fastest alternative is ",[13,670,79],{},". At 16 it is ",[13,673,674],{},"23%","; at 64, ",[13,677,678],{},"15%",". The gain narrows as batching makes the baseline more efficient, but remains present across this matrix.",[62,681,683],{"id":682},"reading-a-long-prompt-sooner","Reading a long prompt sooner",[10,685,686],{},"Prefill is the work before generation: reading the input and building the state needed to answer. Here, Paiton is almost twice as fast as the selected stock vLLM baseline. SGLang is much closer, especially on the longer prompt.",[328,688],{"kind":689},"prefill",[10,691,692],{},[231,693,694],{},"Single-prompt input tokens\u002Fs at nominal 8K and 64K depths; higher is better. Actual tokenized lengths are approximately 8,200 and 64,100 tokens. These are separate from decode throughput.",[126,696,697,709],{},[129,698,699],{},[132,700,701,703,706],{},[135,702,270],{},[135,704,705],{"align":140},"Prefill 8K, tok\u002Fs",[135,707,708],{"align":140},"Prefill 64K, tok\u002Fs",[149,710,711,727,737,747,757],{},[132,712,713,717,722],{},[154,714,715],{},[13,716,283],{},[154,718,719],{"align":140},[13,720,721],{},"13,709",[154,723,724],{"align":140},[13,725,726],{},"14,286",[132,728,729,731,734],{},[154,730,513],{},[154,732,733],{"align":140},"12,069",[154,735,736],{"align":140},"11,585",[132,738,739,741,744],{},[154,740,301],{},[154,742,743],{"align":140},"11,535",[154,745,746],{"align":140},"13,827",[132,748,749,751,754],{},[154,750,518],{},[154,752,753],{"align":140},"12,184",[154,755,756],{"align":140},"13,401",[132,758,759,761,764],{},[154,760,292],{},[154,762,763],{"align":140},"7,009",[154,765,766],{"align":140},"7,224",[10,768,769,770,773,774,777,778,781],{},"For the approximately 64K-token prompt, measured time to first token is ",[13,771,772],{},"4.49 seconds with Paiton",", ",[13,775,776],{},"4.64 seconds with SGLang MTP16",", and ",[13,779,780],{},"8.88 seconds with vLLM",". Against the best alternative, the prefill advantage is 13% at 8K and 3% at 64K. The larger, roughly 2× gain is specifically against the selected vLLM baseline.",[62,783,785],{"id":784},"about-22-million-kv-cache-tokens-on-one-gpu","About 2.2 million KV-cache tokens on one GPU",[10,787,788],{},"The KV cache is working memory for the tokens a model is processing. More capacity can accommodate longer requests or more requests, subject to the engine's other memory and scheduling requirements.",[328,790],{"kind":791},"cache",[10,793,794],{},[231,795,796],{},"Runtime-reported GPU KV-cache capacity, not additional physical memory or measured simultaneous long-conversation capacity.",[126,798,799,811],{},[129,800,801],{},[132,802,803,805,808],{},[135,804,270],{},[135,806,807],{"align":140},"KV-cache tokens",[135,809,810],{"align":140},"32K blocks, memory-only estimate",[149,812,813,829,839,849],{},[132,814,815,819,824],{},[154,816,817],{},[13,818,283],{},[154,820,821],{"align":140},[13,822,823],{},"2,190,540",[154,825,826],{"align":140},[13,827,828],{},"66",[132,830,831,833,836],{},[154,832,292],{},[154,834,835],{"align":140},"1,795,019",[154,837,838],{"align":140},"54",[132,840,841,843,846],{},[154,842,513],{},[154,844,845],{"align":140},"367,424",[154,847,848],{"align":140},"11",[132,850,851,853,856],{},[154,852,301],{},[154,854,855],{"align":140},"64,704",[154,857,156],{"align":140},[10,859,860,861,864,865,868],{},"Paiton reports ",[13,862,863],{},"22% more token capacity than vLLM"," and approximately ",[13,866,867],{},"6× the default eight-slot SGLang MTP configuration",". Increasing SGLang's running-request allowance reserves more per-request state, leaving less space for the KV cache.",[10,870,871,872,875],{},"Here, ",[13,873,874],{},"32K means 32,768 tokens",". The last column simply divides capacity by that number and rounds down. It does not establish that 66 conversations can run simultaneously: Paiton is configured for 64 running sequences, generated tokens need room, and other per-request state also matters. We did not benchmark 64 simultaneous 32K prompts.",[10,877,878,879,882],{},"All stacks were configured with a ",[13,880,881],{},"262,144-token