The fastest GPU of each vendor we measured, by the geometric mean of its throughput at the ranked size over the four workloads: NVIDIA B200 for NVIDIA, AMD Instinct MI355X for AMD. Measured FP64 FMA ceiling: 36,800 against 71,000 GFLOP/s. Memory bandwidth: 8,000 GB/s each. One GPU each, code tuned for each.
NVIDIA B200: one whole GPU, on the 96 host threads of its memory node. Default power limit, ECC on, clocks not locked. AMD Instinct MI355X: one whole GPU, unpartitioned, with 192 host threads. Power cap 1,400 W, the maker's rated board power, ECC on, clocks not locked.
The hardware
- Cores
- 256 (CU)
- Clock
- 2,400 MHz
- Measured FP64 FMA ceiling, GFLOP/s
- 71,000
- Memory bandwidth
- 8,000 GB/s
- Memory
- 288 GB (maker figure)
- Power limit
- 1,400 W
- Cores
- 148 (SM)
- Clock
- 1,965 MHz
- Measured FP64 FMA ceiling, GFLOP/s
- 36,800
- Memory bandwidth
- 8,000 GB/s
- Memory
- 179.1 GiB
- Power limit
- 1,000 W
Results
| Workload | AMD Instinct MI355X | NVIDIA B200 |
|---|---|---|
| Swaption portfolio, LMM paths per second at 1,024K paths | 192M/s | 78.0M/s |
| Barrier options, Monte Carlo options per second at 16,000 options | 2.71M/s | 1.67M/s |
| American options, binomial lattice options per second at 512K options | 7.26M/s | 1.85M/s |
| European options, Black-Scholes formula options per second at 128M options | 68.6B/s | 52.7B/s |
Tuned code: tuned for each chip under the same rules. Every size, and every platform, is in the results; each platform name links to its measurement record.
AMD Instinct MI355X is 2.5× faster than NVIDIA B200 on the LMM swaption portfolio; AMD Instinct MI355X is 1.6× faster than NVIDIA B200 on barrier options (Monte Carlo); AMD Instinct MI355X is 3.9× faster than NVIDIA B200 on American options (binomial lattice); AMD Instinct MI355X is 1.3× faster than NVIDIA B200 on European options (Black-Scholes formula).
Energy per result
The LMM swaption portfolio: AMD Instinct MI355X uses 15% less energy per result than NVIDIA B200; Barrier options (Monte Carlo): NVIDIA B200 uses 13% less energy per result than AMD Instinct MI355X; American options (binomial lattice): AMD Instinct MI355X uses 41% less energy per result than NVIDIA B200; European options (Black-Scholes formula): NVIDIA B200 uses 20% less energy per result than AMD Instinct MI355X.
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.