Compare the three rack teardowns
Every figure below comes from the vendor's own spec sheet or product page (see each teardown's own sourcing) — nothing here is estimated except the two rows marked "computed", which are Google's own per-chip numbers multiplied by its own 64-chip rack count (Google doesn't publish an aggregate at that unit, only per-chip and per-superpod; the arithmetic checks against Google's superpod figure exactly).
| GB300 NVL72 | AMD Helios | Google TPU7x | |
|---|---|---|---|
| Accelerator | NVIDIA Blackwell Ultra (B300) GPU | AMD Instinct MI455X GPU | Google TPU7x (Ironwood) ASIC |
| Host CPU | Grace CPU (36 per rack, Arm Neoverse V2) | EPYC "Venice" CPU (18 per rack, Zen 6, up to 256 cores) | x86 host per 4-chip set confirmed to exist (224 vCPUs); vendor not disclosed |
| Chips per rack | 72 | 72 | 64 |
| Memory type | HBM3e | HBM4 | HBM3E |
| Memory per chip | 288 GB | 432 GB | 192 GB |
| Memory bandwidth per chip | 8 TB/s | 23.3 TB/s | 7.37 TB/s |
| Memory total per rack | 20.7 TB | 31 TB | 12.3 TB (computed) |
| FP8 compute per rack | 720 PFLOPS | 1,400 PFLOPS | 295 PFLOPS (computed) |
| FP4 compute per rack | 1,440 PFLOPS | 2,900 PFLOPS | — |
| Scale-up bandwidth | 130 TB/s | 260 TB/s | 1.2 TB/s † |
| Scale-up fabric | NVLink 5 + NVSwitch (electrical switch chips) | UALink-over-Ethernet on Broadcom Tomahawk 6 ASICs (electrical switch chips) | 3D-torus Inter-Chip Interconnect (ICI), direct chip-to-chip, no switch chip |
| Scale-out bandwidth | — | 43 TB/s | — |
| Rack power | 135 kW design, up to 155 kW peak | 225–245 kW (DC busbar capacity) | not disclosed per-rack; ~10 MW for a full 9,216-chip superpod |
| Rack mechanical standard | NVIDIA proprietary liquid-cooled cabinet | OCP Open Rack Wide (open standard, double-wide) | Google proprietary |
Google's own docs call a 64-chip rack a "cube" — there is no 72-chip TPU rack; 64 is the largest unit Google itself calls a rack.
† per-chip bidirectional ICI, not a rack aggregate — Google's fabric is a direct chip-to-chip mesh, not a switched network, so there is no single "aggregate bandwidth" figure comparable to NVLink/UALink
Google's TPU7x also scales far past a single rack in a way the other two don't have a published equivalent for: a 256-chip "pod" (four cubes) and a 9,216-chip "superpod" (144 cubes) connected by the same reconfigurable optical circuit switch described on its teardown page, delivering 42.5 exaFLOPS of FP8 at full superpod scale — an order of magnitude past what either GB300 or Helios racks are wired together into today.