[{"data":1,"prerenderedAt":1981},["ShallowReactive",2],{"blog-post-en-\u002Fblog\u002Fpaiton-kimi-k31-pro-amd-instinct-mi355x":3,"blog-posts-sidebar-en":1395},{"id":4,"title":5,"body":6,"categories":1376,"date":1380,"description":1381,"extension":1382,"heading":1383,"image":1384,"meta":1385,"navigation":241,"originalUrl":1386,"path":1387,"seo":1388,"slug":1389,"socialImage":1390,"stem":1391,"updated":1392,"video":1393,"__hash__":1394},"blog\u002Fblog\u002Fpaiton-kimi-k31-pro-amd-instinct-mi355x.md","Kimi-K3.1-Pro on 6 MI355X GPUs at 800 W with Paiton",{"type":7,"value":8,"toc":1353},"minimark",[9,30,33,36,42,48,51,103,107,122,224,231,236,249,254,265,268,276,281,400,405,408,412,490,496,505,508,512,562,567,570,573,576,579,588,649,655,663,667,670,728,735,739,760,877,884,887,890,893,896,900,903,950,956,964,967,976,979,1029,1032,1035,1038,1041,1044,1047,1100,1105,1114,1119,1127,1130,1133,1142,1202,1207,1210,1220,1233,1236,1243,1258,1265,1311,1316,1329,1341,1345],[10,11,12,19,20,24,25,29],"p",{},[13,14,18],"a",{"href":15,"rel":16},"https:\u002F\u002Fhuggingface.co\u002Fmoonshotai\u002FKimi-K3.1-Pro",[17],"nofollow","Kimi-K3.1-Pro"," has 2.8 trillion parameters. With ",[13,21,23],{"href":22},"\u002Fproducts\u002Fpaiton","Paiton",", our AI optimization framework, it runs on ",[26,27,28],"strong",{},"six AMD Instinct MI355X GPUs capped at 800 W each",". Two GPUs in the eight-GPU node stay unused by the model.",[10,31,32],{},"On 8K-input\u002F1K-output requests at sampling temperature 0 and 800 W per GPU, we measured 5,661 total tokens\u002Fs, or 944 per GPU. AMD publishes 936 per GPU for ATOM on eight MI355X GPUs. Paiton's six GPUs drew an average of 4.18 kW during that workload. A separate test also fit a million-token prompt on the same six GPUs at the same power cap.",[10,34,35],{},"There are limits to the comparison. AMD does not state the power limits or sampling temperatures for its throughput tables. Its eight-GPU nodes have higher total throughput, and vLLM remains ahead per GPU on the 128K-input workload. Throughout this article, total tokens means input plus output, not generated tokens alone.",[10,37,38,41],{},[26,39,40],{},"Update, 12 November 2026:"," We fixed a bug in how Paiton read one attention parameter. It hurt recall past about 100K tokens and could cause repetition in long agent sessions. After the fix, the model recalled 72 of 72 planted codes in tests at 31K, 120K and 228K tokens; before, it recalled 31 of 48 in the 120K and 228K tests. Speed at 800 W is unchanged. Our exact reference and the free community build had the same bug. The accuracy section now uses the corrected reference. The community fix is in the plugin repo; its container image is being rebuilt.",[10,43,44],{},[45,46,47],"em",{},"The header is an illustration. The terminal captures and benchmark recordings below are supplied run evidence.",[10,49,50],{},"In this article",[52,53,54,61,67,73,79,85,91,97],"ul",{},[55,56,57],"li",{},[13,58,60],{"href":59},"#the-setup-and-recorded-run","The setup and recorded run",[55,62,63],{},[13,64,66],{"href":65},"#throughput-and-latency","Throughput and latency",[55,68,69],{},[13,70,72],{"href":71},"#gpu-power-and-cost-scenarios","GPU power and cost scenarios",[55,74,75],{},[13,76,78],{"href":77},"#why-six-gpus-and-800-w","Why six GPUs and 800 W",[55,80,81],{},[13,82,84],{"href":83},"#sampling-and-accuracy","Sampling and accuracy",[55,86,87],{},[13,88,90],{"href":89},"#running-on-four-gpus","Running on four GPUs",[55,92,93],{},[13,94,96],{"href":95},"#try-the-community-build","Try the community build",[55,98,99],{},[13,100,102],{"href":101},"#measurement-details-and-a-second-benchmark","Measurement details and a second benchmark",[104,105,60],"h2",{"id":106},"the-setup-and-recorded-run",[10,108,109,110,115,116,121],{},"The original Paiton speed and power measurements below were recorded on 10-11 November 2026. AMD's ",[13,111,114],{"href":112,"rel":113},"https:\u002F\u002Fwww.amd.com\u002Fen\u002Fdeveloper\u002Fresources\u002Ftechnical-articles\u002F2026\u002Fkimi-k3-1-pro-on-amd-instinct-gpus.html",[17],"4 November article"," reports ATOM and vLLM results; ",[13,117,120],{"href":118,"rel":119},"https:\u002F\u002Fvllm.ai\u002Fblog\u002F2026-11-04-kimi-k31-pro",[17],"vLLM's own article"," reports the vLLM figures. These are separate runs on different nodes, not a controlled software comparison on one machine.",[123,124,125,143],"table",{},[126,127,128],"thead",{},[129,130,131,135,137,140],"tr",{},[132,133,134],"th",{},"Setting",[132,136,23],{},[132,138,139],{},"ATOM, published by AMD",[132,141,142],{},"vLLM, published by AMD",[144,145,146,160,173,185,198,210],"tbody",{},[129,147,148,152,155,158],{},[149,150,151],"td",{},"GPUs",[149,153,154],{},"6 × MI355X",[149,156,157],{},"8 × MI355X",[149,159,157],{},[129,161,162,165,168,171],{},[149,163,164],{},"Power limit per GPU",[149,166,167],{},"800 W",[149,169,170],{},"Not stated",[149,172,170],{},[129,174,175,178,181,183],{},[149,176,177],{},"Combined GPU power limit",[149,179,180],{},"4.8 kW",[149,182,170],{},[149,184,170],{},[129,186,187,190,193,196],{},[149,188,189],{},"Measured GPU draw, 8K\u002F1K, T=0",[149,191,192],{},"4.18 kW, about 696 W\u002FGPU",[149,194,195],{},"Not published",[149,197,195],{},[129,199,200,203,206,208],{},[149,201,202],{},"Sampling temperature",[149,204,205],{},"0 and 1.0, reported separately",[149,207,170],{},[149,209,170],{},[129,211,212,215,218,221],{},[149,213,214],{},"Speculative decoding",[149,216,217],{},"DSpark, Inferact drafter",[149,219,220],{},"DSpark, five drafts",[149,222,223],{},"Throughput-table setting not stated",[10,225,226],{},[227,228],"img",{"alt":229,"src":230},"Terminal capture of the six-GPU 8K\u002F1K run, with amd-smi showing GPUs 2-7 active near their 800 W caps and GPUs 0-1 idle.