AMD Helios moves the constraint from chips to facility economics
The Helios rack carries 31 TB of HBM4 across 72 MI455X GPUs, and that memory stack sets the scale of the decision.
| AMD Helios | |
|---|---|
| GPUs per rack | 72 AMD Instinct MI455X GPUs |
| HBM4 memory | 31 TB |
| Peak OCP MXFP4 | up to 2.9 exaFLOPS |
| Peak OCP MXFP8 | up to 1.4 exaFLOPS |
| Memory bandwidth | 1.67 PB/s |
| GPU HBM4 | 432 GB |
| GPU memory bandwidth | 23.3 TB/s |
| Scale-up bandwidth per GPU | up to 3.6 TB/s |
The rack's peak OCP MXFP4 figure is 2.9 exaFLOPS, with 1.67 PB/s of memory bandwidth. Each MI455X carries 432 GB of HBM4 at 23.3 TB/s.
The page describes the platform as co-designed with EPYC CPUs and Pensando networking. It also names CoWoS-L packaging and 3D hybrid bonded compute dies.
The unmeasured parts are tokens per second, watt per token, and the price of a populated rack. Yield, TDP, and ship date remain unstated.
The buyer must decide how much of the peak figure survives real model shapes, interconnect contention, and cooling. The rack turns a GPU spec into a facility decision.
The spec proves a 72-GPU configuration, not sustained throughput, power draw, or price.
After amd.com. We did not report this. The pictures, if any, are theirs.