context limit",". A configured limit is not evidence that every configuration can serve a request that large. In particular, SGLang's 16-slot MTP cache is smaller than that limit.",[62,884,886],{"id":885},"load-changes-the-result","Load changes the result",[10,888,889,890,773,893,777,896,899],{},"The serving sweep adds input processing and longer outputs. We tested three shapes: ",[13,891,892],{},"1K \u002F 1K",[13,894,895],{},"1K \u002F 8K",[13,897,898],{},"8K \u002F 1K"," input\u002Foutput tokens, at one to 64 concurrent requests. Here K means 1,024 tokens.",[328,901],{"kind":902},"serving",[10,904,905],{},[231,906,907,908,24],{},"Aggregate output versus per-request decode in the serving sweep. These are different measurements from LocalMaxxing. Full latency, throughput and power records are included in the ",[43,909,911],{"href":910},"#inspect-the-measurements","downloads",[10,913,914,915,918,919,257,922,925,926,121,929,24],{},"For one request producing an 8K-token answer, Paiton's median completion time is ",[13,916,917],{},"22.6 seconds",", versus ",[13,920,921],{},"43.1 seconds for SGLang",[13,923,924],{},"67.0 seconds for vLLM",". At 64 concurrent requests in that shape, total output is ",[13,927,928],{},"1,937 tok\u002Fs",[13,930,931],{},"1,737 tok\u002Fs for vLLM",[10,933,934,935,938,939,942],{},"Long inputs put more pressure on scheduling. At 64 requests with 8K input \u002F 1K output, Paiton's median first-token time rises to ",[13,936,937],{},"9.6 seconds",". That is lower than vLLM's ",[13,940,941],{},"37.6 seconds",", but it is still a substantial wait. A high aggregate score does not erase that latency.",[62,944,946],{"id":945},"energy-more-output-per-gpu-watt","Energy: more output per GPU watt",[10,948,949,950,953,954,957],{},"A power limit is not measured consumption. During the 1K \u002F 1K sweep, Paiton's mean GPU reading rises from approximately ",[13,951,952],{},"308 W"," at one request to ",[13,955,956],{},"445 W"," at 64. Combining that reading with throughput gives the following efficiency result.",[328,959],{"kind":960},"energy",[10,962,963],{},[231,964,965],{},"Output tokens\u002Fs per kW of measured GPU power; higher is better. Excludes CPU, system RAM, power-supply losses and cooling.",[126,967,968,983],{},[129,969,970],{},[132,971,972,974,977,980],{},[135,973,270],{},[135,975,976],{"align":140},"1 request",[135,978,979],{"align":140},"16 requests",[135,981,982],{"align":140},"64 requests",[149,984,985,1004,1017,1031],{},[132,986,987,991,996,999],{},[154,988,989],{},[13,990,283],{},[154,992,993],{"align":140},[13,994,995],{},"1,076",[154,997,998],{"align":140},"2,577",[154,1000,1001],{"align":140},[13,1002,1003],{},"4,021",[132,1005,1006,1008,1011,1014],{},[154,1007,292],{},[154,1009,1010],{"align":140},"413",[154,1012,1013],{"align":140},"2,120",[154,1015,1016],{"align":140},"3,354",[132,1018,1019,1021,1024,1029],{},[154,1020,301],{},[154,1022,1023],{"align":140},"645",[154,1025,1026],{"align":140},[13,1027,1028],{},"2,720",[154,1030,638],{"align":140},[132,1032,1033,1035,1037,1039],{},[154,1034,518],{},[154,1036,638],{"align":140},[154,1038,638],{"align":140},[154,1040,1041],{"align":140},"2,493",[10,1043,1044,1045,1048,1049,1052,1053,1056],{},"Paiton delivers ",[13,1046,1047],{},"2.60× the output per unit of GPU energy"," of the vLLM baseline at one request and ",[13,1050,1051],{},"20% more"," at 64. SGLang's 16-slot profile is ",[13,1054,1055],{},"5.6% more efficient"," than Paiton at 16 requests. That exception matters when choosing a target operating load.",[10,1058,1059],{},"These are GPU-only figures from AMD SMI, not wall-plug efficiency or a whole-server cost calculation. Host energy is particularly relevant here because part of the model resides in system memory.",[62,1061,1063],{"id":1062},"a-separate-accuracy-check-of-the-engines","A separate accuracy check of the engines",[10,1065,1066,1067,1070],{},"The weight evaluation at the start of this article tests the quantization. We also checked the execution paths, comparing next-token distributions at ",[13,1068,1069],{},"130 teacher-forced