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fw6-8k1k-midrun-amdsmi.png",[10,232,233],{},[45,234,235],{},"Original mid-run capture, temperature 0. The instantaneous readings near 800 W differ from the run's average draw of about 696 W per active GPU.",[10,237,238],{},[239,240],"video",{"controls":241,"playsInline":241,"preload":242,"width":243,"height":244,"ariaLabel":245,"ariaDescribedBy":246,"poster":230,"src":248},true,"none",1920,1080,"Accelerated recording of the six-GPU Kimi-K3.1-Pro benchmark and GPU power readings",[247],"kimi-six-gpu-caption","\u002Fasset\u002Fvideos\u002Fblog\u002Fpaiton-kimi31\u002Fw6-8k1k-t0-run.mp4",[10,250,251],{"id":247},[45,252,253],{},"Benchmark recording at 4× speed, not real time. Six MI355X GPUs, 800 W cap per GPU, sampling temperature 0.",[10,255,256,259,260,264],{},[13,257,258],{"href":248},"Open the six-GPU recording"," or inspect the ",[13,261,263],{"href":262},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fw6-8k1k-final-result.png","original result screen",".",[104,266,66],{"id":267},"throughput-and-latency",[10,269,270,271,275],{},"We used the stock ",[272,273,274],"code",{},"vllm bench serve"," client with random prompts, 64 requests and at most 16 requests in flight, matching AMD's published workload shape. Prefix caching was off and EOS was ignored. The throughput runs used a maximum sequence length of 140,000 tokens; the million-token test used a separate configuration.",[277,278,280],"h3",{"id":279},"_8k-input-and-1k-output","8K input and 1K output",[123,282,283,308],{},[126,284,285],{},[129,286,287,290,292,296,299,302,305],{},[132,288,289],{},"Runtime and GPU power cap",[132,291,202],{},[132,293,295],{"align":294},"right","Total tok\u002Fs, node",[132,297,298],{"align":294},"Total tok\u002Fs\u002FGPU",[132,300,301],{"align":294},"Output tok\u002Fs, node",[132,303,304],{"align":294},"Median TTFT",[132,306,307],{"align":294},"Median TPOT",[144,309,310,333,355,378],{},[129,311,312,315,318,321,324,327,330],{},[149,313,314],{},"Paiton, 6 GPUs × 800 W",[149,316,317],{},"0",[149,319,320],{"align":294},"5,661",[149,322,323],{"align":294},"944",[149,325,326],{"align":294},"629",[149,328,329],{"align":294},"0.65 s",[149,331,332],{"align":294},"20.4 ms",[129,334,335,337,340,343,346,349,352],{},[149,336,314],{},[149,338,339],{},"1.0",[149,341,342],{"align":294},"4,535",[149,344,345],{"align":294},"756",[149,347,348],{"align":294},"504",[149,350,351],{"align":294},"1.23 s",[149,353,354],{"align":294},"30.6 ms",[129,356,357,360,363,366,369,372,375],{},[149,358,359],{},"ATOM, 8 GPUs, cap unstated",[149,361,362],{},"Unstated",[149,364,365],{"align":294},"7,484",[149,367,368],{"align":294},"936",[149,370,371],{"align":294},"831.6",[149,373,374],{"align":294},"0.821 s",[149,376,377],{"align":294},"16.36 ms",[129,379,380,383,385,388,391,394,397],{},[149,381,382],{},"vLLM, 8 GPUs, cap unstated",[149,384,362],{},[149,386,387],{"align":294},"6,015",[149,389,390],{"align":294},"752",[149,392,393],{"align":294},"668.4",[149,395,396],{"align":294},"1.875 s",[149,398,399],{"align":294},"22.06 ms",[10,401,402],{},[45,403,404],{},"TTFT: time to first token. TPOT: time per output token after the first. Paiton T=0 at 800 W per GPU is the mean of two runs, 5,673 and 5,648 total tok\u002Fs; T=1.0 is one run. Per-GPU figures are rounded. AMD uses TP8\u002FDCP8, concurrency 16 and 64 prompts.",[10,406,407],{},"At T=0, Paiton's per-GPU throughput is close to ATOM's published result. At T=1.0, it is close to vLLM's. These pairings are illustrative: AMD's unreported sampling temperatures prevent a temperature-matched comparison. In those pairings, Paiton's output-token latency is higher and its first-token latency lower.",[277,409,411],{"id":410},"_128k-input-and-100-output","128K input and 100 output",[123,413,414,428],{},[126,415,416],{},[129,417,418,420,422,424,426],{},[132,419,289],{},[132,421,202],{},[132,423,295],{"align":294},[132,425,298],{"align":294},[132,427,304],{"align":294},[144,429,430,445,460,475],{},[129,431,432,434,436,439,442],{},[149,433,314],{},[149,435,317],{},[149,437,438],{"align":294},"12,768",[149,440,441],{"align":294},"2,128",[149,443,444],{"align":294},"122 s",[129,446,447,449,451,454,457],{},[149,448,314],{},[149,450,339],{},[149,452,453],{"align":294},"12,885",[149,455,456],{"align":294},"2,148",[149,458,459],{"align":294},"100 s",[129,461,462,464,466,469,472],{},[149,463,359],{},[149,465,362],{},[149,467,468],{"align":294},"14,005",[149,470,471],{"align":294},"1,751",[149,473,474],{"align":294},"27.9 s",[129,476,477,479,481,484,487],{},[149,478,382],{},[149,480,362],{},[149,482,483],{"align":294},"17,678",[149,485,486],{"align":294},"2,210",[149,488,489],{"align":294},"105.5 s",[10,491,492],{},[227,493],{"alt":494,"src":495},"Total input-plus-output tokens per second per GPU for the 8K\u002F1K and 128K\u002F100 workloads, separating Paiton's two sampling temperatures from AMD's unstated temperatures.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fper-gpu.webp",[10,497,498],{},[45,499,500,501,264],{},"Paiton: six MI355X GPUs at 800 W each. Published AMD results: eight MI355X GPUs, power limits and sampling temperatures unstated. The tables above contain the chart values. ",[13,502,504],{"href":503},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fper-gpu.svg","Download the SVG",[10,506,507],{},"Paiton at T=0 and 800 W per GPU is about 22% ahead of ATOM per GPU on 128K\u002F100, and about 4% behind vLLM. ATOM has much lower first-token latency. In Paiton's run, 16 long requests arrive together and their prompts are processed sequentially, so requests wait behind earlier prefills. A good tokens-per-GPU figure does not remove that wait.",[277,509,511],{"id":510},"one-request-and-a-million-token-prompt","One request and a million-token prompt",[123,513,514,530],{},[126,515,516],{},[129,517,518,521,523,525,528],{},[132,519,520],{},"Paiton, 6 MI355X GPUs × 800 W",[132,522,202],{},[132,524,307],{"align":294},[132,526,527],{"align":294},"Approximate output tok\u002Fs",[132,529,304],{"align":294},[144,531,532,548],{},[129,533,534,537,539,542,545],{},[149,535,536],{},"8K\u002F1K, one request at a time",[149,538,317],{},[149,540,541],{"align":294},"4.35 ms",[149,543,544],{"align":294},"230",[149,546,547],{"align":294},"0.61 