--- title: Why One GPU Isn't Enough: The Rise of AMD Helios | daily.dev description: AMD unveiled Helios, its first rack-scale reference design combining 72 Instinct MI455X GPUs (with HBM4 memory), EPYC 'Venice' CPUs, ULink GPU interconnect,... canonical: https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb twitter:card: summary_large_image twitter:site: @dailydotdev og:type: website og:site_name: daily.dev og:title: Why One GPU Isn't Enough: The Rise of AMD Helios | daily.dev og:description: AMD unveiled Helios, its first rack-scale reference design combining 72 Instinct MI455X GPUs (with HBM4 memory), EPYC 'Venice' CPUs, ULink GPU interconnect,... og:url: https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb og:image: https://api.daily.dev/og/posts/HZLWCC9jB.png og:image:alt: Why One GPU Isn't Enough: The Rise of AMD Helios 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. # Why One GPU Isn't Enough: The Rise of AMD Helios **[ByteMonk](https://daily.dev/sources/bytemonk)** · 9 min read · 2 upvotes · 0 comments ## Summary AMD unveiled Helios, its first rack-scale reference design combining 72 Instinct MI455X GPUs (with HBM4 memory), EPYC 'Venice' CPUs, ULink GPU interconnect, and Pensando Ethernet-based networking built on Ultra Ethernet standards. It's positioned as a challenger to Nvidia's tightly integrated NVLink/NVSwitch-based NVL72 racks and CUDA software moat. Microsoft's upcoming Azure service and Oracle have committed to running production AI workloads on Helios, a strong validation signal. Separately, OpenAI—already committed to deploying up to 6GW of AMD GPUs—is also building its own inference-focused chip, Jalapeno, with Broadcom, reporting better throughput-per-watt and lower latency than commercial systems in its own benchmarks. The piece frames AI compute as splitting into specialized systems for training vs. inference rather than one GPU architecture dominating everything. ## Full article daily.dev links to this article rather than hosting it. Read it at the original source: ## Questions this post answers ### What is AMD's Helios rack-scale system and what GPUs does it use? Helios is AMD's first rack-scale reference design, packing 72 Instinct MI455X GPUs with HBM4 memory, EPYC 'Venice' CPUs as hosts, UALink for GPU-to-GPU interconnect within the rack, and Pensando networking with Ultra Ethernet support between racks. The full rack offers roughly 31 terabytes of memory, about 50% more per GPU than Nvidia's upcoming Vera Rubin system. _Teams comparing rack-scale AI hardware options can track how these architectures evolve on daily.dev._ ### Is Microsoft actually using AMD Helios in production? Yes, Microsoft confirmed an upcoming Azure service will run AI models at production scale on Helios, following Oracle's earlier commitment to the system. This matters because Azure handles some of the most demanding AI workloads, making Microsoft's adoption a significant validation of AMD's rack-scale approach beyond just GPU sales. _Engineers weighing cloud AI infrastructure vendors can follow this validation story on daily.dev._ ### Why is OpenAI building its own chip called Jalapeno if it already buys GPUs from AMD and Nvidia? OpenAI is building Jalapeno with Broadcom specifically for inference workloads, where latency, throughput, and power efficiency matter most, while still relying on AMD and Nvidia GPUs for training and other needs. OpenAI's own benchmarks reported better throughput per kilowatt and lower latency than commercial systems, though it continues deploying accelerators from multiple partners rather than replacing them. _Anyone deciding between GPU vendors and custom silicon for inference can follow these tradeoffs on daily.dev._ ## Similar posts on daily.dev - [AMD’s Helios puts 72 GPUs and 31 terabytes of HBM4 in one rack. It is AMD’s answer to Nvidia’s NVL72.](https://daily.dev/posts/amd-s-helios-puts-72-gpus-and-31-terabytes-of-hbm4-in-one-rack-it-is-amd-s-answer-to-nvidia-s-nvl72-rt5ma3xnm) · The Next Web · 1 upvotes · 0 comments - [AMD puts 32,768 CPU cores in a rack built for AI agents](https://daily.dev/posts/amd-puts-32-768-cpu-cores-in-a-rack-built-for-ai-agents-zwab5jr9l) · The Next Web · 0 upvotes · 0 comments - [The infrastructure lock-in costing AI companies hundreds of millions](https://daily.dev/posts/the-infrastructure-lock-in-costing-ai-companies-hundreds-of-millions-wuyswiobz) · The New Stack · 0 upvotes · 0 comments --- Tags: [#openai](https://daily.dev/tags/openai), [#nvidia](https://daily.dev/tags/nvidia), [#amd](https://daily.dev/tags/amd), [#ai-infrastructure](https://daily.dev/tags/ai-infrastructure) [View this post on daily.dev](https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb) ```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":"Why One GPU Isn't Enough: The Rise of AMD Helios","url":"https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb","mainEntityOfPage":{"@type":"WebPage","@id":"https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb"},"datePublished":"2026-08-31T04:19:49.689Z","dateModified":"2026-09-13T19:22:08.640Z","description":"AMD unveiled Helios, its first rack-scale reference design combining 72 Instinct MI455X GPUs (with HBM4 memory), EPYC 'Venice' CPUs, ULink GPU interconnect,...","image":"https://i.ytimg.com/vi/abDpesRn-0U/sddefault.jpg","thumbnailUrl":"https://i.ytimg.com/vi/abDpesRn-0U/sddefault.jpg","isAccessibleForFree":true,"articleSection":"ByteMonk","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":"ByteMonk","logo":"https://media.daily.dev/image/upload/s--n36_-Uve--/f_auto/v1752670492/logos/bytemonk","url":"https://daily.dev/sources/bytemonk"},"commentCount":0,"discussionUrl":"https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb","interactionStatistic":[{"@type":"InteractionCounter","interactionType":{"@type":"LikeAction"},"userInteractionCount":2},{"@type":"InteractionCounter","interactionType":{"@type":"CommentAction"},"userInteractionCount":0}],"keywords":"openai,nvidia,amd,ai-infrastructure","timeRequired":"PT9M","video":{"@type":"VideoObject","name":"Why One GPU Isn't Enough: The Rise of AMD Helios","description":"AMD unveiled Helios, its first rack-scale reference design combining 72 Instinct MI455X GPUs (with HBM4 memory), EPYC 'Venice' CPUs, ULink GPU interconnect,...","thumbnailUrl":"https://i.ytimg.com/vi/abDpesRn-0U/sddefault.jpg","uploadDate":"2026-08-31T04:19:49.689Z","duration":"PT9M","url":"https://api.daily.dev/r/HZLWCC9jB","embedUrl":"https://www.youtube.com/embed/abDpesRn-0U"}} {"@context":"https://schema.org","@type":"BreadcrumbList","itemListElement":[{"@type":"ListItem","position":1,"name":"Home","item":"https://daily.dev"},{"@type":"ListItem","position":2,"name":"ByteMonk","item":"https://daily.dev/sources/bytemonk"},{"@type":"ListItem","position":3,"name":"Why One GPU Isn't Enough: The Rise of AMD Helios"}]} {"@context":"https://schema.org","@type":"FAQPage","@id":"https://daily.dev/posts/why-one-gpu-isn-t-enough-the-rise-of-amd-helios-hzlwcc9jb#faq","mainEntity":[{"@type":"Question","name":"What is AMD's Helios rack-scale system and what GPUs does it use?","acceptedAnswer":{"@type":"Answer","text":"Helios is AMD's first rack-scale reference design, packing 72 Instinct MI455X GPUs with HBM4 memory, EPYC 'Venice' CPUs as hosts, UALink for GPU-to-GPU interconnect within the rack, and Pensando networking with Ultra Ethernet support between racks. The full rack offers roughly 31 terabytes of memory, about 50% more per GPU than Nvidia's upcoming Vera Rubin system. Teams comparing rack-scale AI hardware options can track how these architectures evolve on daily.dev."}},{"@type":"Question","name":"Is Microsoft actually using AMD Helios in production?","acceptedAnswer":{"@type":"Answer","text":"Yes, Microsoft confirmed an upcoming Azure service will run AI models at production scale on Helios, following Oracle's earlier commitment to the system. This matters because Azure handles some of the most demanding AI workloads, making Microsoft's adoption a significant validation of AMD's rack-scale approach beyond just GPU sales. Engineers weighing cloud AI infrastructure vendors can follow this validation story on daily.dev."}},{"@type":"Question","name":"Why is OpenAI building its own chip called Jalapeno if it already buys GPUs from AMD and Nvidia?","acceptedAnswer":{"@type":"Answer","text":"OpenAI is building Jalapeno with Broadcom specifically for inference workloads, where latency, throughput, and power efficiency matter most, while still relying on AMD and Nvidia GPUs for training and other needs. OpenAI's own benchmarks reported better throughput per kilowatt and lower latency than commercial systems, though it continues deploying accelerators from multiple partners rather than replacing them. Anyone deciding between GPU vendors and custom silicon for inference can follow these tradeoffs on daily.dev."}}]} ```