positions following 8,128-token prompts",". This uses a different metric: the reference's top 1,024 tokens plus the remaining probability mass as one tail term.",[328,1072],{"kind":1073},"accuracy",[10,1075,1076],{},[231,1077,1078],{},"Narrow next-token diagnostic, not answer accuracy. Lower KL is closer to BF16. Paiton and AMD MXFP4 run on MI350P; the BF16 control runs on MI355X. The harness paths differ.",[126,1080,1081,1093],{},[129,1082,1083],{},[132,1084,1085,1087,1090],{},[135,1086,270],{},[135,1088,1089],{"align":140},"KL divergence",[135,1091,1092],{"align":140},"Top-1 token agreement",[149,1094,1095,1106,1123],{},[132,1096,1097,1100,1103],{},[154,1098,1099],{},"BF16 control via vLLM, MI355X",[154,1101,1102],{"align":140},"0.0077",[154,1104,1105],{"align":140},"95.4%",[132,1107,1108,1113,1118],{},[154,1109,1110],{},[13,1111,1112],{},"Paiton A5 engine test path, MI350P",[154,1114,1115],{"align":140},[13,1116,1117],{},"0.0219",[154,1119,1120],{"align":140},[13,1121,1122],{},"93.1%",[132,1124,1125,1128,1131],{},[154,1126,1127],{},"AMD Quark MXFP4 via vLLM + AITER, MI350P",[154,1129,1130],{"align":140},"0.1141",[154,1132,1133],{"align":140},"80.8%",[10,1135,1136,1137,1140],{},"Paiton is closer to the reference on this small sample. It does ",[13,1138,1139],{},"not"," mean that 93.1% of answers are correct, nor that one setup is five times more accurate on user tasks.",[10,1142,1143],{},"Paiton is measured through its engine test path with MTP off; the other rows use a vLLM API client. Short-prompt rows were excluded because the BF16 control showed inconsistencies there. The comparison also combines checkpoint and kernel differences. Published logs support these selected rows, but the bundle does not contain the full reference distributions or evaluation client needed to rerun this diagnostic independently.",[62,1145,1147],{"id":1146},"from-rdna4-on-radeon-to-cdna4-on-instinct","From RDNA4 on Radeon to CDNA4 on Instinct",[10,1149,1150,1151,1155,1156,257,1159,1162],{},"This work builds on our Radeon releases. With ",[43,1152,1154],{"href":1153},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","3-bit Qwen3.8 27B on one Radeon AI PRO R9700",", Paiton reached ",[13,1157,1158],{},"492.1 output tokens\u002Fs across eight concurrent requests",[13,1160,1161],{},"19.9% faster weighted decode"," than the preceding MXFP4 release. That article covers the settings and accuracy tradeoffs of the comparison.",[10,1164,1165,1166,1170,1171,1174,1175,1178],{},"Our ",[43,1167,1169],{"href":1168},"\u002Fblog\u002Fpaiton-qwen38-3bit-accuracy-prefix-caching-mtp","later Flash Next update on two R9700s"," improved 3-bit accuracy and added prefix caching with MTP. Its optional native worker reached ",[13,1172,1173],{},"258.7 tokens\u002Fs weighted single-request decode",". In a separate test, a repeated 200K-token conversation started responding in ",[13,1176,1177],{},"0.68 seconds on a cache hit, instead of 25.81 seconds cold",". That requires the unchanged conversation prefix to remain cached.",[10,1180,1181,1182,257,1187,1192,1193,1196],{},"The ",[43,1183,1186],{"href":1184,"rel":1185},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-27B-W3Rot-INT3-Paiton-RDNA4",[47],"27B W3Rot weights",[43,1188,1191],{"href":1189,"rel":1190},"https:\u002F\u002Fhuggingface.co\u002FEliovpAI\u002FQwen3.8-Flash-Next-W3A8-Paiton-RDNA4",[47],"Flash Next W3A8 weights"," recorded ",[13,1194,1195],{},"more than 2,200 combined Hugging Face downloads in the preceding 30 days",", checked on 30 October 2026. That counts downloads, not unique users.",[10,1198,1199],{},"The aim is the same across these releases: get more from AMD hardware by optimizing the model and its execution together. The R9700 and MI350P figures come from separate tests with different models and settings; this is not a direct comparison between the cards.",[62,1201,1203],{"id":1202},"a-look-at-the-benchmark-output","A look at the benchmark output",[10,1205,1206],{},[1207,1208],"img",{"alt":1209,"src":1210},"Unaltered terminal output from a separate Paiton run on one MI350P: 262.6 tokens\u002Fs at one request, 1,274.3 at sixteen, and 14,243 input tokens\u002Fs for prefill.