s",[129,549,550,552,554,557,560],{},[149,551,536],{},[149,553,339],{},[149,555,556],{"align":294},"11.5 ms",[149,558,559],{"align":294},"87",[149,561,547],{"align":294},[10,563,564],{},[45,565,566],{},"Eight random-prompt requests per configuration. Output speed is the reciprocal of median TPOT.",[10,568,569],{},"Random prompts favor speculative decoding at T=0 because their greedy continuations become repetitive. On real text at 800 W per GPU, the same path produced about 100 output tokens\u002Fs for one request. AMD's single-request figures use synthetic draft acceptance, so we have not put them in this table.",[10,571,572],{},"In the separate long-context test on six GPUs at 800 W each, one request with 1,040,000 input tokens and 64 output tokens took 90 seconds to its first token, then 19.9 ms per output token. The maximum sequence length and KV pool were both 1,052,672 tokens. GPU memory use reached roughly 98%. This demonstrates that the context fits on six GPUs at 800 W each; it is not the configuration used for the throughput tables.",[104,574,72],{"id":575},"gpu-power-and-cost-scenarios",[10,577,578],{},"Our six GPUs averaged 4.18 kW on 8K\u002F1K and 4.58 kW on 128K\u002F100 at T=0, below their combined 4.8 kW cap. These readings cover GPUs only, not the host or cooling.",[10,580,581,582,587],{},"AMD does not publish measured draw or configured caps for its runs. For the calculations below, we assume 1,400 W per GPU, or 11.2 kW for eight GPUs, using the ",[13,583,586],{"href":584,"rel":585},"https:\u002F\u002Fwww.amd.com\u002Fen\u002Fproducts\u002Faccelerators\u002Finstinct\u002Fmi350\u002Fmi355x.html",[17],"MI355X's rated typical board power",". That specification is not evidence that AMD ran at that cap or consumed that power.",[123,589,590,603],{},[126,591,592],{},[129,593,594,597,600],{},[132,595,596],{},"Total tok\u002Fs per kW",[132,598,599],{"align":294},"8K\u002F1K",[132,601,602],{"align":294},"128K\u002F100",[144,604,605,616,627,638],{},[129,606,607,610,613],{},[149,608,609],{},"Paiton T=0, 6 GPUs × 800 W, actual 4.8 kW GPU cap",[149,611,612],{"align":294},"1,179",[149,614,615],{"align":294},"2,660",[129,617,618,621,624],{},[149,619,620],{},"Paiton T=0, 6 GPUs × 800 W, measured 4.18\u002F4.58 kW GPU draw",[149,622,623],{"align":294},"1,354",[149,625,626],{"align":294},"2,788",[129,628,629,632,635],{},[149,630,631],{},"ATOM, assumed 11.2 kW GPU cap",[149,633,634],{"align":294},"668",[149,636,637],{"align":294},"1,250",[129,639,640,643,646],{},[149,641,642],{},"vLLM, assumed 11.2 kW GPU cap",[149,644,645],{"align":294},"537",[149,647,648],{"align":294},"1,578",[10,650,651],{},[227,652],{"alt":653,"src":654},"Throughput per kilowatt using Paiton's configured GPU cap and an assumed 1,400 W per GPU for the published AMD results; this is a power-limit scenario.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fper-kw.webp",[10,656,657],{},[45,658,659,660,264],{},"Input plus output tokens. Paiton T=0, six GPUs × 800 W; AMD sampling temperatures and actual caps are unstated. Dividing by assumed caps does not measure comparative energy efficiency. ",[13,661,504],{"href":662},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fper-kw.svg",[277,664,666],{"id":665},"gpu-electricity","GPU electricity",[10,668,669],{},"At an assumed $0.10\u002FkWh, these are GPU-only costs per million total tokens. CPUs, networking and cooling are excluded.",[123,671,672,683],{},[126,673,674],{},[129,675,676,679,681],{},[132,677,678],{},"Electricity scenario",[132,680,599],{"align":294},[132,682,602],{"align":294},[144,684,685,696,706,717],{},[129,686,687,690,693],{},[149,688,689],{},"Paiton T=0, 6 GPUs at their 4.8 kW cap",[149,691,692],{"align":294},"$0.024",[149,694,695],{"align":294},"$0.010",[129,697,698,701,704],{},[149,699,700],{},"Paiton T=0, measured GPU draw",[149,702,703],{"align":294},"$0.021",[149,705,695],{"align":294},[129,707,708,711,714],{},[149,709,710],{},"ATOM, 8 GPUs assumed at 11.2 kW",[149,712,713],{"align":294},"$0.042",[149,715,716],{"align":294},"$0.022",[129,718,719,722,725],{},[149,720,721],{},"vLLM, 8 GPUs assumed at 11.2 kW",[149,723,724],{"align":294},"$0.052",[149,726,727],{"align":294},"$0.018",[10,729,730,731,734],{},"The calculation is ",[272,732,733],{},"kW × price per kWh × 1,000,000 ÷ (tokens\u002Fs × 3,600)",". At continuous GPU draw of 4.8 or 11.2 kW, annual electricity would be $4,205 or $9,811. These scenarios cannot establish a measured electricity saving against AMD.",[277,736,738],{"id":737},"gpu-rental","GPU rental",[10,740,741,742,747,748,753,754,759],{},"For an illustrative GPU-cost allocation, we use published rates checked on 11 November 2026: ",[13,743,746],{"href":744,"rel":745},"https:\u002F\u002Ftensorwave.com\u002Fproducts\u002Faccelerators\u002Famd-mi355x",[17],"TensorWave starts at $2.95\u002FGPU-hour",", ",[13,749,752],{"href":750,"rel":751},"https:\u002F\u002Fwww.daytona.io\u002Fpricing",[17],"Daytona lists $5.99\u002FGPU-hour"," with CPU, RAM and storage priced separately, and ",[13,755,758],{"href":756,"rel":757},"https:\u002F\u002Fwww.oracle.com\u002Fapac\u002Fa\u002Focom\u002Fdocs\u002Fcorporate\u002Fpricing\u002Foracle-paas-and-iaas-global-price-list.pdf",[17],"Oracle lists $8.60\u002FGPU-hour"," for bare-metal service. These different offers are not interchangeable whole-node quotes. All rows use total tokens, including prompt tokens.",[123,761,762,778],{},[126,763,764],{},[129,765,766,769,772,775],{},[132,767,768],{},"Cost per million total tokens",[132,770,771],{"align":294},"At $2.95\u002FGPU-h",[132,773,774],{"align":294},"At $5.99\u002FGPU-h",[132,776,777],{"align":294},"At $8.60\u002FGPU-h",[144,779,780,794,808,822,836,850,864],{},[129,781,782,785,788,791],{},[149,783,784],{},"8K\u002F1K, Paiton, 6 GPUs × 800 W, T=0",[149,786,787],{"align":294},"$0.87",[149,789,790],{"align":294},"$1.76",[149,792,793],{"align":294},"$2.53",[129,795,796,799,802,805],{},[149,797,798],{},"8K\u002F1K, Paiton, 6 GPUs × 800 