- AMD’s Helios beats Vera Rubin NVL72 in rack-level memory, with 31TB of HBM4 vs. about 20.7TB
- AMD also claims to outperform the competition by 15% in FP4 computing “per GPU,” even as Nvidia comes out ahead in the published “per rack” numbers.
- AMD claims that Helios also offers 30% more tokens per dollar spent compared to the competition.
AMD has released its Helios rackscale offering, outlining five comparisons in which it claims to win over what it calls “the leading competitive solution.”
While the company omitted to name Nvidia, the market leader’s Vera Rubin-based NVL72 rack-scale solution is the only real competitor to Helios and one it continues to be compared to.
While its memory claims hold up, its FP4 GPU claim could fall short when comparing rack to rack, and many of its calculations are based on peak performance rather than Nvidia’s published numbers, making the Helios an interesting “win” but one that encourages potential adopters to look closer.
Latest videos ofTechnologyRadar
A numbers game that continues to get complex even as AMD scores some wins
While AMD claims a 15% win over Nvidia’s Rubin by GPU for FP4 compute, its 72-GPU rack-scale solution falls short of Nvidia’s published rack-level numbers: 2.9 exaflops versus 3.6. AMD doesn’t specify whether it’s counting Nvidia’s individual dies or its two-die bundles, and the distinction matters: against dies the gap runs in AMD’s favor by well over 15%, while against AMD bundles it lags behind.
The two figures also use different formats, AMD’s MXFP4 versus Nvidia’s NVFP4, so they don’t measure identical arithmetic.
However, this could be indicative of a very real situation hampering both vendors’ headline numbers: real-world FP4 workloads rarely reach maximum hardware ratings due to memory movement limitations, scheduling overhead, and software kernel efficiency. AMD admitted as much at its own event, putting the FP4’s measured performance at about half of its maximum rating.
Nvidia is able to maintain more of its peak thanks to its custom Vera CPU, a mature NVLink 6 software stack, and a larger pool of what it calls fast memory, 75TB per rack once you count 54TB of LPDDR5X along with 20.7TB of HBM4. AMD’s 31TB are all HBM, better suited to models that need to be kept entirely in high-bandwidth memory, and the Helios offers higher capacity and bandwidth per accelerator at 432GB and 23.3TB/s.
In the raw scalable fabric, the two are level and both offer 3.6 TB/s per accelerator and 260 TB/s per rack.
Despite this, Helios is an exceptionally strong product on paper, and may be the first time AMD has produced a credible rack-scale answer to Nvidia since the AI race began. Seventy-two MI455X accelerators, 18 EPYC Venice CPUs, 31TB of HBM4, UALink over in-rack Ethernet, and Ultra Ethernet output, in an OCP Open Rack Wide chassis with commercial Broadcom switch silicon, is a serious response to a company that had a two-year lead in form factor.
More importantly, its fabric specifications are publicly available, allowing hyperscalers to create custom variants that meet their requirements. Nvidia’s platform does not offer equivalent latitude.
AMD frames openness as the platform’s core advantage, and Vamsi Boppana, senior vice president of AI at AMD, says Helios “brings together leadership computing, high-performance networking, and open software into a unified rack-scale platform.”
However, the biggest issue for AMD might be memory supply. Nvidia has had Vera Rubin in full production since Q1 with partner availability in this half, while AMD’s first Helios deployments won’t happen until Q4. AMD may be further impacted by HBM4 supply, much of which is reportedly already committed to hyperscalers, which could keep its deployment volumes well below Nvidia’s this year.
Follow TechRadar on Google News and add us as a preferred source to receive news, reviews and opinions from our experts in your feeds.
