Platforms · Measurement record
Recordmi350x-2026-11-05Outputs validatedFixed
AMD Instinct MI350X
Measured on 5 November 2026 on one GPU, with the published method. 26 double-precision measurements of the tuned code for this platform: five launches each, every launch recorded.
One whole GPU, unpartitioned, with 192 host threads. Power cap 1,000 W, the maker's rated board power, ECC on, clocks not locked.
- Published method
- Run and checked by Xcelerit
- Suite version: November 2026 round
- Double precision
- 1 measurement repeated on a second machine
- Accelerator
- AMD Instinct MI350X
- Power limit
- 1,000 W
- Driver
- amdgpu 7.1.3
- Host CPU
- AMD EPYC 9005 (Turin), 192 threads
- Host memory
- 2,015 GiB
- Operating system
- Ubuntu 24.04.5 LTS
- Compiler
- ROCm 7.2.4, gfx950
- Libraries
- hipRAND/rocRAND
Results at a glance
The sizes the results rank at. Time is the mean of the middle three of five launches; throughput is the problem size over that time (Black-Scholes: options × 100 passes, the work of one call). Sizes in K and M are binary (1K = 1,024); throughput is per second, in decimal units. Tuned code: tuned for this chip under the same rules. Select a row to see the launches behind it.
| Workload | Size | Time (ms) | Throughput (per s) | Energy | Within-launch CI | Launch spread | |
|---|---|---|---|---|---|---|---|
| Swaption portfolio, LMM | 1,024K paths | 7.12 | 147,223,852 | 1.826 Wh/bn | ±1.00% | 0.67% | Inspect |
| Barrier options, Monte Carlo | 16,000 options | 6.57 | 2,433,982 | 0.09377 Wh/M | ±0.39% | 0.10% | Inspect |
| American options, binomial lattice | 512K options | 79.63 | 6,583,793 | 35.27 Wh/bn | ±0.15% | 0.39% | Inspect |
| European options, Black-Scholes formula | 128M options | 233.98 | 57,362,711,072 | 3.492 Wh/tn | ±1.00% | 0.63% | Inspect |
All 26 measurements
| Workload | Code | Size | Time (ms) | Within-launch CI | Launch spread | Average warm calls per launch | Published from | Other machine (ms) |
|---|---|---|---|---|---|---|---|---|
| Swaption portfolio, LMM | Tuned code | 16K | 0.472 | ±0.10% | 2.77% | 4,242 | – | – |
| Swaption portfolio, LMM | Tuned code | 32K | 0.480 | ±0.05% | 2.61% | 4,171 | – | – |
| Swaption portfolio, LMM | Tuned code | 64K | 0.697 | ±0.07% | 1.17% | 2,869 | – | – |
| Swaption portfolio, LMM | Tuned code | 128K | 0.949 | ±0.43% | 1.06% | 8,329 | machine 1 | machine 2: 0.946 (-0.3%, launch spread 1.6%) |
| Swaption portfolio, LMM | Tuned code | 256K | 1.727 | ±1.00% | 0.48% | 8,808 | – | – |
| Swaption portfolio, LMM | Tuned code | 512K | 3.436 | ±0.72% | 1.16% | 3,704 | – | – |
| Swaption portfolio, LMM | Tuned code | 1,024K | 7.122 | ±1.00% | 0.67% | 2,000 | – | – |
| Swaption portfolio, LMM | Tuned code | 2,048K | 14.285 | ±0.44% | 0.19% | 203 | – | – |
| Barrier options, Monte Carlo | Tuned code | 2,000 | 1.716 | ±0.17% | 0.74% | 1,166 | – | – |
| Barrier options, Monte Carlo | Tuned code | 4,000 | 2.311 | ±0.16% | 0.02% | 866 | – | – |
| Barrier options, Monte Carlo | Tuned code | 8,000 | 3.704 | ±0.25% | 0.13% | 540 | – | – |
| Barrier options, Monte Carlo | Tuned code | 16,000 | 6.574 | ±0.39% | 0.10% | 305 | – | – |
| Barrier options, Monte Carlo | Tuned code | 32,000 | 12.067 | ±0.74% | 0.34% | 166 | – | – |
| American options, binomial lattice | Tuned code | 16K | 9.607 | ±0.26% | 1.00% | 209 | – | – |
| American options, binomial lattice | Tuned code | 32K | 12.389 | ±0.31% | 1.98% | 162 | – | – |
