{"slug": "warning-shots-fired-as-amd-announces-new-data-center-gpus", "title": "Warning Shots Fired as AMD Announces New Data Center GPUs", "summary": "AMD CEO Lisa Su announced new data center GPUs and CPUs at the Advancing AI conference, including the Helios 72-GPU rack system with MI455 GPUs delivering up to 18× more tokens per dollar versus MI355. Anthropic will buy up to 2 GW of Helios racks, with 1 GW delivered by the first half of 2027, and AMD will invest $5 billion in Anthropic as part of the deal.", "body_md": "SAN FRANCISCO, Calif. – AMD CEO Lisa Su headed up disclosures at the company’s Advancing AI developer conference that included competitive new CPU and GPU chips and racks, a partnership with Cerebras for disaggregated inference, and a major deployment at frontier lab Anthropic.\n\nThis Advancing AI event is AMD’s answer to Nvidia’s GTC, complete with a two-hour CEO keynote and vibrant show floor, but without the Jensen-style theatrics. Su’s style is at the opposite end of the spectrum; quietly confident as she unveiled graph after graph showing off the company’s new CPUs and GPUs with notable performance advantages over the green team. AMD’s keynote is not the Lisa Su Show; customer after customer was wheeled out to sing AMD’s praises before other AMD executives came on to share the spotlight, notably AMD senior VP Vamsi Boppana, who presented significant progress on AMD’s software stack, ROCm.\n\nWhile the style may have been different, the message was clear: AMD is stepping out of Nvidia’s shadow with this next generation of CPUs and GPUs and aiming for a clear leadership position in terms of performance. The company has notable deployments for its GPU racks with major frontier labs, and any perceived technology moat Nvidia had is shrinking fast.\n\n“There’s no one company that can solve it all, but this is the opportunity to bring the best and brightest together, and we have never been in a better position to lead,” Su said. “We have the broadest product portfolio, we have the strongest roadmaps we’ve ever had, and the thing that I’m most proud of is we have the deepest partnerships with the companies that are building this future.”\n\n[View All](https://www.eetimes.com/category/sponsored-content/)\n\nSu unveiled Helios, AMD’s 72-GPU rack system, intended to compete with Nvidia’s Vera Rubin NVL72 racks.\n\n“It’s simply the best AI rack in the world,” she said.\n\nHelios uses AMD’s new MI455 GPU, which includes12 compute and I/O chiplets on TSMC 2 nm and 3 nm, offering 20 PFLOPS FP8 compute with 432 GB of HBM4 memory. MI455 delivers 4× the token throughput of MI355, but what customers really care about is cost per token, Su said.\n\n“With every generation of Instinct [GPUs], we’re driving the cost per token down, and that’s with more memory, more bandwidth, and much more compute,” she said. “We’re taking another major step with MI455, delivering up to 18× more tokens per dollar [versus MI355], so customers can serve far more users with the same investment.”\n\nFor the same power budget, Helios offers 10 to 15× better performance than the competition, achieving up to 30% more tokens per dollar versus Nvidia’s Vera Rubin NVL72 for Kimi K2.\n\n“Customer demand for Helios is extremely strong,” Su said. “From the largest AI labs to hyperscale and real-time providers, we are working across the entire ecosystem to enable Helios together with our OEM and ODM partners.”\n\n**Anthropic deal**\n\nAnthropic co-founder Tom Brown joined Su on stage after being unveiled as a customer days earlier. This deal will see Anthropic buy up to 2 GW of Helios racks, with 1 GW to be delivered in the first half of 2027. AMD will also invest $5 billion in Anthropic as part of the deal.\n\n“The scale that we’re growing as an industry is enormous,” Brown said. “So, we have been working to make sure that we have the best of different chips, and we can use the best chips for the best workloads.”\n\nBringing up a new hardware platform is usually a headache, Brown said, but with AMD’s previous-generation MI355, a single engineer connected the rack to Anthropic’s Claude AI assistant. After asking Claude to bring up the machine, it was left running over the weekend to continue to optimize performance.\n\n“I think that that’s a testament to the open platform that [AMD has] made where anyone, human or AI, can now build real models on [AMD’s] platform,” Brown said.\n\nThese words were carefully chosen. The biggest part of Nvidia’s moat is a combination of its installed base and its CUDA software stack. AMD closing the gap on CUDA with its ROCm stack is hugely significant. (More on ROCm below.)\n\nOpenAI head of industrial compute Sachin Katti came on to talk about the 6 GW of AMD hardware being deployed on its behalf. OpenAI has had Helios racks in its data centers for three months, he said, and GPT-class workloads are up and running on MI455s.\n\n“We’re really excited about the capabilities Helios is already showing us, and we expect that we’ll be deploying Helios at massive scale starting towards the end of this year and accelerating towards 2027,” Katti said.\n\nKatti spoke about recursion—where AI designs the systems it runs on and can even do its own research. AI is showing the potential to automate a significant portion of kernel writing, model-hardware optimisation, compilers, communications, libraries and system considerations, he said.