AIBRIEF
AMD's 72 GPU Helios rack platform carries 31TB of HBM4 high bandwidth memory versus 20.7TB on Nvidia's competing 72 GPU Vera Rubin NVL72 rack, but trails on per rack FP4 (4 bit floating point) compute.
AMD launched Helios at its Advancing AI event on July 23, a 72-GPU rack built on MI455X Instinct accelerators and 6th-gen EPYC CPUs, pitched as a direct rival to Nvidia's Vera Rubin NVL72. The headline numbers split by the way you measure them.
At the GPU, AMD claims Helios beats Rubin's per-chip FP4 throughput by 15%. At the rack, AMD posts 2.9 exaflops of FP4 against Nvidia's 3.6 exaflops. On memory, AMD's 31TB of HBM4 outruns NVL72's 20.7TB; Nvidia's 75TB "fast memory" figure only leads once 54TB of LPDDR5X system memory is folded in alongside the HBM.
Two qualifiers matter. AMD's per-GPU figure uses MXFP4, Nvidia's uses NVFP4; the two are not identical arithmetic, and AMD did not say whether Nvidia's rack number counts individual dies or two-die packages. AMD's own engineers conceded at the event that measured FP4 lands near half of peak on real workloads, blaming memory movement, scheduling, and kernel efficiency.
AMD's 30%-more-tokens-per-dollar claim and its open-rack pitch are the only public numbers to weigh against Nvidia's rack totals until third-party benchmarks arrive.