AI Accelerator Power Dataset: GPUs, Servers and Racks
Board power of current AI accelerators from NVIDIA, AMD and Intel, and the rated power of the servers and racks built on them, in one reference table.
How many watts does an H100, a B200 or an MI355X draw, and what does that become once it sits in a server or a rack? This dataset collects the board power of current AI accelerators and the rated power of the systems built on them, and the same numbers drive our AI workload energy calculator.
- The highest board power in this dataset is AMD Instinct MI355X at 1,400 W; the lowest is NVIDIA L4 (PCIe) at 72 W.
- NVIDIA DGX H100 is rated at 10.2 kW for 8 NVIDIA H100 (SXM) accelerators: 1.27 kW each, 1.82 times their 700 W board power.
- NVIDIA DGX B200 is rated at 14.3 kW for 8 NVIDIA B200 (SXM) accelerators: 1.79 kW each, 1.79 times their 1,000 W board power.
- Intel Gaudi 3 HLB-325 baseboard is rated at 7.6 kW for 8 Intel Gaudi 3 (OAM) accelerators: 0.95 kW each, 1.06 times their 900 W board power.
- The NVIDIA GB200 NVL72 rack supports up to 132 kW per rack (rack capacity of the reference power and cooling design), about 1.83 kW per accelerator across its 72 GPUs.
- Board power excludes host CPUs, memory, networking and cooling, so a server draws well above the sum of its GPUs.
Accelerator board power
| Accelerator | Board power | Relative to the highest | Calculator |
|---|---|---|---|
| AMD Instinct MI355X | 1,400 W | Estimate a job | |
| AMD Instinct MI325X | 1,000 W | Estimate a job | |
| NVIDIA B200 (SXM) | 1,000 W | Estimate a job | |
| Intel Gaudi 3 (OAM) | 900 W | Estimate a job | |
| AMD Instinct MI300X | 750 W | Estimate a job | |
| NVIDIA H100 (SXM) | 700 W | Estimate a job | |
| NVIDIA H200 (SXM) | 700 W | Estimate a job | |
| NVIDIA A100 (SXM) | 400 W | Estimate a job | |
| NVIDIA H100 NVL (PCIe) | 400 W | Estimate a job | |
| NVIDIA L40S (PCIe) | 350 W | Estimate a job | |
| NVIDIA L4 (PCIe) | 72 W | Estimate a job |
Board power is the accelerator's own rating (TDP or total board power). Several parts are configurable below their maximum. The calculator link opens the AI workload calculator with that accelerator selected.
Servers and racks
| System | Rated power | GPUs | Per GPU | Rating |
|---|---|---|---|---|
| NVIDIA DGX H100 | 10.2 kW | 8 | 1.27 kW 1.82× board | maximum system power |
| NVIDIA DGX B200 | 14.3 kW | 8 | 1.79 kW 1.79× board | maximum system power |
| Intel Gaudi 3 HLB-325 baseboard | 7.6 kW | 8 | 0.95 kW 1.06× board | subsystem TDP (baseboard and 8 OAMs, no host) |
| NVIDIA GB200 NVL72 rack | 132 kW | 72 | 1.83 kW | rack capacity of the reference power and cooling design (up to) |
Per accelerator divides the rated power by the number of accelerators, so it includes CPUs, memory, networking and fans. The rack figure is a facility design capacity, not a measured draw.
Cite as: ScienceShot, “AI accelerator power dataset”, https://scienceshot.com/ai-accelerator-power (checked 2026-10-10). © ScienceShot. You may quote individual figures with a link to this page.
ScienceShot dataset, checked 10 Oct 2026. How we build it: Method and standards.