--- title: Running Kimi K3 on MI355X at Better Performance per... description: Wafer benchmarks Kimi K3, a 2.8T-parameter open source LLM, on AMD MI355X GPUs and finds it delivers 952 tok/s per node at ~48 tok/s per dollar — far ahead of... canonical: https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 | daily.dev og:description: Wafer benchmarks Kimi K3, a 2.8T-parameter open source LLM, on AMD MI355X GPUs and finds it delivers 952 tok/s per node at ~48 tok/s per dollar — far ahead of... og:url: https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd og:image: https://api.daily.dev/og/posts/BzrW6kipd.png og:image:alt: Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 og:image:width: 1200 og:image:height: 630 og:locale: en --- > ## Documentation Index > Fetch the complete documentation index at: https://daily.dev/llms.txt > Use this file to discover all available pages before exploring further. # Running Kimi K3 on MI355X at Better Performance per Dollar Than B300 **[Hacker News](https://daily.dev/sources/hn)** · 6 min read · 0 upvotes · 0 comments ## Summary Wafer benchmarks Kimi K3, a 2.8T-parameter open source LLM, on AMD MI355X GPUs and finds it delivers 952 tok/s per node at ~48 tok/s per dollar — far ahead of NVIDIA B200 and B300 on performance-per-dollar. The post details two key engineering fixes required on ROCm: patching a missing top-k renorm probability function in sglang's speculative decoding path (a pure PyTorch fix, no custom kernel needed), and zero-padding attention heads to enable AMD's fast AITER MLA prefill kernel, which improved prefill speed 2–3×. The conclusion is that AMD's HBM capacity advantage becomes practically decisive at Kimi K3's scale, and the CUDA software moat is narrowing faster than expected. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Community take How the wider developer community reacted, aggregated from 1 discussion and 42 comments across hackernews (as of 2026-08-03). **TL;DR:** The community is largely dismissive of this Wafer.ai post, calling it AI-generated slop with inaccurate or misleading benchmark comparisons and unfair GPU pricing figures; a secondary debate erupts over whether releasing model weights qualifies as 'open source.' **Sentiment:** 10% positive · 30% mixed · 60% skeptical **The case for** - The B300 does outperform the MI355X on most per-GPU throughput metrics according to the data presented. - Some commenters acknowledge the post at least made a light effort at readability compared to typical AI slop. **The pushback** - The article is widely labeled AI-generated slop, with em-dashes left in and no apparent editorial review. - GPU pricing comparisons are seen as cherry-picked and inaccurate — using rental aggregator prices rather than real-world or TCO figures. - The B200 multi-node comparison is criticized as unfair since the model doesn't fit in a single node, inflating latency. - Wafer.ai is accused of a pattern of hype, exaggerated benchmarks, and previously rugpulling a token subscription plan. - The 'open source model' framing is contested — many argue only weights were released, making it 'open weights' at best. - No reproducibility details or code were shared to replicate the benchmark setup. **By community** - hackernews (skeptical): Commenters broadly dismiss the post as AI slop with misleading pricing and benchmark comparisons, while a side thread debates 'open source' vs 'open weights' terminology. **Hottest debate:** Whether releasing model weights without training data or code qualifies as 'open source' or should only be called 'open weights.' **Open questions** - Are the MI355X rental prices cited ($2.50/GPU-hr) real market prices or outliers, and what does actual TCO look like? - Was the zero-padding kernel fix validated for correctness, or could it produce incoherent model outputs? - Why didn't Wafer.ai share code or methodology to allow independent replication of their results? **Highlights** > Lol, such a lazily written article by wafer.ai GPUs. 8× MI355X (TP8) B300 (TP8+DCP8) Decode tok/s per stream 118 tok/s 172 tok/s Peak aggregate. 952 tok/s 1,568 tok/s Peak aggregate per GPU 119 tok/s 196 tok/s On every row the B300 beat the MI355X The B200 is being forcefully compared against something which is not gonna fit within it's memory in a single node & not much details about multi-node interconnectivity, disagg or not. As expected of a shoddy slop. The only point it won is of cost per hour is one aggregation website for rentals, the premium a B300 commands against the $3/hr AMD chip which no provider has in abundance. Never bothered to do TCO of owning the hardware either. > — [villgax on hackernews · 1 comments](https://news.ycombinator.com/item?id=49141794) > Wafer is making themselves synonymous with slop in the inference space. Exaggerated unfair comparisons in all their results, twitter hype posts with alarm emojis etc. > $2.50/GPU-hr for the MI355X, $6.00 for the B300, and $4.25 for the B200. This is not an accurate price comparison for real terms. > — [inferencecoder on hackernews · 2 comments](https://news.ycombinator.com/item?id=49141817) > This is a company that launched a token subscription (WaferPass), before weeks