AMD Takes On Nvidia with MI455X GPUs and Helios Racks

retained text

Genesis Mission Summit 2026: Big Ambitions and a Familiar Blind Spot
The first Genesis Mission Summit was, by design, a day about momentum. Conner Prochaska, ARPA-E...
During its Advancing AI 2026 launch event in San Francisco yesterday, AMD formally unveiled its latest generation GPUs, including the AI-oriented Instinct MI455X and the HPC-oriented Instinct MI430X. Surrounding the GPUs will be new Zen6 CPUs and a new 800Gb Pensando NIC, providing UALink-compatible scale-up Helios rack to go up against Nvidia’s proprietary NVLink systems.
The new Instinct MI455X GPU is geared at emerging AI workloads, including AI model training and serving AI agents in production. To that end, it’s loaded with lower precision CDNA cores that drive 40 petafalops of FP4 performance and 20 petaflops of FP8 performance. The chips, which are built at TSMC on 2nm and 3nm processes and feature 320 billion transistors, contain 432 GB of HBM4 capacity and deliver 23.2 TB per second of memory bandwidth.
On paper, the GPU figures stack up well against Nvidia latest Rubin chip, which delivers up to 50 petaflops of FP4 inference performance, 35 petaflops of PF4 training performance, and anywhere from 4 to 17.5 petaflops of FP8 performance, on top of 288 GB of HBM4 memory and 22 TB per second of memory bandwidth.
But sheer GPU horsepower does not an AI juggernaut make. Getting the most out of those GPUs requires a balanced system that’s capable of moving data at high speeds, in addition to processing it quickly.
To that end, AMD had several other cards up its sleeves at Advancing AI 2026 to build a complete scale-up rack based on its new Helios architecture, including new CPUs. As AMD CEO Lisa Su showed today during her keynote, AMD is packaging its new chiplet-based MI455X accelerators into what AMD calls an Enhanced Accelerator Module (EAM).
“It’s a production module that goes into Helios and it is much more than the GPU package,” Su said. “The EAM actually integrates the GPU, memory, power delivery, high speed interfaces, system management, and cold plates for liquid cooling, all into one compact serviceable module.”
The EAM is being paired with the new sixth-generation EPYC 9006 CPUs, codenamed “Venice,” which feature up to 512-core processors and DDR5 memory. It’s also being paired with AMD’s latest generation Pensando network interface cards (NICs), the Pensando Vulcano 800, which delivers 800Gb of scale-up network capability. AMD is also launching Pensando Salina data processing units (DPUs), which handle front-end network virtualization, security, and KV cache management at 400Gb speeds.
Each EAM contains a single MI455X GPU and a single EPYC “Venice” CPU, for a total of 12 compute and I/O chiplets and 432GB of stacked HBM4 memory. AMD is putting four EAMs in each Helios compute tray, and it’s putting 18 compute trays in a single rack, which means each Helios rack has 72 GPUs and 72 CPUs.
On the networking side, AMD is placing six Vulcano NICs on each board, with two Vulcano NIC boards per Helios tray, providing up to 2.4 Tbps of bandwidth per GPU via UALink for scale-up reach.
“Each Helios compute tray includes two of these Vulcano boards, and that allows us to have the scale out and scale up networking overall,” Su said. “You put all that inside the rack, and we have an additional six dedicated networking trays that handle the scale up networking. And those are all connecting 72 GPUs together with UALink over Ethernet with silicon from our partners, and that’s what we mean by an open ecosystem.” (NVLink, by comparison, is a proprietary interface owned by Nvidia for scale-up networking.)
All told, each Helios rack features 18,000 CDNA5 GPU compute units, more than 4,600 Zen6 CPU cores, and 31TB of HBM, according to Su.
“That’s what it takes to run agentic AI at scale,” she said. “I can say for sure it’s the highest performance AI accelerator in the industry. When you compare Helios to the competition, we’re delivering 15% more compute, 50% more HBM for memory capacity and memory bandwidth, and 50% more scale out bandwidth. And what that means is that every Helios can deliver more performance for the largest models, more capacity for longer context, and the bandwidth to scale across thousands of racks.”
Helios is in full production. AMD is on track to start deliveries at the end of the third quarter, ramping up into the fourth quarter. Among the customers are Anthropic and Microsoft. Anthropic plans to deploy 2 gigawatts worth of Helios, while Microsoft plans to deploy Helios at scale in its Azure cloud.
AMD also had an update on the Instinct MI430X, the HPC-oriented GPU that will be going into several upcoming supercomputers, including the new Discovery leadership-class supercomputer at Oak Ridge National Lab and Alice Recoque, the new exascale machine going into the CEA’s TGCC supercomputing center in France.
When it ships in early 2027, the new MI430X will feature 288 teraflops of double-precision FP64 compute, 432 GB of HBM4 capacity, and 23.3 TB per second of memory bandwidth. That is more native FP64 capacity (i.e. not emulated using Ozaki scheme), more HBM, and more memory bandwidth than we anticipated the new chip would have when we wrote about it back in March.
“Just like MI455 leads in AI compute with FP4 and memory performance, MI430x leads HPC compute with FP64 and memory performance,” said Vamsi Boppana, senior vice president of the Artificial Intelligence Group at AMD.
“This is hardware double precision, not emulation, for the scientific community,” Boppana said. “That distinction is everything. This product delivers 288 teraflops of FP64 compute, nearly nine times better than the competition, and the same leadership in memory capacity and bandwidth of MI455.”

document fetched 2026-07-25 07:57 · canonical https://hpcwire.com/2026/07/24/amd-takes-on-nvidia-with-mi455-gpus-and-helios-racks