W, T=1.0",[149,800,801],{"align":294},"$1.08",[149,803,804],{"align":294},"$2.20",[149,806,807],{"align":294},"$3.16",[129,809,810,813,816,819],{},[149,811,812],{},"8K\u002F1K, ATOM, 8 GPUs, cap\u002FT unstated",[149,814,815],{"align":294},"$0.88",[149,817,818],{"align":294},"$1.78",[149,820,821],{"align":294},"$2.55",[129,823,824,827,830,833],{},[149,825,826],{},"8K\u002F1K, vLLM, 8 GPUs, cap\u002FT unstated",[149,828,829],{"align":294},"$1.09",[149,831,832],{"align":294},"$2.21",[149,834,835],{"align":294},"$3.18",[129,837,838,841,844,847],{},[149,839,840],{},"128K\u002F100, Paiton, 6 GPUs × 800 W, T=0",[149,842,843],{"align":294},"$0.39",[149,845,846],{"align":294},"$0.78",[149,848,849],{"align":294},"$1.12",[129,851,852,855,858,861],{},[149,853,854],{},"128K\u002F100, ATOM, 8 GPUs, cap\u002FT unstated",[149,856,857],{"align":294},"$0.47",[149,859,860],{"align":294},"$0.95",[149,862,863],{"align":294},"$1.36",[129,865,866,869,872,875],{},[149,867,868],{},"128K\u002F100, vLLM, 8 GPUs, cap\u002FT unstated",[149,870,871],{"align":294},"$0.37",[149,873,874],{"align":294},"$0.75",[149,876,801],{"align":294},[10,878,879,880,883],{},"Rental cost is ",[272,881,882],{},"GPU count × hourly rate × 1,000,000 ÷ (tokens\u002Fs × 3,600)",". If you rent an indivisible eight-GPU node, leaving two GPUs unused will not reduce the invoice. They are capacity for another workload, subject to the host's remaining resources. Test combined workloads before counting that capacity as revenue.",[104,885,78],{"id":886},"why-six-gpus-and-800-w",[10,888,889],{},"The model's roughly 1.56 TB of weights and a million-token KV pool fit on six MI355X GPUs. Paiton's own sharding supports that GPU count even though the model's main dimensions do not divide evenly by six. The framework runs its own HIP kernels on the GPUs.",[10,891,892],{},"Two GPUs and 576 GB of rated HBM capacity remain outside the Kimi process. Embeddings, reranking or a second model could use them, but those combinations were not benchmarked here. Six GPUs also have less compute and memory bandwidth than eight. The published eight-GPU results lead in total node throughput.",[10,894,895],{},"The 800 W limit came from our single-phase PDU. Testing at 1,000 W per GPU brought the feed close to its limit; sustained eight-GPU operation at 1,400 W was beyond that supply. All end-to-end Paiton benchmark results in this article use 800 W per GPU.",[277,897,899],{"id":898},"what-1400-w-might-change","What 1,400 W might change",[10,901,902],{},"We could not run the full model at 1,400 W. The following ranges are projections from limited component tests, assuming faster prefill and at most 10% faster decode. They have not been validated at that power limit.",[123,904,905,918],{},[126,906,907],{},[129,908,909,912,915],{},[132,910,911],{},"Paiton on 6 GPUs, total tok\u002Fs\u002FGPU",[132,913,914],{"align":294},"Measured at 800 W\u002FGPU",[132,916,917],{"align":294},"Estimated at 1,400 W\u002FGPU",[144,919,920,930,940],{},[129,921,922,925,927],{},[149,923,924],{},"8K\u002F1K, T=0",[149,926,323],{"align":294},[149,928,929],{"align":294},"1,035-1,105",[129,931,932,935,937],{},[149,933,934],{},"8K\u002F1K, T=1.0",[149,936,345],{"align":294},[149,938,939],{"align":294},"830-885",[129,941,942,945,947],{},[149,943,944],{},"128K\u002F100, T=0",[149,946,441],{"align":294},[149,948,949],{"align":294},"2,780-2,790",[10,951,952],{},[227,953],{"alt":954,"src":955},"Paiton's measured total tokens per second per GPU at 800 W beside estimated ranges at 1,400 W, with estimates visibly distinguished.","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fpower-estimate.webp",[10,957,958],{},[45,959,960,961,264],{},"The 1,400 W values are extrapolations from a component test and assumed decode scaling. They are not measured whole-model results or a performance promise. ",[13,962,504],{"href":963},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fpower-estimate.svg",[104,965,84],{"id":966},"sampling-and-accuracy",[10,968,969,970,975],{},"Paiton uses the public ",[13,971,974],{"href":972,"rel":973},"https:\u002F\u002Fhuggingface.co\u002FInferact\u002FKimi-K3.1-Pro-DSpark-preview",[17],"Inferact DSpark drafter",". Sampling temperature changes how many drafted tokens the target model accepts. At T=0, repetitive random-prompt continuations are easier to predict; at T=1.0, more drafts are rejected. This is why we report both temperatures without assigning either one to AMD's unpublished settings.",[10,977,978],{},"The following results are from the fixed six-GPU build at 800 W per GPU, with a 1M-token context, greedy decoding and thinking off. We used the same harness and test sets as before; MMLU-Pro is new to this post.",[123,980,981,994],{},[126,982,983],{},[129,984,985,988,991],{},[132,986,987],{},"Benchmark",[132,989,990],{},"Fixed build, 12 November 2026",[132,992,993],{},"Previously reported, affected build",[144,995,996,1007,1018],{},[129,997,998,1001,1004],{},[149,999,1000],{},"GSM8K, five-shot",[149,1002,1003],{},"197\u002F200 (98.5%)",[149,1005,1006],{},"185\u002F200 (92.5%)",[129,1008,1009,1012,1015],{},[149,1010,1011],{},"HumanEval, pinned OpenAI set",[149,1013,1014],{},"159\u002F164 (97.0%)",[149,1016,1017],{},"160\u002F164 (97.6%)",[129,1019,1020,1023,1026],{},[149,1021,1022],{},"MMLU-Pro",[149,1024,1025],{},"1,157\u002F1,400 (82.6%)",[149,1027,1028],{},"Not previously reported",[10,1030,1031],{},"HumanEval is unchanged within measurement noise. These checks do not establish accuracy on every task.",[10,1033,1034],{},"Our earlier KL figures used a flawed reference and are no longer valid. Against the corrected exact model, the served settings have a prefill-path KL divergence of 0.0078 across 47 held-out rows and 85,152 positions, with 96.4% top-1 agreement. Against that same corrected reference, the old build scores 0.109 and stock vLLM scores 0.0030. Lower KL means closer to the reference distribution, not necessarily better answers on every task.",[10,1036,1037],{},"We have not re-measured decode-path KL.",[104,1039,90],{"id":1040},"running-on-four-gpus",[10,1042,1043],{},"The