AMD Contemplates And Engineers Yottascale AI Compute

Under Lisa Su’s more than eleven years as AMD’s chief executive officer, the company has returned from Opteron exile to be a formidable CPU foe to Intel in the datacenter, due in large part to the innovation around its Zen microarchitecture and Epyc server chips driven by Su, chief technology officer Mark Papermaster, and countless others. And, as we have noted before, thanks to missteps by Intel itself.

In this age of AI, the vendor is now looking to carve out a larger space for itself in a GPU systems market dominated by pacesetter Nvidia.

AMD still has a way to go, but it’s driving hard to put its Instinct GPUs into a place where it’s viewed as an attractive alternative to Nvidia. Its recently announced partnership with OpenAI, which kicked off the current generative AI trend three years ago with the launch of ChatGPT and – like its competitors – is pushing the agentic AI, should only help AMD, joining it with a company that is begging for more compute power and seemingly has the money to pay for it. It follows others AI-driven partnerships for AMD, including one announced early last spring with Oracle.

It was against that backdrop that Su took the stage, giving the opening keynote at the annual Consumer Electronics Show 2026 mega-event in Las Vegas and noting the rapidly increasing demand for compute power from the AI world. AMD, she said, is the best suited and most ready to deliver that power, adding that it not only can provide the GPUs and CPUs, but also the neural processing units (NPUs) and system architecture needed to meet a demand for compute that has increased more than four times a year, keeping pace with the rise off smarter and more useful models, a jump in inferencing, and the rise of agents.

“To keep up with this compute demand, you really need the entire ecosystem to come together,” Su said. “What we like to say is, the real challenge is how do we put AI infrastructure at yottascale? That requires more than just raw performance. It starts with leadership compute, CPUs, GPUs, networking coming together. It takes an open modular rack design that can evolve over product generations. It requires high-speed networking to connect thousands of accelerators into a single unified system. And it has to be really easy to deploy.”

Enter Helios, AMD’s upcoming next-generation rack-scale platform that the CEO said is designed for the Yotta-scale era. It’s a double-wide design based on the Open Compute Project’s Open Rack Wide standard that weighs almost 7,000 pounds. Helios, which will available this year, will also be armed with AMD’s latest AI GPU, the Instinct MI455X and next-generation Epyc “Venice” server CPUs, both of which Su showed off from the keynote stage.

“For those of you who have not seen a rack before, let me tell you, Helios is a monster of a rack,” she said. “This is no regular rack, okay? It is actually more than two compact cars.”

Su also laid out AMD’s datacenter GPU roadmap for the next two years, starting with the Instinct MI455X that is due out this year. It’s a GPU socket with 320 billion transistors – 70 percent more than the MI355X, which was released about six months ago – 2 nanometer and 3 nanometer architecture, and 432 GB of HBM4 stacked memory, all connected via AMD’s 3D chip-stacking technology.

The MI455X also accelerates the performance boost for AMD’s GPUs. With the launch of the MI355X, the chip maker delivered up to three times the inference throughput of the previous generation. The MI455X provides up to 10 times more performance than the MI355X.

The trend will speed up further next year with the release of the MI500 series, is being developed now. The GPU will be built on AMD’s next-generation CDNA 6 architecture, manufactured on a 2 nanometer process technology, and use the faster HBM4E (High-Bandwidth Memory 4 Extended). When it hits the market in 2027, it will mean that AMD was able to deliver a 1,000-times increase in AI performance over four years, she said.

Helios also will feature AMD’s latest Epyc CPU, codenamed Venice, which is built on the 2-nm process technology and includes up to 256 Zen 6 cores.

“We doubled the memory and GPU bandwidth from our prior generation so Venice can feed MI455 with data at full speed, even at rack-scale,” Su said. “This is really about co-engineering. We tie it all together with our 800 Gb/sec Ethernet Pensando “Vulcano” and “Salina” networking chips, delivering ultra-high bandwidth as well as ultra-low latency.”

Each liquid-cooled tray within the Helios system will include four MI455X GPUs, Venice CPUs, and Pensando Vulcano and Salina networking chips.

“Tens of thousands of Helios racks can scale across the datacenter,” Su said. “That means that each Helios rack has more than 18,000 CDNA5 GPU compute units and more than 4,600 Zen 6 CPU cores delivering up to 2.9 exaflops of performance. Each rack also includes 31 terabytes of HBM4 memory, an industry-leading 260 terabytes-per-second of scale-up bandwidth, and 43 terabytes-per-second of aggregate scale-out bandwidth to move data in and out incredibly fast.”

AMD is taking is swings at Nvidia and has a roadmap that says it will continue to do so. But it will be swinging at a fast-moving market leader. As we wrote this week, Nvidia at CES is laying out the stats for its own upcoming “Rubin” GPU accelerators and “Vera” Arm-based server CPUs, not to mention is scale-up NVLink memory fabric and scale-out Spectrum Ethernet interconnect, not to mention new NICs and DPUs.

Let the CES games begin.