Inside the AI data centre · Compute hall
The Compute Rack
The two 2026 racks of 72 GPUs side by side: NVIDIA Vera Rubin NVL72 and AMD Helios. Like-for-like rows first; every figure from the vendor.
| Rack | NVIDIAVera Rubin NVL72NVIDIA’s 2026 rack | AMDHeliosAMD’s 2026 rack |
|---|---|---|
| Timing and form | ||
| Announced → available | January 2026 → production May 2026, 2H 20261,2 | Previewed June 2025, launched July 2026 → volume 2H 20263,4,5 |
| Rack | Third-generation MGX NVL72, single-wide (19-inch)6,7 | OCP Open Rack Wide, double-wide; reference design built by OEMs5,8 |
| Trays | 18 compute (4 GPU + 2 CPU) · 9 NVLink switch7,9Per-tray count derived from the superchip make-up. | 18 compute (4 GPU + 1 CPU) · 6 switch trays · 4 rear cable cartridges4,10One CPU per tray is derived (18 CPUs, 18 trays). |
| Chips and memory | ||
| GPUs | 72 × Rubin6 | 72 × Instinct MI455X5 |
| CPUsDifferent ratios: Vera Rubin pairs one CPU with two GPUs, Helios one with four. | 36 × Vera6 | 18 × EPYC “Venice”8 |
| HBM per GPU | 288 GB HBM46 | 432 GB HBM411 |
| HBM per rack | 20.7 TB6 | 31 TB5 |
| HBM bandwidth per rack | 1.4 PB/s6 | 1.67–1.7 PB/s11,12 |
| Fabrics | ||
| Scale-upInside the rack: the fabric that makes 72 GPUs one domain. | NVLink 6 · 216 TB/s6,13Launch pages said 260 TB/s. | UALink over Ethernet · 260 TB/s12,5 |
| Scale-outBetween racks. Vendors count directions differently; Helios’s 43 TB/s counts both. | 32.4 TB/s bi-directional · ConnectX-96 | 43 TB/s · Pensando “Vulcano”, up to 2.4 Tb/s per GPU5 |
| Peak AI compute (rack) | ||
| FP4, denseLike with like: dense figures only. The formats differ (NVFP4 vs MXFP4). | 2.52 EFLOPS6NVIDIA’s NVFP4 training figure. | 2.9 EFLOPS12 |
| FP4, headlineThe figure each vendor leads with. Not comparable: NVIDIA’s is an inference figure, AMD’s is dense. | 3.6 EFLOPS6NVFP4 inference. | 2.9 EFLOPS12Dense; AMD publishes no sparse figure. |
| FP8, dense | 1.26 EFLOPS6 | 1.4 EFLOPS11Peak MXFP8. |
| Facility and ecosystem | ||
| Rack power | Not published7 | Not published (NDA)12 |
| Cooling | 100% liquid, 45 °C inlet13 | Liquid, quick-disconnects to every tray5 |
| Software | DGX OS, NVIDIA AI Enterprise, Mission Control14 | ROCm, open source; Day-0 PyTorch, JAX, vLLM, Triton5 |
| Openness | NVIDIA’s NVLink and rack design6 | OCP rack, UALink protocol, Ethernet scale-out5 |
What decides between them
- Memory capacity and bandwidth. Decide whether a model fits on one rack without splitting, how many users’ context it can hold, and how fast tokens come out. Helios leads on capacity per GPU.
- Scale-up fabric. Decides how well 72 GPUs act as one for training and large-model inference. NVLink is mature and proprietary; Helios carries the open UALink protocol over Ethernet silicon, new in this generation.
- Software maturity. CUDA has the larger installed base of tuned libraries and tools; ROCm is open source and supports the main frameworks from day one, but porting and tuning effort is real and workload-specific.
- Power, cooling and the facility. Both are liquid-cooled, and neither vendor publishes a rack power figure: plan the facility with the supplier. Helios also needs a double-width rack position.
- Openness against integration. Helios is an OCP reference design built by several OEMs; NVL72 is one vendor’s integrated system. More suppliers to choose from, and more parties to integrate.
What this table does not tell you
- Peak figures are not benchmarks. Usable performance per watt, cost per token and time to train depend on the workload, the software and the cluster around the rack.
- AMD’s own comparison with Vera Rubin NVL72 (up to 15% more compute, 50% more HBM, 50% more scale-out bandwidth) is AMD’s claim, not an independent result.
- Headline FP4 figures mix inference and dense numbers, and NVFP4 is not MXFP4. Compare the dense row.
- Availability is not the same as volume: both racks are ramping in the second half of 2026, and most fleets today still run the previous generation.
Sources
- 1NVIDIA Kicks Off the Next Generation of AI With Rubin · NVIDIA Newsroom · 2026-01-05
- 2NVIDIA Vera Rubin Ramps Into Full Production to Power Agentic AI Factories Worldwide · NVIDIA Newsroom · 2026-05-31
- 3AMD Unveils Vision for an Open AI Ecosystem, Detailing New Silicon, Software and Systems at Advancing AI 2025 · AMD (press release) · 2025-06-12
- 4AAI 2026: AMD Launches AMD Helios Rackscale Solution for Frontier AI · AMD Newsroom · 2026-07-23
- 5AMD Helios · AMD (product page)
- 6NVIDIA Vera Rubin NVL72 · NVIDIA (product page)
- 7NVIDIA Vera Rubin POD: Seven Chips, Five Rack-Scale Systems, One AI Supercomputer · NVIDIA Technical Blog · 2026-03-16
- 8AAI 2026: AMD Delivers Full-Stack Compute for the Agentic AI Era · AMD (press release) · 2026-07-23
- 9Inside the NVIDIA Vera Rubin Platform: Six New Chips, One AI Supercomputer · NVIDIA Technical Blog · 2026-01-05
- 10AMD Helios rack, front and rear (labelled diagram) · AMD, shown in its Helios architecture deep dive at Advancing AI 2026; reproduced by ServeTheHome · 2026-08-03
- 11AMD Instinct MI400 Series · AMD (product page)
- 12AMD Launches Helios: The Highest Performing Rackscale AI Infrastructure Solution · AMD blog · 2026-07-23
- 13NVIDIA Vera Rubin datasheet · NVIDIA (PDF) · 2026-09
- 14NVIDIA DGX Vera Rubin NVL72 · NVIDIA
Vendor figures, read on the vendors’ own pages; derived values are marked. Not independently verified.