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Flive-benchmark-poster.webp",[10,1212,1213],{},[231,1214,1215,1216,24],{},"Final frame of the original terminal capture, unchanged. Click to enlarge. This separate run uses 20-second cells, not the headline benchmark's 30-second cells. The source recording and timing logs remain in the ",[43,1217,1219],{"href":1218},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fproof\u002Faccuracy-recording.tar.gz","accuracy and recording download",[62,1221,1223],{"id":1222},"inspect-the-measurements","Inspect the measurements",[10,1225,1226,1227,1230],{},"The evidence is split into manageable downloads, with raw matrices, per-request records, commands and checksums. The supplied SHA-256 manifest's ",[13,1228,1229],{},"177 entries were checked successfully"," before publication. That verifies file integrity, not independent experimental validation.",[10,1232,1233,1236,1237,1240],{},[13,1234,1235],{},"Why these results are not on LocalMaxxing:"," we would like to publish the numbers there as well, but its GPU detection did not recognize the MI350P in this run, which prevented publication. The report's ",[484,1238,1239],{},"CPU_ONLY"," label is a detection limitation, not how the benchmark ran: the model ran on the MI350P. We are sharing the raw results here so readers can inspect them directly.",[1242,1243,1244,1251,1257,1263,1268,1274],"ul",{},[1245,1246,1247],"li",{},[43,1248,1250],{"href":1249},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fproof\u002Fpaiton.tar.gz","Paiton benchmark evidence",[1245,1252,1253],{},[43,1254,1256],{"href":1255},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fproof\u002Fvllm.tar.gz","vLLM benchmark evidence",[1245,1258,1259],{},[43,1260,1262],{"href":1261},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fproof\u002Fsglang.tar.gz","SGLang benchmark evidence",[1245,1264,1265],{},[43,1266,1267],{"href":1218},"Accuracy logs and terminal recording evidence",[1245,1269,1270],{},[43,1271,1273],{"href":1272},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fproof\u002Fmethods.tar.gz","Methods, commands, serving client and checksum manifest",[1245,1275,1276],{},[43,1277,1278],{"href":236},"Chart data and selected source paths",[434,1280,1282],{"id":1281},"download-the-stock-baseline-container","Download the stock baseline container",[10,1284,1181,1285,1290,1291,1294,1295,1298],{},[43,1286,1289],{"href":1287,"rel":1288},"https:\u002F\u002Fgithub.com\u002Fusers\u002FEliovp\u002Fpackages\u002Fcontainer\u002Fpaiton-vllm-plugin\u002F1472686429?tag=cdna4-mxfp4-default-20261028",[47],"public CDNA4 container"," packages ",[13,1292,1293],{},"stock vLLM 0.32 + AITER",", configured to serve ",[13,1296,1297],{},"AMD's Quark MXFP4 checkpoint",". Pull the published image by digest:",[1300,1301,1306],"pre",{"className":1302,"code":1303,"language":1304,"meta":1305,"style":1305},"language-bash shiki shiki-themes github-light github-dark","docker pull ghcr.io\u002Feliovp\u002Fpaiton-vllm-plugin@sha256:20b729e6c7b0f2de101fd27d080d91291a7d37c4002312fe9ee5aeb5ff131b19\n","bash","",[484,1307,1308],{"__ignoreMap":1305},[1309,1310,1313,1317,1321],"span",{"class":1311,"line":1312},"line",1,[1309,1314,1316],{"class":1315},"sScJk","docker",[1309,1318,1320],{"class":1319},"sZZnC"," pull",[1309,1322,1323],{"class":1319}," ghcr.io\u002Feliovp\u002Fpaiton-vllm-plugin@sha256:20b729e6c7b0f2de101fd27d080d91291a7d37c4002312fe9ee5aeb5ff131b19\n",[10,1325,1326,1327,1330,1331],{},"This downloads the baseline container; it does not start a server or supply the proprietary engine. The ",[43,1328,1329],{"href":1272},"methods download"," contains the original benchmark commands. ",[13,1332,1333],{},"This image is not a serving path for the A5 weights and will not reproduce the Paiton column.",[434,1335,1337],{"id":1336},"measurement-settings","Measurement settings",[10,1339,1340,1343],{},[13,1341,1342],{},"SGLang profiles."," The serving sweep uses MTP with 16 slots for 1K \u002F 1K and eight slots for the