The fastest CPU against the fastest GPU on each workload, one GPU against a whole two-socket CPU server, all 6 platforms measured the same way, each with code tuned for it.
Fastest of each
| Workload | Fastest CPU | Fastest GPU | GPU against CPU |
|---|---|---|---|
| Swaption portfolio, LMM | AMD EPYC Turin, 44.0M/s | AMD Instinct MI355X, 192M/s | 4.4× |
| Barrier options, Monte Carlo | AMD EPYC Turin, 1.04M/s | AMD Instinct MI355X, 2.71M/s | 2.6× |
| American options, binomial lattice | AMD EPYC Turin, 4.94M/s | AMD Instinct MI355X, 7.26M/s | 1.5× |
| European options, Black-Scholes formula | Intel Xeon 6 (Granite Rapids), 23.2B/s | AMD Instinct MI355X, 68.6B/s | 3.0× |
On the LMM swaption portfolio the fastest GPU (AMD Instinct MI355X) is 4.4× faster than the fastest CPU (AMD EPYC Turin); on barrier options (Monte Carlo) the fastest GPU (AMD Instinct MI355X) is 2.6× faster than the fastest CPU (AMD EPYC Turin); on American options (binomial lattice) the fastest GPU (AMD Instinct MI355X) is 1.5× faster than the fastest CPU (AMD EPYC Turin); on European options (Black-Scholes formula) the fastest GPU (AMD Instinct MI355X) is 3.0× faster than the fastest CPU (Intel Xeon 6 (Granite Rapids)).
The hardware
- Cores
- 2 × 96
- Threads
- 192
- Memory
- 375 GiB
- Peak power (measured)
- 320 W per socket
- 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
- 2 × 96
- Threads
- 192
- Memory
- 755 GiB
- Power limit
- 600 W per socket
Energy per result
Energy is compared GPU with GPU and CPU with CPU, not between the two: GPUs read the board sensor alone, host excluded; CPUs the processor package alone, memory excluded. Each figure is on the results and in its measurement record.
No single ratio sizes a grid: the gap depends on the workload, so measure the ones that make up the bill.
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.