timings in this section were measured before the fix. The corrected six-GPU accuracy results above do not validate this four-GPU configuration.",[10,1045,1046],{},"Paiton also runs Kimi-K3.1-Pro on four MI355X GPUs at 800 W each. The weights do not fit entirely in their HBM, so the node's two AMD EPYC 9575F CPUs compute part of the expert workload.",[123,1048,1049,1066],{},[126,1050,1051],{},[129,1052,1053,1056,1058,1060,1063],{},[132,1054,1055],{},"Paiton, 4 GPUs × 800 W plus CPU experts",[132,1057,295],{"align":294},[132,1059,298],{"align":294},[132,1061,1062],{"align":294},"Median TPOT, 16 concurrent",[132,1064,1065],{"align":294},"Single-request TPOT",[144,1067,1068,1084],{},[129,1069,1070,1072,1075,1078,1081],{},[149,1071,924],{},[149,1073,1074],{"align":294},"915",[149,1076,1077],{"align":294},"229",[149,1079,1080],{"align":294},"152 ms",[149,1082,1083],{"align":294},"26.1 ms, about 38 output tok\u002Fs",[129,1085,1086,1088,1091,1094,1097],{},[149,1087,934],{},[149,1089,1090],{"align":294},"781",[149,1092,1093],{"align":294},"195",[149,1095,1096],{"align":294},"179 ms",[149,1098,1099],{"align":294},"26.4 ms, about 38 output tok\u002Fs",[10,1101,1102],{},[45,1103,1104],{},"Same 64-request, concurrency-16 workload; T=0 is the mean of two runs. Single-request timing is a separate test. GPUs 2-5 are active.",[10,1106,1107],{},[239,1108],{"controls":241,"playsInline":241,"preload":242,"width":243,"height":244,"ariaLabel":1109,"ariaDescribedBy":1110,"poster":1112,"src":1113},"Accelerated recording of the four-GPU Kimi-K3.1-Pro benchmark with CPU experts",[1111],"kimi-four-gpu-caption","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fw4-8k1k-midrun-amdsmi.png","\u002Fasset\u002Fvideos\u002Fblog\u002Fpaiton-kimi31\u002Fw4-8k1k-t0-run.mp4",[10,1115,1116],{"id":1111},[45,1117,1118],{},"Accelerated benchmark recording, not real time. Four MI355X GPUs at an 800 W cap each, with CPU experts, sampling temperature 0.",[10,1120,1121,259,1124,264],{},[13,1122,1123],{"href":1113},"Open the four-GPU recording",[13,1125,263],{"href":1126},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fw4-8k1k-final-result.png",[10,1128,1129],{},"This uses fewer GPUs but loses substantial throughput while waiting for CPU work. The four unused GPUs share that busy host. Two concurrent Kimi instances have not been validated.",[104,1131,96],{"id":1132},"try-the-community-build",[10,1134,1135,1136,1141],{},"The free ",[13,1137,1140],{"href":1138,"rel":1139},"https:\u002F\u002Fgithub.com\u002FEliovp-BV\u002Fpaiton-vllm-plugin\u002Ftree\u002Fmain\u002Fmodels\u002FKimi-K3.1-Pro",[17],"community build and launcher"," run the model on four supported AMD Instinct CDNA4 GPUs. It is plain Python\u002FPyTorch, with no Paiton kernels, batching or speculative decoding. It is a way to inspect the model and try your own prompts; it does not reproduce Paiton's performance.",[123,1143,1144,1156],{},[126,1145,1146],{},[129,1147,1148,1151,1154],{},[132,1149,1150],{},"Four MI355X GPUs at 800 W each, T=0",[132,1152,1153],{"align":294},"Community build",[132,1155,23],{"align":294},[144,1157,1158,1169,1180,1191],{},[129,1159,1160,1163,1166],{},[149,1161,1162],{},"One request, time per output token",[149,1164,1165],{"align":294},"1.9-4.6 s",[149,1167,1168],{"align":294},"About 26 ms",[129,1170,1171,1174,1177],{},[149,1172,1173],{},"Prefill",[149,1175,1176],{"align":294},"About 5 input tok\u002Fs",[149,1178,1179],{"align":294},"About 3,200 input tok\u002Fs",[129,1181,1182,1185,1188],{},[149,1183,1184],{},"Concurrent requests",[149,1186,1187],{"align":294},"1",[149,1189,1190],{"align":294},"16",[129,1192,1193,1196,1199],{},[149,1194,1195],{},"Complete 8K\u002F1K request",[149,1197,1198],{"align":294},"About 1-2 hours, estimated",[149,1200,1201],{"align":294},"29 s",[10,1203,1204],{},[45,1205,1206],{},"Historical measurements before the fix, at 800 W per GPU. Community prefill: 1,175 tokens in 239 seconds; Paiton prefill: roughly 8K tokens in 2.6 seconds. The community 8K\u002F1K duration is estimated from measured phase speeds, not a completed two-hour benchmark. These phase tests differ in prompt length.",[10,1208,1209],{},"As of 12 November 2026, the community fix is in the plugin repo and the container image is being rebuilt. The tag below identifies the original release; it is not confirmation that a pull includes the fix.",[1211,1212,1218],"pre",{"className":1213,"code":1215,"language":1216,"meta":1217},[1214],"language-text","ghcr.io\u002Feliovp\u002Fpaiton-vllm-plugin:kimi31pro-rocm10.1.0-pytorch2.13.0-cdna4-v1\n","text","",[272,1219,1215],{"__ignoreMap":1217},[10,1221,1222,1223,1226,1227,1232],{},"Read the guide before downloading. It requires supported ",[272,1224,1225],{},"gfx950"," GPUs, at least 768 GB of host RAM and storage for roughly 1.6 TB of weights plus the approximately 37 GB image. It serves one request at a time with a default maximum length of 32,768 tokens; this is not a validated million-token community setup. The code's Apache-2.0 license and ",[13,1228,1231],{"href":1229,"rel":1230},"https:\u002F\u002Fhuggingface.co\u002Fmoonshotai\u002FKimi-K3.1-Pro\u002Fblob\u002Fmain\u002FLICENSE",[17],"Kimi's model terms"," are separate. Check the model's commercial conditions for your use.",[104,1234,102],{"id":1235},"measurement-details-and-a-second-benchmark",[10,1237,1238,1239,1242],{},"Our node has eight MI355X GPUs, two 64-core EPYC 9575F CPUs and 2.2 TB of RAM. Six-GPU tests used GPUs 2-7; four-GPU tests used GPUs 2-5. GPU power was sampled with ",[272,1240,1241],{},"amd-smi metric -p"," every two to three seconds, with configured limits saved alongside the results.",[10,1244,1245,1246,1249,1250,1253,1254,1257],{},"The serving client came from ",[272,1247,1248],{},"vllm\u002Fvllm-openai-rocm:kimi-k31",". For 8K\u002F1K it used ",[272,1251,1252],{},"--dataset-name random --random-input-len 8192 --random-output-len 1024 --num-prompts 64 --max-concurrency 16 --ignore-eos",". T=0 requests explicitly set ",[272,1255,1256],{},"--temperature 0","; the other run used