A.I.
Hardware · Chips
Instinct MI455X and Helios
The MI455X is AMD's current top Instinct GPU, or graphics processing unit, built on AMD's fifth-generation CDNA design. Helios is AMD's rack-scale system, meaning the whole server rack is designed as one machine, with 72 MI455X GPUs. AMD says Helios volume deployments are expected in the second half of 2026 (AMD, Helios). The wider Instinct line is on the Instinct page.
The chip
| Figure | One MI455X GPU |
|---|---|
| Compute, FP4 | 40 PFLOPS |
| Compute, FP8 | 20.1 PFLOPS |
| Compute, FP16 or BF16, dense | 5 PFLOPS |
| Memory | 432 GB HBM4, 12 stacks |
| Memory bandwidth | 23.3 TB/s |
| Transistors | 320 billion |
| Scale-up link | 3.6 TB/s, both directions |
| Cooling | Direct liquid cooling |
Figures are from AMD's MI455X product page (AMD, MI455X). The Helios page also says "up to 19.6 TB/s" per GPU in one spot, so AMD's own pages disagree on bandwidth. The 23.3 TB/s figure appears on both pages and is used here. Dense means every number is counted; AMD also prints 10.1 PFLOPS FP16 with structured sparsity, a shortcut that skips zeros.
The Helios rack
| Figure | One Helios rack |
|---|---|
| GPUs | 72 MI455X |
| CPUs | AMD EPYC Venice |
| Compute | 2.9 EFLOPS FP4; 1.4 EFLOPS FP8 |
| Memory | 31 TB HBM4 |
| Scale-up bandwidth | 260 TB/s |
| Scale-out bandwidth | 43 TB/s |
Scale-up is the fast network that ties the GPUs inside the rack together; Helios uses UALink over Ethernet (UALoE), an open standard. Scale-out is the network between racks, here AMD Pensando Vulcano network cards at 800 Gb/s. The rack follows Meta's Open Rack Wide design, a double-wide rack standard submitted to the Open Compute Project.
What AMD claims
- AMD compares MI455X with Nvidia Vera Rubin on its product page: 432 GB against 288 GB of memory and 23.3 TB/s against 22.0 TB/s. Those are AMD's own figures for both chips. Nvidia's own Rubin figures are on the Rubin page.
Units on this page: a PFLOPS is a thousand trillion math operations per second; an EFLOPS is a thousand PFLOPS. FP4 and FP8 store each number in 4 or 8 bits, and MX in front (MXFP4, MXFP8) means a shared-scale version of that format. GB is gigabytes, GiB is about 7% more than a GB, and TB is a thousand GB. TB/s and PB/s are terabytes and petabytes per second, how fast data moves. W is watts and kW is a thousand watts. TDP is the most power a chip is built to draw.