Energy per result on each workload: every GPU in the 5 November 2026 round, measured the same way, each with code tuned for it. Each GPU reads its own board sensor, with the host CPU and the rest of the server left out. Results within 10% of each other read as about the same, since sensors differ between makers.
Energy per result
| Workload | AMD Instinct MI350X | AMD Instinct MI355X | NVIDIA B200 | NVIDIA H200 |
|---|---|---|---|---|
| Swaption portfolio, LMM Wh per billion paths | 1.83 Wh | 1.73 Wh | 2.02 Wh | 2.00 Wh |
| Barrier options, Monte Carlo Wh per million options | 0.0938 Wh | 0.0974 Wh | 0.0850 Wh | 0.0899 Wh |
| American options, binomial lattice Wh per billion options | 35.3 Wh | 37.3 Wh | 63.1 Wh | 57.4 Wh |
| European options, Black-Scholes formula Wh per trillion options | 3.49 Wh | 3.51 Wh | 2.81 Wh | 2.43 Wh |
Watt-hours per result at the ranked size; the lowest in bold, with any within 10% of it.
On the LMM swaption portfolio AMD Instinct MI355X uses the least energy per result, about the same as AMD Instinct MI350X; AMD Instinct MI355X uses 14% less energy per result than NVIDIA H200; on barrier options (Monte Carlo) NVIDIA B200 uses the least energy per result, about the same as NVIDIA H200; NVIDIA B200 uses 9% less energy per result than AMD Instinct MI350X; on American options (binomial lattice) AMD Instinct MI350X uses the least energy per result, about the same as AMD Instinct MI355X; AMD Instinct MI350X uses 39% less energy per result than NVIDIA H200; on European options (Black-Scholes formula) NVIDIA H200 uses the least energy per result; NVIDIA H200 uses 14% less energy per result than NVIDIA B200.
The hardware
- Power limit
- 1,000 W
- Energy read from
- GPU board sensor, host excluded
- Cores
- 256 (CU)
- Memory
- 288 GB (maker figure)
- Measured FP64 FMA ceiling, GFLOP/s
- 59,700
- Power limit
- 1,400 W
- Energy read from
- GPU board sensor, host excluded
- Cores
- 256 (CU)
- Memory
- 288 GB (maker figure)
- Measured FP64 FMA ceiling, GFLOP/s
- 71,000
- Power limit
- 1,000 W
- Energy read from
- GPU board sensor, host excluded
- Cores
- 148 (SM)
- Memory
- 179.1 GiB
- Measured FP64 FMA ceiling, GFLOP/s
- 36,800
- Power limit
- 700 W
- Energy read from
- GPU board sensor, host excluded
- Cores
- 132 (SM)
- Memory
- 140.4 GiB
- Measured FP64 FMA ceiling, GFLOP/s
- 33,300
How these were measured
Each platform ran code tuned for that chip under the same rules, in double precision: five launches per measurement, the middle three averaged, warm calls in every launch until the 95% confidence interval of their mean was within ±1% or until the call limit. Any test spreading more than 5% is repeated on a second machine of the same type, and the steadier of the two is published. A Black-Scholes call makes 100 passes over the whole batch, each counted as work, so its options per second are options × 100 passes over the time of a call. At every problem size, the option price is validated against the reference implementation's before any time counts. The full method is on How we measure.