PML 3.0.0 large-triplet timings¶
These measurements compare v3.0.0 with v2.9.0. For the current release, see the 3.0.1 performance overview and 3.0.1 timings and memory measurements.
3 fresh processes per configuration, each averaging 3 blocks of 10 forward/backward pairs after warmup. Times are milliseconds per pair; peak memory is total PyTorch allocation in MiB. Encoder, optimizer, and data loading are excluded. This campaign measures timing and memory; full-gradient agreement and capacity require separate evidence.
| GPU | GPU memory (GiB) | Host CPU | PyTorch | CUDA |
|---|---|---|---|---|
| RTX 4090 | 23.52 | AMD EPYC 7352 24-Core Processor | 2.14.0+cu130 | 13.0 |
| RTX 5090 | 31.36 | AMD Ryzen 9 9950X 16-Core Processor | 2.14.0+cu130 | 13.0 |
| L40S | 44.39 | AMD EPYC 9354 32-Core Processor | 2.14.0+cu130 | 13.0 |
| A100-SXM4-80GB | 79.25 | AMD EPYC 7742 64-Core Processor | 2.14.0+cu130 | 13.0 |
| H100 80GB HBM3 | 79.18 | Intel(R) Xeon(R) Platinum 8470 | 2.14.0+cu130 | 13.0 |
CSV measurements retain full precision and statuses.
standard: 128 dimensions, 4 samples per class for pair/triplet losses, 1,024 classes for proxy losses.
All-triplet · standard · float32 · batch 8,192¶
| GPU | Version | Status | Total | Forward | Backward | Process range | Peak MiB |
|---|---|---|---|---|---|---|---|
| RTX 4090 | 2.9.0 | ok | 3,041 | 2,448 | 593 | 3,027–3,058 | 16,528 |
| RTX 4090 | 3.0.0 | ok | 289 | 91.6 | 198 | 289–289 | 15,950 |
| RTX 5090 | 2.9.0 | ok | 838 | 619 | 219 | 837–840 | 16,566 |
| RTX 5090 | 3.0.0 | ok | 194 | 54.6 | 139 | 194–194 | 15,989 |
| L40S | 2.9.0 | ok | 1,530 | 1,161 | 368 | 1,477–1,557 | 16,528 |
| L40S | 3.0.0 | ok | 375 | 126 | 249 | 375–375 | 15,950 |
| A100-SXM4-80GB | 2.9.0 | ok | 2,540 | 2,137 | 402 | 2,532–2,553 | 16,528 |
| A100-SXM4-80GB | 3.0.0 | ok | 293 | 54.9 | 238 | 293–293 | 15,950 |
| H100 80GB HBM3 | 2.9.0 | ok | 1,316 | 1,110 | 206 | 1,273–1,392 | 16,566 |
| H100 80GB HBM3 | 3.0.0 | ok | 158 | 32.4 | 126 | 158–158 | 15,989 |
Raw evidence¶
RTX 4090: raw records · configuration and hardware
RTX 5090: raw records · configuration and hardware
L40S: raw records · configuration and hardware
A100-SXM4-80GB: raw records · configuration and hardware
H100 80GB HBM3: raw records · configuration and hardware