longer shapes, selected by peak throughput for each shape. At 32 and 64 requests, it uses the non-MTP configuration.",[10,1345,1346,1349],{},[13,1347,1348],{},"Common conditions."," One MI350P, a configured 450 W GPU limit, fans at full speed, one benchmark at a time and the client on the server host.",[10,1351,1352,1355,1356,1358,1359,1362],{},[13,1353,1354],{},"LocalMaxxing."," ",[484,1357,486],{}," 0.1.54 with ",[484,1360,1361],{},"llm-inference-bench"," 0.7.7, 30-second cells, empty-context decode capped at 8,192 output tokens, and nominal 8K\u002F64K prefill. After warmup, the main configurations run three matrices; we select the whole run with median single-request decode. Comparator c32\u002Fc64 measurements are separate single runs, not three-run medians.",[10,1364,1365,1368,1369,1372,1373,1376],{},[13,1366,1367],{},"Serving sweep."," A streaming client uses greedy generation, ignores EOS and gives each request a distinct source-text window. After two short-output warmup rounds, it starts up to ",[484,1370,1371],{},"max(8, 2 × concurrency)"," requests, stopping new starts after 100 seconds and allowing in-flight requests to finish. Cells contain ",[13,1374,1375],{},"2 to 128 measured requests",", all with zero errors and no short outputs. Small-sample p99 values are close to the observed maximum, not reliable population percentiles.",[10,1378,1379,1382,1383,1386,1387,1390],{},[13,1380,1381],{},"Selection detail."," The displayed vLLM 1K \u002F 1K c32 result uses the earlier recorded cell, ",[13,1384,1385],{},"1,169.2 tok\u002Fs","; a later repeat produced ",[13,1388,1389],{},"1,064.0 tok\u002Fs",". Both are in the download. Other c32\u002Fc64 sweep rows use the later batch. The source paths in the chart data make that choice explicit.",[10,1392,1393],{},"The MI350P results show what the Paiton framework delivers on this CDNA4 setup: its largest speed advantage is at low concurrency, with gains continuing under heavier load. Our next optimization task is to make speculative decoding pay off above four running requests.",[10,1395,1396,1397,1401,1402,24],{},"Read the ",[43,1398,1400],{"href":45,"rel":1399},[47],"public weights and quantization recipe"," or learn more about ",[43,1403,283],{"href":430},[10,1405,1406],{},[231,1407,1408],{},"Cover: AI-generated editorial illustration of an MI350P, not a product photograph or benchmark capture.",[62,1410,1412],{"id":1411},"running-a-token-factory-get-more-from-your-amd-gpus","Running a token factory? Get more from your AMD GPUs.",[10,1414,1415],{},"Already serving models, or interested in starting a token factory? We're here to help you increase throughput, lower your cost per token and maximize the return on your AMD hardware.",[10,1417,1418,24],{},[43,1419,1421],{"href":1420},"\u002Fcontact","Let's talk about your token factory",[1423,1424,1425],"style",{},"html pre.shiki code .sScJk, html code.shiki .sScJk{--shiki-default:#6F42C1;--shiki-dark:#B392F0}html pre.shiki code .sZZnC, html code.shiki .sZZnC{--shiki-default:#032F62;--shiki-dark:#9ECBFF}html .default .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .shiki span {color: var(--shiki-default);background: var(--shiki-default-bg);font-style: var(--shiki-default-font-style);font-weight: var(--shiki-default-font-weight);text-decoration: var(--shiki-default-text-decoration);}html .dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}html.dark .shiki span {color: var(--shiki-dark);background: var(--shiki-dark-bg);font-style: var(--shiki-dark-font-style);font-weight: var(--shiki-dark-font-weight);text-decoration: var(--shiki-dark-text-decoration);}",{"title":1305,"searchDepth":1427,"depth":1427,"links":1428},2,[1429,1430,1431,1435,1436,1437,1438,1439,1440,1441,1442,1443,1447],{"id":64,"depth":1427,"text":65},{"id":316,"depth":1427,"text":317},{"id":411,"depth":1427,"text":412,"children":1432},[1433],{"id":436,"depth":1434,"text":437},3,{"id":478,"depth":1427,"text":479},{"id":682,"depth":1427,"text":683},{"id":784,"depth":1427,"text":785},{"id":885,"depth":1427,"text":886},{"id":945,"depth":1427,"text":946},{"id":1062,"depth":1427,"text":1063},{"id":1146,"depth":1427,"text":1147},{"id":1202,"depth":1427,"text":1203},{"id":1222,"depth":1427,"text":1223,"children":1444},[1445,1446],{"id":1281,"depth":1434,"text":1282},{"id":1336,"depth":1434,"text":1337},{"id":1411,"depth":1427,"text":1412},[283,1449,1450,1451,1452],"AMD