Paiton's server default of 1.0. Prefix caching was disabled and prefill chunks were 8,192 tokens.",[10,1259,1260,1261,1264],{},"We also ran LocalMaxxing's open CLI, ",[272,1262,1263],{},"lmx 0.1.60",", against Paiton on six MI355X GPUs capped at 800 W each with the million-token configuration. Its prompts and timing differ from the random-token benchmark.",[123,1266,1267,1277],{},[126,1268,1269],{},[129,1270,1271,1274],{},[132,1272,1273],{},"LocalMaxxing CLI test, 6 GPUs × 800 W",[132,1275,1276],{},"Recorded result",[144,1278,1279,1287,1295,1303],{},[129,1280,1281,1284],{},[149,1282,1283],{},"One request, canonical reasoning prompt",[149,1285,1286],{},"183 output tok\u002Fs; 116 ms to first token",[129,1288,1289,1292],{},[149,1290,1291],{},"One cold request, prefill",[149,1293,1294],{},"12,844 input tok\u002Fs at 8K; 13,370 at 128K, 9.5 s",[129,1296,1297,1300],{},[149,1298,1299],{},"Sustained decode, 16 requests, short prompts",[149,1301,1302],{},"741 aggregate output tok\u002Fs; median 46 tok\u002Fs per request",[129,1304,1305,1308],{},[149,1306,1307],{},"Sustained decode, 16 requests, 64K prompts",[149,1309,1310],{},"663 aggregate output tok\u002Fs; median 28 tok\u002Fs per request",[10,1312,1313],{},[45,1314,1315],{},"Aggregate output is counted across each 30-second decode cell. Per-request speed is the median reciprocal inter-token latency, not aggregate throughput divided by concurrency.",[10,1317,1318,1319,1323,1324,1328],{},"These are our local CLI results, not a published leaderboard entry. Submission was rejected because the accepted AMD Instinct list stopped at MI300X and MI325X and did not include MI355X. The ",[13,1320,1322],{"href":1321},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Flocalmaxxing-results.html","saved results page"," and ",[13,1325,1327],{"href":1326},"\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Flocalmaxxing-results.png","original screenshot"," are available here.",[10,1330,1331,1332,1323,1336,1340],{},"Thanks to Moonshot AI for the released weights, Inferact for the drafter, and the AMD ATOM and vLLM teams for publishing their measurements. Our earlier ",[13,1333,1335],{"href":1334},"\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p","MI350P results",[13,1337,1339],{"href":1338},"\u002Fblog\u002Fpaiton-qwen38-mxfp4-dflash2-r9700","R9700 work"," cover other model and hardware budgets.",[277,1342,1344],{"id":1343},"running-a-token-factory","Running a token factory?",[10,1346,1347,1348,1352],{},"Or planning one? ",[13,1349,1351],{"href":1350},"\u002Fcontact","Bring your model, latency target and hardware budget",". We can test the tradeoffs with you.",{"title":1217,"searchDepth":1354,"depth":1354,"links":1355},2,[1356,1357,1363,1367,1370,1371,1372,1373],{"id":106,"depth":1354,"text":60},{"id":267,"depth":1354,"text":66,"children":1358},[1359,1361,1362],{"id":279,"depth":1360,"text":280},3,{"id":410,"depth":1360,"text":411},{"id":510,"depth":1360,"text":511},{"id":575,"depth":1354,"text":72,"children":1364},[1365,1366],{"id":665,"depth":1360,"text":666},{"id":737,"depth":1360,"text":738},{"id":886,"depth":1354,"text":78,"children":1368},[1369],{"id":898,"depth":1360,"text":899},{"id":966,"depth":1354,"text":84},{"id":1040,"depth":1354,"text":90},{"id":1132,"depth":1354,"text":96},{"id":1235,"depth":1354,"text":102,"children":1374},[1375],{"id":1343,"depth":1360,"text":1344},[23,1377,1378,1379],"AMD Instinct","Inference","Kimi","2026-11-11T12:00:00Z","Kimi-K3.1-Pro on six MI355X GPUs at 800 W each. Paiton throughput, power measurements, a separate 1M-context test and a four-GPU option.","md","Kimi-K3.1-Pro on 6 MI355X GPUs at 800 W each","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fhero.webp",{},"https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-kimi-k31-pro-amd-instinct-mi355x","\u002Fblog\u002Fpaiton-kimi-k31-pro-amd-instinct-mi355x",{"title":5,"description":1381},"paiton-kimi-k31-pro-amd-instinct-mi355x","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-kimi31\u002Fsocial.webp","blog\u002Fpaiton-kimi-k31-pro-amd-instinct-mi355x","2026-11-12",null,"2r8TCZ_wPsqb4ZpLN7B7tvCDTWzrzAsRWX3XmrOKmZY",[1396,1409,1411,1421,1432,1442,1451,1466,1475,1484,1493,1502,1514,1524,1536,1545,1559,1568,1580,1611,1622,1644,1662,1681,1699,1716,1733,1744,1760,1775,1787,1796,1804,1819,1831,1842,1853,1863,1876,1886,1899,1910,1920,1931,1940,1952,1963,1972],{"path":1397,"title":1398,"description":1399,"date":1400,"slug":1401,"image":1402,"originalUrl":1403,"categories":1404},"\u002Fblog\u002Fpaiton-minimax-h3-image-radeon-ai-pro-r9700","MiniMax-H3-Image on a single R9700","MiniMax-H3-Image on one Radeon AI PRO R9700: 2048 x 2048 in 6.9 seconds, 4096 x 4096 in 33.4 seconds. FP8 weights, image editing and setup.","2026-11-12T10:00:00Z","paiton-minimax-h3-image-radeon-ai-pro-r9700","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-minimax-h3-image\u002F01-hero-concert.jpg","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-minimax-h3-image-radeon-ai-pro-r9700",[23,1405,1406,1407,1408],"AMD Radeon","Local AI","Image Generation","MiniMax",{"path":1387,"title":5,"description":1381,"date":1380,"slug":1389,"image":1384,"originalUrl":1386,"categories":1410},[23,1377,1378,1379],{"path":1412,"title":1413,"description":1414,"date":1415,"slug":1416,"image":1417,"originalUrl":1418,"categories":1419},"\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",[23,1405,1406,1420],"3D Generation",{"path":1334,"title":1422,"description":1423,"date":1424,"slug":1425,"image":1426,"originalUrl":1427,"categories":1428},"Deepseek V4.1 Flash on a single MI350P","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.","2026-11-03T12:00:00Z","paiton-deepseek-v41-flash-amd-instinct-mi350p","\u002Fasset\u002Fimages\u002Fblog\u002Fpaiton-deepseek-v41\u002Fhero.webp","https:\u002F\u002Feliovp.com\u002Fblog\u002Fpaiton-deepseek-v41-flash-amd-instinct-mi350p",[23,1377,1429,1430,1431],"AI