later, rugpulling the plan for being unsustainable while simultaneously claiming they achieved incredible inference efficiency gains worthy of paying them mind. > — [greyb on hackernews](https://news.ycombinator.com/item?id=49142117) > > Is it the preferred way only because training is expensive? No, it's the preferred way because that's literally how you train it. Contrary to popular misconceptions, you don't "compile" data into weights. You initialize a model (based on architecture, config, etc) and then you modify it via training. But crucially they (i.e. model creators) modify it the same way (technically speaking) as you would. That's what the license grants you. Nothing less, nothing more. The "how" as in knowhow has never been something covered by a license. Open-weight is something dreamt up by people misunderstanding the basics of model creation and training, and having ideological biases against AI and/or LLMs. It is what it is, but you should know that you are technically wrong. A model released under an open source license (Apache, MIT, etc) is an open source model. Because the weights are the source of the models. Training is not "compilation". Training is the "how" as in knowhow to modify the model. Training deals with values. Source deals with everything, including values. In the past, if someone would have released a piece of software (say a PID controller algo) w/ hardcoded values, no-one would bat an eye. LLMs are just that, with billions of hardcoded values. Nothing less nothing more. > — [NitpickLawyer on hackernews](https://news.ycombinator.com/item?id=49142172) > They even left the em-dashes. Absolute slop. " The only problem with AMD is software support — slower kernels and less day-0 support on inference frameworks make serving frontier models on AMD a real engineering effort." > — [muragekibicho on hackernews](https://news.ycombinator.com/item?id=49141788) **Source threads** - [hackernews](https://news.ycombinator.com/item?id=49141073) · 100 points · 42 comments ## Similar posts on daily.dev - [Beyond Porting: How vLLM Orchestrates High-Performance Inference on AMD ROCm](https://daily.dev/posts/beyond-porting-how-vllm-orchestrates-high-performance-inference-on-amd-rocm-9sjfha0v1) · vLLM · 2 upvotes · 0 comments - [Attention Decode on AMD MI450 GPUs: A Gluon Kernel Optimization Guide](https://daily.dev/posts/attention-decode-on-amd-mi450-gpus-a-gluon-kernel-optimization-guide-xpcetmubd) · Hacker News · 0 upvotes · 0 comments - [Efficient MoE Pre-training at Scale on 1K AMD GPUs with TorchTitan – PyTorch](https://daily.dev/posts/efficient-moe-pre-training-at-scale-on-1k-amd-gpus-with-torchtitan-pytorch-xscr4vica) · PyTorch · 0 upvotes · 0 comments --- Tags: [#ai-inference](https://daily.dev/tags/ai-inference), [#kimi-k3](https://daily.dev/tags/kimi-k3), [#rocm](https://daily.dev/tags/rocm) [View this post on daily.dev](https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd) ```json {"@context":"https://schema.org","@graph":[{"@type":"Organization","@id":"https://daily.dev/#organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180},"sameAs":["https://twitter.com/dailydotdev","https://github.com/dailydotdev","https://www.linkedin.com/company/daily-dev-ltd"]},{"@type":"WebSite","@id":"https://daily.dev/#website","url":"https://daily.dev","name":"daily.dev","publisher":{"@id":"https://daily.dev/#organization"},"potentialAction":{"@type":"SearchAction","target":{"@type":"EntryPoint","urlTemplate":"https://daily.dev/search?q={search_term_string}"},"query-input":"required name=search_term_string"}}]} {"@context":"https://schema.org","@type":"TechArticle","headline":"Running Kimi K3 on MI355X at Better Performance per Dollar Than B300","url":"https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd"},"datePublished":"2026-08-02T06:02:15.866Z","dateModified":"2026-08-03T20:46:02.857Z","description":"Wafer benchmarks Kimi K3, a 2.8T-parameter open source LLM, on AMD MI355X GPUs and finds it delivers 952 tok/s per node at ~48 tok/s per dollar — far ahead of...","image":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/48775388813736c2d79549317b215621?_a=AQAEuop","thumbnailUrl":"https://media.daily.dev/image/upload/f_auto,q_auto/v1/posts/48775388813736c2d79549317b215621?_a=AQAEuop","isAccessibleForFree":true,"articleSection":"Hacker News","inLanguage":"en","publisher":{"@type":"Organization","name":"daily.dev","url":"https://daily.dev","logo":{"@type":"ImageObject","url":"https://daily.dev/apple-touch-icon.png","width":180,"height":180}},"author":{"@type":"Organization","name":"Hacker News","logo":"https://media.daily.dev/image/upload/t_logo,f_auto/v1/logos/hn","url":"https://daily.dev/sources/hn"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/running-kimi-k3-on-mi355x-at-better-performance-per-dollar-than-b300-bzrw6kipd","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":0},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"ai-inference,kimi-k3,rocm","timeRequired":"PT6M"} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"Hacker News","item":"https://daily.dev/sources/hn"},{"@type":"ListItem","position":3,"name":"Running Kimi K3 on MI355X at Better Performance per Dollar Than B300"}]} ```