| American options, binomial lattice | Tuned code | 64K | 17.262 | ±0.24% | 1.60% | 116 | – | – |
| American options, binomial lattice | Tuned code | 128K | 26.901 | ±0.25% | 1.08% | 75 | – | – |
| American options, binomial lattice | Tuned code | 256K | 44.993 | ±0.23% | 0.49% | 45 | – | – |
| American options, binomial lattice | Tuned code | 512K | 79.633 | ±0.15% | 0.39% | 26 | – | – |
| American options, binomial lattice | Tuned code | 1,024K | 149.394 | ±0.15% | 0.30% | 14 | – | – |
| American options, binomial lattice | Tuned code | 2,048K | 289.576 | ±0.21% | 0.11% | 10 | – | – |
| European options, Black-Scholes formula | Tuned code, pinned host memory | 16M | 32.696 | ±1.00% | 0.12% | 236 | – | – |
| European options, Black-Scholes formula | Tuned code, pinned host memory | 32M | 61.663 | ±1.00% | 0.12% | 181 | – | – |
| European options, Black-Scholes formula | Tuned code, pinned host memory | 64M | 119.016 | ±1.00% | 0.10% | 135 | – | – |
| European options, Black-Scholes formula | Tuned code, pinned host memory | 128M | 233.981 | ±1.00% | 0.63% | 158 | – | – |
| European options, Black-Scholes formula | Tuned code, pinned host memory | 256M | 463.583 | ±0.92% | 0.99% | 88 | – | – |
Outputs validated before speed
A time only counts if the output is right. At every problem size, the option price the tuned code computes is validated against the reference implementation's on the same machine.
| Workload | Against the reference code | Result |
|---|---|---|
| Swaption portfolio, LMM | Within 0.91% at 128K paths, on this platform's own random-number generator | Pass |
| Barrier options, Monte Carlo | Within 0.79% at 32,000 options, on this platform's own random-number generator | Pass |
| American options, binomial lattice | Identical to double precision | Pass |
| European options, Black-Scholes formula | Identical to double precision | Pass |
Binomial and Black-Scholes are deterministic: the tuned code gives the reference code's prices. Monte Carlo: on GPUs the barrier and LMM codes generate their random numbers differently, so their prices differ only by sampling: within 4 standard errors of the reference implementation's, and fed the same random numbers they agree with the reference implementation's to within 1e-9 (relative).
Every launch behind one measurement
Pick a workload and a size. Each dot is one launch: the mean of its warm calls, run until the 95% confidence interval was within ±1% or until the call limit. The fastest and slowest launch are dropped and the middle three averaged into the result.
Zoom: drag a box on the chart or use the slider under it; reset brings it back.
| Launch | Warm calls | Warm mean (ms) | Range (ms) | Used |
|---|
How time grows with the problem
Time against size on log scales, for every size measured. The dashed line is perfect linear scaling from the smallest size: points under it mean the machine was not yet full at small sizes.
What the record holds
The measurement record behind this page: every launch, the machine and the output validation, and the results as a table.
Full record (JSON) Results (CSV)
Free to cite with attribution: Xcelerit Finance Benchmarks, record mi350x-2026-11-05.
SHA-2569978f5c1203b007a25f816154024cea8709c7b94f97631d2e9f4402ea8b2f46d
Records are fixed: once published, a record is never edited. A correction is published as a new record, linked from the old one. Checksums of every record