Instinct","AI Inference","Qwen","Quantization","2026-10-30T12:00:00Z","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.","md","Qwen3.8 Flash Next on an AMD MI350P: twice the speed","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fhero.webp",{},true,"https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p","\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p",{"title":5,"description":1454},"paiton-qwen38-flash-next-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-mi350p\u002Fsocial.webp","blog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p",null,"7c96FNh8gJ37RIj7cm8TXFzOwt9KH73Hticiv33yIHc",[1469,1471,1481,1496,1505,1514,1522,1532,1545,1555,1567,1576,1590,1599,1611,1643,1654,1676,1694,1713,1731,1748,1765,1776,1792,1807,1819,1828,1836,1851,1863,1874,1885,1895,1908,1918,1931,1942,1952,1963,1972,1984,1995,2004],{"path":1461,"title":5,"description":1454,"date":1453,"slug":1463,"image":1457,"originalUrl":1460,"categories":1470},[283,1449,1450,1451,1452],{"path":1168,"title":1472,"description":1473,"date":1474,"slug":1475,"image":1476,"originalUrl":1477,"categories":1478},"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",[283,1479,1480,1451,1452],"AMD Radeon","Local AI",{"path":1482,"title":1483,"description":1484,"date":1485,"slug":1486,"image":1487,"originalUrl":1488,"categories":1489},"\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",[1490,1491,1492,1493,1494,1495],"Modular DC","AI Infrastructure","ModFlex","NVIDIA","Vera Rubin","Liquid Cooling",{"path":1497,"title":1498,"description":1499,"date":1500,"slug":1501,"image":1502,"originalUrl":1503,"categories":1504},"\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",[283,1479,1480,1451,1452],{"path":1506,"title":1507,"description":1508,"date":1509,"slug":1510,"image":1511,"originalUrl":1512,"categories":1513},"\u002Fblog\u002Fpaiton-qwen38-flash-next-radeon-ai-pro-r9700","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",[283,1479,1480,1451,1452],{"path":1153,"title":1515,"description":1516,"date":1517,"slug":1518,"image":1519,"originalUrl":1520,"categories":1521},"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",[283,1479,1480,1451,1452],{"path":1523,"title":1524,"description":1525,"date":1526,"slug":1527,"image":1528,"originalUrl":1529,"categories":1530},"\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",[283,1479,1480,1531,1451],"Image Generation",{"path":1533,"title":1534,"description":1535,"date":1536,"slug":1537,"image":1538,"originalUrl":1539,"categories":1540},"\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",[283,1479,1541,1542,1543,1544],"vLLM","Qwen3.8","Inference Optimization","DFlash2",{"path":1546,"title":1547,"description":1548,"date":1549,"slug":1550,"image":1551,"originalUrl":1552,"categories":1553},"\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",[283,1479,1480,1554,1541],"GGUF",{"path":1556,"title":1557,"description":1558,"date":1559,"slug":1560,"image":1561,"originalUrl":1562,"categories":1563},"\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",[283,1479,1480,1564,1565,1566],"Video Generation","MiniMax H3","ComfyUI",{"path":1568,"title":1569,"description":1570,"date":1571,"slug":1572,"image":1573,"originalUrl":1568,"categories":1574},"\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",[283,1479,1480,1531,1575,1566],"FLUX",{"path":1577,"title":1578,"description":1579,"date":1580,"slug":1581,"image":1582,"originalUrl":1583,"categories":1584},"\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",[283,1585,1479,1450,1586,1587,1543,1588,1541,1589],"Artificial