{"id":700,"date":"2026-09-06T15:15:57","date_gmt":"2026-09-06T18:15:57","guid":{"rendered":"https:\/\/www.linuxpro.com.br\/2026\/09\/amd-vs-nvidia-fosso-cuda\/"},"modified":"2026-09-06T15:28:45","modified_gmt":"2026-09-06T18:28:45","slug":"amd-vs-nvidia-fosso-cuda","status":"publish","type":"post","link":"https:\/\/www.linuxpro.com.br\/en\/2026\/09\/amd-vs-nvidia-fosso-cuda\/","title":{"rendered":"AMD beats Nvidia on every specification \u2014 and still loses"},"content":{"rendered":"<p><img loading=\"lazy\" decoding=\"async\" src=\"https:\/\/www.linuxpro.com.br\/wp-content\/uploads\/2026\/09\/capa-amd-vs-nvidia.webp\" alt=\"AMD vs Nvidia: o fosso do CUDA\" width=\"1020\" height=\"620\" \/><\/p>\n<blockquote>\n<p><strong>Video summary:<\/strong> this article is a detailed summary of the video <em>\u201cAMD Beats Nvidia On Every Spec And Still Loses\u201d<\/em>, from the channel <strong>The Stack<\/strong>.<br \/>\n<strong>Video URL:<\/strong> <a href=\"https:\/\/www.youtube.com\/watch?v=bOHhUVSDmrw\" rel=\"noopener\" target=\"_blank\">https:\/\/www.youtube.com\/watch?v=bOHhUVSDmrw<\/a><\/p>\n<\/blockquote>\n<div class=\"video-container\" style=\"position:relative;padding-bottom:56.25%;height:0;overflow:hidden;margin:1.5em 0;\">\n  <iframe src=\"https:\/\/www.youtube.com\/embed\/bOHhUVSDmrw\" title=\"AMD Beats Nvidia On Every Spec And Still Loses\" frameborder=\"0\" allow=\"accelerometer; autoplay; clipboard-write; encrypted-media; gyroscope; picture-in-picture; web-share\" allowfullscreen style=\"position:absolute;top:0;left:0;width:100%;height:100%;\"><\/iframe>\n<\/div>\n<p>On paper, there should not even be a contest: the accelerator <strong>AMD MI300X<\/strong> carries <strong>192 GB of memory versus 80 GB<\/strong> on Nvidia's H100, moves <strong>5,3 TB\/s<\/strong> of bandwidth, and delivers <strong>1,31 petaflop<\/strong> \u2014 winning line by line on the datasheet. AMD itself raised the stakes in an official announcement: the <strong>Instinct MI325X outperforms the H200<\/strong>, precisely the chip the industry was lining up to buy. Even so, Nvidia closed 2025 with <strong>92% of the enterprise GPU market<\/strong> \u2014 of every 100 GPUs in server racks, 92 are green. The Stack's video dissects why better specifications lose to network effects \u2014 and the answer is not in the silicon, it is in the software.<\/p>\n<h2>The moat is called CUDA<\/h2>\n<p>Anyone buying a cluster worth tens of millions of dollars is not buying a chip \u2014 they are buying compatibility. Nvidia launched <strong>CUDA<\/strong> in <strong>2007<\/strong>, when GPUs for general-purpose computing were still an odd bet. Almost twenty years later, that bet became the industry's most effective castle moat: <strong>PyTorch, TensorFlow, and JAX<\/strong> \u2014 the three frameworks supporting modern AI \u2014 assume CUDA <em>by default<\/em>. The result: an estimated <strong>90% of all AI training workloads<\/strong> run on Nvidia.<\/p>\n<p>AMD's answer, <strong>ROCm<\/strong>, only arrived in <strong>2016<\/strong> \u2014 nine years late. In hardware, missing a window costs one product cycle; in software, nine years is a geological era: enough time for the incumbent to dictate how an entire generation of engineers learns to program, how universities teach, and how the surrounding tooling is built. Then there is the gravity of the installed base: with <strong>half the world's installed AI chips<\/strong> on Nvidia silicon, every optimization trick and every weekend project on GitHub is born tested on the architecture the developer has at hand \u2014 making the platform even safer for the next buyer, who expands the base again. It is a self-reinforcing cycle.<\/p>\n<h2>The three-headed hydra<\/h2>\n<p>Analysts describe Nvidia's structural advantage as a <strong>three-headed hydra<\/strong>: the chip is only one head. The second is the <strong>network layer<\/strong> that stitches thousands of GPUs into a single cluster (NVLink, InfiniBand); the third is the software stack on top of it all. A competitor matching its computing power still needs to beat the other two heads \u2014 and that is where the math falls apart.<\/p>\n<h2>Five months of benchmarking, and a painful verdict<\/h2>\n<p>The most forceful part comes from <strong>SemiAnalysis<\/strong>, which spent <strong>five months<\/strong> testing AMD hardware with direct support from engineers at both manufacturers \u2014 in other words, giving AMD every chance to optimize. The conclusion: the silicon is as fast as advertised, but the software stack is so <em>\u201criddled with bugs\u201d<\/em> that the time saved on hardware evaporated while chasing strange errors, turning extremely expensive engineers into full-time IT support. The published verdict was blunt: <strong>AMD has not yet crossed the CUDA moat<\/strong>. Buying the most powerful chip gets expensive when the real cost is skilled people's hours spent putting out fires.