\n\n“[AI has] dramatically shortened the path from a new model coming out to the end-to-end production, and something that we can much more easily tune to changes in the workload, and that’s a big deal,” he said.\n\nThis is in part enabled by AMD’s open software ecosystem, he noted, without saying that Nvidia’s CUDA is famously less open. OpenAI, also not noted for its commitment to open source, is inventing internally but wants to allow the whole world to use its advancements on all AI models, not just its own, Katti said.\n\n**Inference disaggregation**\n\nOne of Nvidia’s most surprising moves this year has been to all-but acquire startup Groq in order to speed up portions of inference that GPUs don’t do well. In a similar move, AMD is now working with Groq’s fast token competitor Cerebras to improve throughput on Cerebras’ low-latency tokens; the combination can offer 5× more throughput on the fastest tokens.\n\nBig GPUs like MI455 offer tremendous throughput, which is a compute-bound problem, while non-GPU competitors like Cerebras have been able to excel at latency, which is a memory-bound problem. A balance of throughput and latency is required for a new class of applications that needs very fast tokens, Su said. The best way to do this in a GPU-based system today is to have GPUs run the prefill stage of LLM inference, with a dedicated accelerator for the decode stage.\n\n“You can actually take this a step further,” Su said. “If you know what workloads you’re trying to run, you can actually let customers tune each of these pieces independently.”\n\nCerebras CEO Andrew Feldman, formerly a vocal critic of Nvidia’s disaggregated inference approach, said that the combination of Helios’ memory capacity combined with Cerebras’ SRAM bandwidth would allow the companies to build an unmatched solution.\n\n“To serve this ultra-low latency market, we saw a partnership that could extend our capabilities and our footprint, and there was no better answer than AMD,” he said.\n\nJoint AMD-Cerebras solutions will deploy later this year in the Cerebras cloud, presumably responding to demands from large Cerebras customers including OpenAI and AWS.\n\n“Until recently, customers could have high throughput or extraordinary speed,” Feldman said. “By bringing together the Helios rack with the Cerebras wafer-scale engine, we give you five times the throughput while continuing to deliver this extraordinary speed. It’s really something amazing.”\n\nAsked about the future of the partnership during the press Q&A, Su said the Cerebras partnership was not merely a stopgap while AMD works on its own decode acceleration chip.\n\n“We don’t start anything with the notion that it’s going to be a stopgap,” she said. “Cerebras has a very interesting technology, and it works very well with Helios. I think the idea of our open ecosystem is frankly that we will work with a number of different companies that may have technology that can be useful.”\n\nWe can expect to see more workload disaggregation going forward, with the rate and pace of developments driven by cost structures and price points, according to Su.\n\n**Venice CPUs**\n\nServer CPUs for agentic AI run two main workloads. In GPU servers, host CPUs need speed—the highest frequency cores and fastest I/O—to keep GPUs fully fed. In agentic servers, there is a new workload AMD calls “sandboxes”—executing code, calling tools, querying data outside the model—so the priority is density and the highest performance per watt. These categories are distinct from general-purpose enterprise servers, which need to run a range of workloads.\n\n“Epyc is the only CPU portfolio that leads across all three [workloads],” Su said.\n\nVenice, with its new generation of core, will use a new chiplet that supports up to 512 threads per socket to offer the largest generation performance uplift in the history of Epyc, Su said.\n\nAMD is using its chiplet-based architecture to its advantage, developing a family of Venice CPUs around the same tapeouts. A version of Venice for highest performance on AI host nodes has eight compute chiplets, each with 12 cores running at up to 5 GHz. This is the CPU host used in Helios; it outperforms its x86 competition by 1.8× the tokens per second.\n\nA 256-core version of Venice is for agentic sandboxes. It has eight compute chiplets, each with 32 cores, which scales to 512 threads. This CPU offers the highest compute density in the industry, Su said, leading to twice the agents per watt versus the nearest x86 competitor.\n\nA 128-core Venice CPU is for general-purpose CPU servers, offering twice the performance per Watt versus x86 competitors for a 100 kW rack.\n\nThe gap is even bigger for Arm-based CPUs, Su said, even on performance per Watt. Venice CPUs are in full production today.\n\nContinuing the naming tradition of Italian cities (Naples, Rome, Turin, and Genoa came before Venice), AMD’s next-generation AI host CPUs will be named “Verano” after a suburb of Rome, presumably picked solely for its resemblance to “Vera No” (Vera is Nvidia’s ARM-based GPU-host CPU). Verano will host MI500s in the next generation of Helios in 2027.\n\nNext-generation Venice CPUs, coming in 2027, will use AMD’s 3D vCache memory-stacking technology.