Intelligence","GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1591,"title":1592,"description":1593,"date":1594,"slug":1595,"image":1596,"originalUrl":1597,"categories":1598},"\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",[283,1585,1479,1450,1586,1587,1543,1588,1541,1589],{"path":1600,"title":1601,"description":1602,"date":1603,"slug":1604,"image":1605,"originalUrl":1466,"categories":1606},"\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",[1607,1491,1608,1492,1609,1610],"All","Data Centers","HPC","AMD Helios",{"path":1612,"title":1613,"description":1614,"date":1615,"slug":1616,"image":1617,"originalUrl":1618,"categories":1619},"\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",[1607,1585,283,1620,1621,1622,1623,1624,1625,1626,1627,1628,1629,1630,1631,1632,1633,1634,1635,1636,283,1637,1638,1639,1640,1641,1642],"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":1644,"title":1645,"description":1646,"date":1647,"slug":1648,"image":1649,"originalUrl":1650,"categories":1651},"\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",[1607,1585,1490,1652,1621,1653,1490,1634],"Uncategorized","De Tijd",{"path":1655,"title":1656,"description":1657,"date":1658,"slug":1659,"image":1660,"originalUrl":1661,"categories":1662},"\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",[1607,1585,1663,1652,1664,1665,1666,1667,1668,1669,1670,1671,1627,1672,1628,1673,1674,1675],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1677,"title":1678,"description":1679,"date":1680,"slug":1681,"image":1682,"originalUrl":1683,"categories":1684},"\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",[1607,1685,1686,1687,1669,1688,1689,1690,1672,1691,1692,1693,1674],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1695,"title":1696,"description":1697,"date":1698,"slug":1699,"image":1700,"originalUrl":1701,"categories":1702},"\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",[1607,1585,1703,1663,1704,1705,1706,1707,1708,1709,1710,1711,1637,1712],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1714,"title":1715,"description":1716,"date":1717,"slug":1718,"image":1719,"originalUrl":1720,"categories":1721},"\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",[1607,1585,1663,1491,1722,1723,1724,1725,1726,1727,1728,1729,1730],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1732,"title":1733,"description":1734,"date":1735,"slug":1736,"image":1737,"originalUrl":1738,"categories":1739},"\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",[1607,1490,1652,1740,1491,1741,1742,1495,1743,1744,1745,1746,1747],"150kW Rack","DLC","High Density","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1749,"title":1750,"description":1751,"date":1752,"slug":1753,"image":1754,"originalUrl":1755,"categories":1756},"\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",[1607,1585,283,1652,1757,1585,1758,1759,1760,1761,1762,1763,283,1764],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1766,"title":1767,"description":1768,"date":1769,"slug":1770,"image":1771,"originalUrl":1772,"categories":1773},"\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",[1607,1585,283,1450,1449,1757,1589,1774,1543,1763,283,1775,1541],"High Throughput","SGLang",{"path":1777,"title":1778,"description":1779,"date":1780,"slug":1781,"image":1782,"originalUrl":1783,"categories":1784},"\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",[1607,1585,283,1785,1757,1786,1543,1787,1788,1789,1790,283,1791],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":1793,"title":1794,"description":1795,"date":1796,"slug":1797,"image":1798,"originalUrl":1799,"categories":1800},"\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",[1607,1585,1703,1652,1704,1801,1802,1803,1804,1709,1711,1637,1805,1806],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":1808,"title":1809,"description":1810,"date":1811,"slug":1812,"image":1813,"originalUrl":1814,"categories":1815},"\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",[1607,1585,283,1652,1816,1621,1622,1817,1818,1633,1634,283,1541],"AI","H200","MI300X",{"path":1820,"title":1821,"description":1822,"date":1823,"slug":1824,"image":1825,"originalUrl":1826,"categories":1827},"\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",[1607,1585,1703,1704,1801,1802,1803,1804,1709,1711,1637,1805,1806],{"path":1829,"title":1830,"description