Inference","DeepSeek","Quantization",{"path":1433,"title":1434,"description":1435,"date":1436,"slug":1437,"image":1438,"originalUrl":1439,"categories":1440},"\u002Fblog\u002Fpaiton-qwen38-flash-next-amd-instinct-mi350p","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",[23,1377,1429,1441,1431],"Qwen",{"path":1443,"title":1444,"description":1445,"date":1446,"slug":1447,"image":1448,"originalUrl":1449,"categories":1450},"\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",[23,1405,1406,1441,1431],{"path":1452,"title":1453,"description":1454,"date":1455,"slug":1456,"image":1457,"originalUrl":1458,"categories":1459},"\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",[1460,1461,1462,1463,1464,1465],"Modular DC","AI Infrastructure","ModFlex","NVIDIA","Vera Rubin","Liquid Cooling",{"path":1467,"title":1468,"description":1469,"date":1470,"slug":1471,"image":1472,"originalUrl":1473,"categories":1474},"\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",[23,1405,1406,1441,1431],{"path":1476,"title":1477,"description":1478,"date":1479,"slug":1480,"image":1481,"originalUrl":1482,"categories":1483},"\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",[23,1405,1406,1441,1431],{"path":1485,"title":1486,"description":1487,"date":1488,"slug":1489,"image":1490,"originalUrl":1491,"categories":1492},"\u002Fblog\u002Fpaiton-qwen38-w3a4-radeon-ai-pro-r9700","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",[23,1405,1406,1441,1431],{"path":1494,"title":1495,"description":1496,"date":1497,"slug":1498,"image":1499,"originalUrl":1500,"categories":1501},"\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",[23,1405,1406,1407,1441],{"path":1338,"title":1503,"description":1504,"date":1505,"slug":1506,"image":1507,"originalUrl":1508,"categories":1509},"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",[23,1405,1510,1511,1512,1513],"vLLM","Qwen3.8","Inference Optimization","DFlash2",{"path":1515,"title":1516,"description":1517,"date":1518,"slug":1519,"image":1520,"originalUrl":1521,"categories":1522},"\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",[23,1405,1406,1523,1510],"GGUF",{"path":1525,"title":1526,"description":1527,"date":1528,"slug":1529,"image":1530,"originalUrl":1531,"categories":1532},"\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",[23,1405,1406,1533,1534,1535],"Video Generation","MiniMax H3","ComfyUI",{"path":1537,"title":1538,"description":1539,"date":1540,"slug":1541,"image":1542,"originalUrl":1537,"categories":1543},"\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",[23,1405,1406,1407,1544,1535],"FLUX",{"path":1546,"title":1547,"description":1548,"date":1549,"slug":1550,"image":1551,"originalUrl":1552,"categories":1553},"\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",[23,1554,1405,1429,1555,1556,1512,1557,1510,1558],"Artificial Intelligence","GPU Performance","Inference Latency","Large Language Models","Cost Efficiency",{"path":1560,"title":1561,"description":1562,"date":1563,"slug":1564,"image":1565,"originalUrl":1566,"categories":1567},"\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",[23,1554,1405,1429,1555,1556,1512,1557,1510,1558],{"path":1569,"title":1570,"description":1571,"date":1572,"slug":1573,"image":1574,"originalUrl":1393,"categories":1575},"\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",[1576,1461,1577,1462,1578,1579],"All","Data Centers","HPC","AMD Helios",{"path":1581,"title":1582,"description":1583,"date":1584,"slug":1585,"image":1586,"originalUrl":1587,"categories":1588},"\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",[1576,1554,23,1589,1590,1591,1592,1593,1594,1595,1596,1597,1598,1599,1378,1600,1601,1602,1603,1604,23,1605,1606,1607,1608,1609,1610],"14B","AMD","B200","Benchmarks","Blackwell","Compute","Diffusion","Eliovp","Generative AI","GPU","Hardware","Instinct","MI355x","NVidia","On-Premise","Optimization","Sovereign AI","T2V","Text-to-Video","Tuning","Video-Generation","Wan2.2",{"path":1612,"title":1613,"description":1614,"date":1615,"slug":1616,"image":1617,"originalUrl":1618,"categories":1619},"\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",[1576,1554,1460,1620,1590,1621,1460,1602],"Uncategorized","De Tijd",{"path":1623,"title":1624,"description":1625,"date":1626,"slug":1627,"image":1628,"originalUrl":1629,"categories":1630},"\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",[1576,1554,1631,1620,1632,1633,1634,1635,1636,1637,1638,1639,1596,1640,1597,1641,1642,1643],"Trending","AI Act","Anthropomorphism","AVG","Benelux","ChatGPT","Cybersecurity","Data Governance","Data Security","GDPR","Microsoft Copilot","Privacy","Shadow AI",{"path":1645,"title":1646,"description":1647,"date":1648,"slug":1649,"image":1650,"originalUrl":1651,"categories":1652},"\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",[1576,1653,1654,1655,1637,1656,1657,1658,1640,1659,1660,1661,1642],"AWS","Belgian Mobile ID","Cloud Act","Data Sovereignty","Digital Identity","eIDAS","itsme","Liberty Global","MyGov.be",{"path":1663,"title":1664,"description":1665,"date":1666,"slug":1667,"image":1668,"originalUrl":1669,"categories":1670},"\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",[1576,1554,1671,1631,1672,1673,1674,1675,1676,1677,1678,1679,1605,1680],"Solutions","Agentic AI","AI Engineering","AI Strategy","Autonomous Agents","Enterprise AI","Local LLM","Model Fine-Tuning","On-Premise AI","VRAM Optimization",{"path":1682,"title":1683,"description":1684,"date":1685,"slug":1686,"image":1687,"originalUrl":1688,"categories":1689},"\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",[1576,1554,1631,1461,1690,1691,1692,1693,1694,1695,1696,1697,1698],"AI Neocloud","Circular Financing","GPU Cloud","Investment Risks","Startup Valuation","Synthetic Bubble","Tech Analysis","Vaporware","Venture Capital",{"path":1700,"title":1701,"description":1702,"date":1703,"slug":1704,"image":1705,"originalUrl":1706,"categories":1707},"\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",[1576,1460,1620,1708,1461,1709,1710,1465,1711,1712,1713,1714,1715],"150kW Rack","DLC","High Density","Modular Data Center","NVIDIA Blackwell Ultra","NVIDIA GB300","NVL72","Rapid