<\/p>\n<h2>Attacking the moat from below<\/h2>\n<p>If the problem is software, software can be fixed \u2014 but rewriting code solves only half the equation; the other half is twenty years of accumulated habit. Two efforts attack exactly that:<\/p>\n<ul>\n<li><strong>SCALE<\/strong> (Spectral Compute) \u2014 a compatibility layer that takes your <strong>CUDA binaries without any modification<\/strong> and runs them on AMD GPUs. Translation happens at compile time: PyTorch believes it is talking to an Nvidia chip while an MI300X does the computation underneath. No migrating models, maintaining two codebases, or retraining the team<\/li>\n<li><strong>CANN<\/strong> (Huawei) \u2014 the toolkit for Ascend accelerators, <strong>opened in August 2025<\/strong> specifically to test whether giving the stack away beats developers' muscle memory<\/li>\n<\/ul>\n<p>If \u201crun CUDA anywhere\u201d becomes reliably real, the moat stops protecting the castle \u2014 the same move Wine\/Proton made with Windows games on Linux.<\/p>\n<h2>The market is already cracking \u2014 but not for AMD<\/h2>\n<p>Nvidia's dominance fell from its peak of <strong>87% to a projection of 75% in 2026<\/strong> \u2014 12 points of the planet's most profitable hardware sector looking for a new home. The catch: AMD is not capturing the difference. Even with a quarter of <strong>US$ 5,8 billion<\/strong> in data centers, its volume does not explain the hole in the spreadsheet. What explains it is the <strong>in-house chips<\/strong> from Google (TPU), Meta, Amazon, and Microsoft, growing <strong>44,6% per year versus 16,1%<\/strong> for off-the-shelf GPUs.<\/p>\n<p>And the homegrown chip has an unfair software advantage: it only needs to run <em>the company's own internal models<\/em>. It does not need to support PyTorch for the world or convince the global market to change habits \u2014 its target audience is a single building, and internal engineers optimize the closed stack and ignore the ecosystem. The real threat to Nvidia is not its usual competitor \u2014 it is its biggest customers ceasing to be customers.<\/p>\n<h2>The lesson for builders<\/h2>\n<p>The video's lesson applies to any infrastructure layer: <strong>ecosystem beats specifications<\/strong>. A bug can be fixed with a patch; asking an entire department to switch the tools it has mastered is harder than signing the check for a platform it already trusts. True for GPUs, languages, clouds \u2014 anyone who has migrated distros knows: the cost is never the kernel, it is everything built around it. For those of us in the Linux world, there is an upside \u2014 ROCm, CANN, and SCALE are open bets against a proprietary monopoly, and this entire battle is being fought on the penguin: virtually all AI training on the planet runs on Linux, whether the chip is green, red, or homegrown.<\/p>\n<p>The original video is worth its 17 minutes \u2014 and if you run local AI on your hardware, the post <a href=\"\/en\/2026\/09\/rodando-ia-local-no-linux-com-ollama\/\">Running local AI on Linux with Ollama<\/a> is the practical companion.<\/p>","protected":false},"excerpt":{"rendered":"<p>Video summary: this article is a detailed summary of \u201cAMD Beats Nvidia On Every Spec And Still Loses\u201d from The Stack. Video URL: https:\/\/www.youtube.com\/watch?v=bOHhUVSDmrw On paper, there should not even be a contest: the AMD MI300X accelerator carries 192 GB of memory against 80 GB on the Nvidia H100 and moves 5.3 TB\/s of bandwidth \u2026 <a title=\"AMD beats Nvidia on every specification \u2014 and still loses\" class=\"read-more\" href=\"https:\/\/www.linuxpro.com.br\/en\/2026\/09\/amd-vs-nvidia-fosso-cuda\/\" aria-label=\"Read more about AMD vence a Nvidia em todas as especifica\u00e7\u00f5es \u2014 e mesmo assim perde\">Read more<\/a><\/p>","protected":false},"author":0,"featured_media":0,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[115,25],"tags":[342,153,134,116,152,343],"class_list":["post-700","post","type-post","status-publish","format-standard","hentry","category-ia","category-noticias","tag-amd","tag-cuda","tag-gpu","tag-ia","tag-nvidia","tag-rocm"],"_links":{"self":[{"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/posts\/700","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/types\/post"}],"replies":[{"embeddable":true,"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/comments?post=700"}],"version-history":[{"count":2,"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/posts\/700\/revisions"}],"predecessor-version":[{"id":712,"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/posts\/700\/revisions\/712"}],"wp:attachment":[{"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/media?parent=700"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/categories?post=700"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.linuxpro.com.br\/en\/wp-json\/wp\/v2\/tags?post=700"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}