\n\n**AI software stack**\n\nAMD has been working hard on its software stack. The company’s strategy combines partnering deeply with the open-source ecosystem, which brings velocity, with building the right layers of abstraction to enable developer productivity, said Vamsi Boppana, senior VP of the AI group at AMD.\n\nThe hard work is paying off. AMD is becoming part of the default enablement for the most important AI communities, including HuggingFace, PyTorch, JAX, vLLM, and SGLang. This is why new models now get zero-day support for AMD, Boppana said.\n\nThe AI revolution in software development is coming to GPU programming, Boppana said.\n\n“Over the past year, I’ve seen something remarkable inside AMD,” Boppana said. “Our engineers are using AI models to generate GPU kernels, optimize code, debug issues, and improve performance. In some cases, these AI-generated kernels are shockingly good. Better than what we expected. Sometimes better than the most finely tuned versions. The first time you see this, you’re a little bit skeptical. You run more tests, you try to break it, you look for what went wrong, and then you realize, this is real.”\n\nAMD unveiled ROCm.AI, a new version of ROCm which connects ROCm to Cursor, Claude, Codex, and other AI-assisted coding platforms.\n\n“ROCm.AI helps agents understand AMD platforms, understand ROCm, and helps you build and optimize your workloads,” Boppana said. “In other words, we are making those popular coding agents into ROCm super-users. You should be able to describe the workload you want to run, the performance target you want to hit, and then let the agents help you get there.”\n\nAI agents profiled the workload, proposed new kernels, tested configurations, and actually validated all of the results. This approach improved inference performance on large models like DeepSeek-R1 by 3.3× versus the previous version of ROCm, and training by 2.4×.\n\n**Roadmap**\n\nSu also briefly teased AMD’s roadmap.\n\nNext-generation CPUs will be called Florence, with a new core architecture and new AI extensions, supporting the latest memory technologies. These are due in 2028, along with Ferrara, a rack-scale AI host node, and Fidenza, an agentic sandbox node. CPUs with a next-gen core are already under development for 2030.\n\n“It’s all about product breadth, and having the right foundation, and also the right family for each workload,” Su said. “Florence, Ferrara and Fidenza are built on a full family of CPUs that is optimized for these different workloads.”\n\nAMD’s Instinct GPUs remain on their annual product cadence. MI500, coming next year, will have next-generation HBM4e memory and will enable larger scale-up domains using both copper and optical interconnects. If AMD goes for co-packaged optics (CPO) with this GPU generation, that would be ahead of Nvidia’s published plans to move to CPO with its Feynman generation in 2028.\n\n“MI500 will deliver the largest generational leap in the history of Instinct, putting us on track to deliver more than 2,000 times higher inference throughput in just four years,” Su said. “We are working with a number of our customers already on this design point. MI500 is super exciting, and the feedback we’re getting from customers is just fantastic.\n\nMI600, coming in 2028, will be powered by a future generation of the CDNA architecture.\n\nSu ended by talking about entering the [next phase of AI](https://www.embedded.com/what-will-shape-robotics-in-the-next-3-years/), where we really start to see a real-world impact.\n\n“When I think about where AI is today, the biggest change I’m seeing is we’re no longer talking about what might be possible,” Su said. “We’re actually seeing how AI can have real and significant impact across every industry and every part of our personal lives… I’ve spent my entire career in tech believing that high-performance computing can make the world an incredibly better place, and I have never, ever believed that more than I do today.”\n\n##### Read also:\n\n[What It Takes to Build an AI Chip Startup in Europe](https://www.eetimes.com/podcasts/what-it-takes-to-build-an-ai-chip-startup-in-europe/)\n\n[Etched Raises $300M with $1B in Pre-Orders](https://www.eetimes.com/etched-raises-300m-with-1b-in-pre-orders/)\n\n[From Rhetoric to Metrics: Raghib Hussain’s First Year as Altera CEO](https://www.eetimes.com/from-rhetoric-to-metrics-raghib-hussain-first-year-as-altera-ceo/)", "url": "https://wpnews.pro/news/warning-shots-fired-as-amd-announces-new-data-center-gpus", "canonical_source": "https://www.eetimes.com/warning-shots-fired-as-amd-announces-new-data-center-gpus-at-advancing-ai-event/", "published_at": "2026-07-27 16:58:23+00:00", "updated_at": "2026-07-27 17:28:41.262235+00:00", "lang": "en", "topics": ["artificial-intelligence", "ai-chips", "ai-infrastructure", "ai-products", "ai-startups"], "entities": ["AMD", "Lisa Su", "Anthropic", "Tom Brown", "Nvidia", "Cerebras", "Helios", "MI455"], "alternates": {"html": "https://wpnews.pro/news/warning-shots-fired-as-amd-announces-new-data-center-gpus", "markdown": "https://wpnews.pro/news/warning-shots-fired-as-amd-announces-new-data-center-gpus.md", "text": "https://wpnews.pro/news/warning-shots-fired-as-amd-announces-new-data-center-gpus.txt", "jsonld": "https://wpnews.pro/news/warning-shots-fired-as-amd-announces-new-data-center-gpus.jsonld"}}