Deployment",{"path":1717,"title":1718,"description":1719,"date":1720,"slug":1721,"image":1722,"originalUrl":1723,"categories":1724},"\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",[1576,1554,23,1620,1725,1554,1726,1727,1728,1729,1730,1731,23,1732],"AMD MI300X","CUDA Translation","FP8","GPU Optimization","High Performance Computing","HIP","Kernel Tuning","ROCm",{"path":1734,"title":1735,"description":1736,"date":1737,"slug":1738,"image":1739,"originalUrl":1740,"categories":1741},"\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",[1576,1554,23,1429,1377,1725,1558,1742,1512,1731,23,1743,1510],"High Throughput","SGLang",{"path":1745,"title":1746,"description":1747,"date":1748,"slug":1749,"image":1750,"originalUrl":1751,"categories":1752},"\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",[1576,1554,23,1753,1725,1754,1512,1755,1756,1757,1758,23,1759],"AI Benchmarks","Cost per Token","Mixture of Experts","MoE","NVIDIA B200","NVIDIA H200","Qwen3",{"path":1761,"title":1762,"description":1763,"date":1764,"slug":1765,"image":1766,"originalUrl":1767,"categories":1768},"\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",[1576,1554,1671,1620,1672,1769,1770,1771,1772,1677,1679,1605,1773,1774],"Damage Detection","Document Processing","Email Automation","Invoice Extraction","Ticket Automation","Workflow Automation",{"path":1776,"title":1777,"description":1778,"date":1779,"slug":1780,"image":1781,"originalUrl":1782,"categories":1783},"\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",[1576,1554,23,1620,1784,1590,1591,1785,1786,1601,1602,23,1510],"AI","H200","MI300X",{"path":1788,"title":1789,"description":1790,"date":1791,"slug":1792,"image":1793,"originalUrl":1794,"categories":1795},"\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",[1576,1554,1671,1672,1769,1770,1771,1772,1677,1679,1605,1773,1774],{"path":1797,"title":1798,"description":1799,"date":1800,"slug":1801,"image":1217,"originalUrl":1802,"categories":1803},"\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",[1576,1554,23],{"path":1805,"title":1806,"description":1807,"date":1808,"slug":1809,"image":1810,"originalUrl":1811,"categories":1812},"\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",[1576,1554,23,1620,1429,1725,1813,1727,1814,1815,1816,23,1817,1818],"Cold Start","Graph Compilation","Llama 3.1 405B","LLM Optimization","Startup Latency","Tensor Parallelism",{"path":1820,"title":1821,"description":1822,"date":1823,"slug":1824,"image":1825,"originalUrl":1826,"categories":1827},"\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",[1576,1554,23,1620,1725,1828,1676,1597,1555,1556,1557,1815,1829,1830],"Cold Start Optimization","Model Serving","vLLM Optimization",{"path":1832,"title":1833,"description":1834,"date":1835,"slug":1836,"image":1837,"originalUrl":1838,"categories":1839},"\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",[1576,1554,23,1671,1620,1590,1786,1602,1840,1841],"RX7900XTX","tenstorrent",{"path":1843,"title":1844,"description":1845,"date":1846,"slug":1847,"image":1848,"originalUrl":1849,"categories":1850},"\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",[1576,1554,1460,1671,1591,1785,1851,1602,1852],"MI325x","pnl calculator",{"path":1854,"title":1855,"description":1856,"date":1857,"slug":1858,"image":1859,"originalUrl":1860,"categories":1861},"\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",[1576,1554,23,1784,1590,1785,1862,1602,23,1510],"MI300",{"path":1864,"title":1865,"description":1866,"date":1867,"slug":1868,"image":1869,"originalUrl":1870,"categories":1871},"\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",[1576,1460,1872,1461,1710,1578,1465,1711,1873,1874,1714,1875],"1-2MW Data Center","NVIDIA Blackwell","NVIDIA NVL","Precision Cooling",{"path":1877,"title":1878,"description":1879,"date":1880,"slug":1881,"image":1882,"originalUrl":1883,"categories":1884},"\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",[1576,1554,1671,1620,1784,1590,1885],"Healthcare",{"path":1887,"title":1888,"description":1889,"date":1890,"slug":1891,"image":1892,"originalUrl":1893,"categories":1894},"\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",[1576,1631,1784,1590,1895,1896,1897,1898],"import","Taiwan","Tariffs","Trump",{"path":1900,"title":1901,"description":1902,"date":1903,"slug":1904,"image":1905,"originalUrl":1906,"categories":1907},"\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",[1576,1554,1671,1784,1908,1909],"AI Agents","ERP",{"path":1911,"title":1912,"description":1913,"date":1914,"slug":1915,"image":1916,"originalUrl":1917,"categories":1918},"\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",[1576,1554,1631,1919,1590,1598,1602],"AI news",{"path":1921,"title":1922,"description":1923,"date":1924,"slug":1925,"image":1926,"originalUrl":1927,"categories":1928},"\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",[1576,1554,23,1784,1590,1929,1930,1786,23],"benchmark","LLM",{"path":1932,"title":1933,"description":1934,"date":1935,"slug":1936,"image":1937,"originalUrl":1938,"categories":1939},"\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",[1576,1554,23],{"path":1941,"title":1942,"description":1943,"date":1944,"slug":1945,"image":1946,"originalUrl":1947,"categories":1948},"\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",[1576,1590,1949,1950,1951],"Jim Greene","Podcast","Tech Talk",{"path":1953,"title":1954,"description":1955,"date":1956,"slug":1957,"image":1958,"originalUrl":1959,"categories":1960},"\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",[1576,1554,23,1590,1961,1962,1785,1786,1851,23,1510],"Deepseek","H100",{"path":1964,"title":1965,"description":1966,"date":1967,"slug":1968,"image":1969,"originalUrl":1970,"categories":1971},"\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",[1576,1554,23,1590,1961,1962,1785,1786,1851,23,1510],{"path":1973,"title":1974,"description":1975,"date":1976,"slug":1977,"image":1978,"originalUrl":1979,"categories":1980},"\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",[1576,1554,23,1590,1962,1785,1786,1851,23,1510],1794475005487]