":1831,"date":1832,"slug":1833,"image":1305,"originalUrl":1834,"categories":1835},"\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",[1607,1585,283],{"path":1837,"title":1838,"description":1839,"date":1840,"slug":1841,"image":1842,"originalUrl":1843,"categories":1844},"\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",[1607,1585,283,1652,1450,1757,1845,1759,1846,1847,1848,283,1849,1850],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":1852,"title":1853,"description":1854,"date":1855,"slug":1856,"image":1857,"originalUrl":1858,"categories":1859},"\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",[1607,1585,283,1652,1757,1860,1708,1628,1586,1587,1588,1847,1861,1862],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":1864,"title":1865,"description":1866,"date":1867,"slug":1868,"image":1869,"originalUrl":1870,"categories":1871},"\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",[1607,1585,283,1703,1652,1621,1818,1634,1872,1873],"RX7900XTX","tenstorrent",{"path":1875,"title":1876,"description":1877,"date":1878,"slug":1879,"image":1880,"originalUrl":1881,"categories":1882},"\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",[1607,1585,1490,1703,1622,1817,1883,1634,1884],"MI325x","pnl calculator",{"path":1886,"title":1887,"description":1888,"date":1889,"slug":1890,"image":1891,"originalUrl":1892,"categories":1893},"\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",[1607,1585,283,1816,1621,1817,1894,1634,283,1541],"MI300",{"path":1896,"title":1897,"description":1898,"date":1899,"slug":1900,"image":1901,"originalUrl":1902,"categories":1903},"\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",[1607,1490,1904,1491,1742,1609,1495,1743,1905,1906,1746,1907],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":1909,"title":1910,"description":1911,"date":1912,"slug":1913,"image":1914,"originalUrl":1915,"categories":1916},"\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",[1607,1585,1703,1652,1816,1621,1917],"Healthcare",{"path":1919,"title":1920,"description":1921,"date":1922,"slug":1923,"image":1924,"originalUrl":1925,"categories":1926},"\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",[1607,1663,1816,1621,1927,1928,1929,1930],"import","Taiwan","Tariffs","Trump",{"path":1932,"title":1933,"description":1934,"date":1935,"slug":1936,"image":1937,"originalUrl":1938,"categories":1939},"\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",[1607,1585,1703,1816,1940,1941],"AI Agents","ERP",{"path":1943,"title":1944,"description":1945,"date":1946,"slug":1947,"image":1948,"originalUrl":1949,"categories":1950},"\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",[1607,1585,1663,1951,1621,1629,1634],"AI news",{"path":1953,"title":1954,"description":1955,"date":1956,"slug":1957,"image":1958,"originalUrl":1959,"categories":1960},"\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",[1607,1585,283,1816,1621,1961,1962,1818,283],"benchmark","LLM",{"path":1964,"title":1965,"description":1966,"date":1967,"slug":1968,"image":1969,"originalUrl":1970,"categories":1971},"\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",[1607,1585,283],{"path":1973,"title":1974,"description":1975,"date":1976,"slug":1977,"image":1978,"originalUrl":1979,"categories":1980},"\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",[1607,1621,1981,1982,1983],"Jim Greene","Podcast","Tech Talk",{"path":1985,"title":1986,"description":1987,"date":1988,"slug":1989,"image":1990,"originalUrl":1991,"categories":1992},"\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",[1607,1585,283,1621,1993,1994,1817,1818,1883,283,1541],"Deepseek","H100",{"path":1996,"title":1997,"description":1998,"date":1999,"slug":2000,"image":2001,"originalUrl":2002,"categories":2003},"\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",[1607,1585,283,1621,1993,1994,1817,1818,1883,283,1541],{"path":2005,"title":2006,"description":2007,"date":2008,"slug":2009,"image":2010,"originalUrl":2011,"categories":2012},"\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",[1607,1585,283,1621,1994,1817,1818,1883,283,1541],1793353529564]