Performance
From Gw-qcd-wiki
Revision as of 14:17, 14 July 2010 by 128.164.237.205 (talk)
Carver (Fermi with ECC)
- Tester: Ben Gamari
- Test date: 14 Jul 2010
- Commit: e3e4ffafd158abd004c483694a27f4f6bc7d2185
- Hardware: Tesla C2050 (ECC on)
- CUDA version 3.0
| Kernel | Configuration | Bandwidth | FLOPs |
|---|---|---|---|
| Dslash_cuda | Dslash (24^4) | 73 GB/s | 32 GFLOP/s |
| hopping (24^4) | 74 GB/s | 34 GFLOP/s | |
| Dslash_multi_gpu (double) | 1 node, 24^4 Dslash | 79 GB/s | 35 GFLOP/s |
| 2 nodes, 24^4 Dslash | 145 GB/s | 64 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 256 GB/s | 114 GFLOP/s | |
| Dslash_multi_gpu (double) | 1 node, 24^4 Dslash | 79 GB/s | 76 GFLOP/s |
| 2 nodes, 24^4 Dslash | 156 GB/s | 140 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 283 GB/s | 252 GFLOP/s | |
| Vector utilities | Addition | 82 GB/s | 3.4 GFLOP/s |
| Dot product | 88 GB/s | N/A | |
| Copy | 84 GB/s | N/A |
Carver (Fermi without ECC)
- Tester: Ben Gamari
- Test date: 14 Jul 2010
- Commit: e3e4ffafd158abd004c483694a27f4f6bc7d2185
- Hardware: Tesla C2050 (ECC off)
- CUDA version 3.0
| Kernel | Configuration | Bandwidth | FLOPs |
|---|---|---|---|
| Dslash_cuda | Dslash (24^4) | 85 GB/s | 38 GFLOP/s |
| hopping (24^4) | 86 GB/s | 39 GFLOP/s | |
| Dslash_multi_gpu (double) | 1 node, 24^4 Dslash | 96 GB/s | 43 GFLOP/s |
| 2 nodes, 24^4 Dslash | 179 GB/s | 82 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 304 GB/s | 139 GFLOP/s | |
| Dslash_multi_gpu (single) | 1 node, 24^4 Dslash | 114 GB/s | 101 GFLOP/s |
| 2 nodes, 24^4 Dslash | 210 GB/s | 187 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 279 GB/s | 256 GFLOP/s | |
| Vector utilities | Addition | 110 GB/s | 4.6 GFLOP/s |
| Dot product | 119 GB/s | N/A | |
| Copy | 114 GB/s | N/A |
Carver (Tesla)
- Tester: Ben Gamari
- Test date: 14 Jul 2010
- Commit: e3e4ffafd158abd004c483694a27f4f6bc7d2185
- Hardware: Tesla C1060
- CUDA version 3.0
| Kernel | Configuration | Bandwidth | FLOPs |
|---|---|---|---|
| Dslash_cuda | Dslash (24^4) | 69 GB/s | 31 GFLOP/s |
| hopping (24^4) | 68 GB/s | 31 GFLOP/s | |
| Dslash_multi_gpu (double) | 1 node, 24^4 Dslash | 71 GB/s | 32 GFLOP/s |
| 2 nodes, 24^4 Dslash | 134 GB/s | 60 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 240 GB/s | 107 GFLOP/s | |
| Dslash_multi_gpu (single) | 1 node, 24^4 Dslash | 65 GB/s | 58 GFLOP/s |
| 2 nodes, 24^4 Dslash | 122 GB/s | 112 GFLOP/s | |
| 4 nodes, 24^4 Dslash | 224 GB/s | 199 GFLOP/s | |
| Vector utilities | Addition | 83 GB/s | 3.5 GFLOP/s |
| Dot product | 81 GB/s | N/A | |
| Copy | 83 GB/s | N/A |
Samurai (G80)
- Tester: Ben Gamari
- Test date: 14 Jul 2010
- Commit: e4b3a766ec873f57cc7ef31b4167bc2f032d418e
- Hardware: GeForce GTX 280
- CUDA version 3.1
| Kernel | Configuration | Bandwidth | FLOPs |
|---|---|---|---|
| Dslash_cuda | Dslash (24^4) | 76 GB/s | 34 GFLOP/s |
| hopping (24^4) | 75 GB/s | 34 GFLOP/s | |
| Dslash_multi_gpu (double) | 1 node, 24^4 Dslash | 77 GB/s | 34 GFLOP/s |
| Dslash_multi_gpu (single) | 1 node, 24^4 Dslash | 91 GB/s | 81 GFLOP/s |
| Vector utilities | Addition | 113 GB/s | 4.7 GFLOP/s |
| Dot product | 77 GB/s | N/A | |
| Copy | 121 GB/s | N/A |
