NVIDIA A100 PCIe 80 GB vs NVIDIA Tesla T4
Processeur graphique
| Nom du GPU | GA100 | TU104 |
|---|---|---|
| Architecture | Ampere | Turing |
| Fonderie | TSMC | TSMC |
| Taille du processus | 7 nm | 12 nm |
| Transistors | 54,200 million | 13,600 million |
| Taille des matrices | 826 mm² | 545 mm² |
| GPU variant | — | TU104-895-A1 |
Carte graphique
| Date de publication | Jun 28th, 2021 | Sep 13th, 2018 |
|---|---|---|
| Génération | Tesla | Tesla |
| Production | Active | Active |
| Bus interface | PCIe 4.0 x16 | PCIe 3.0 x16 |
Vitesse clock
| Base clock | 1065 MHz | 585 MHz |
|---|---|---|
| Boost clock | 1410 MHz | 1590 MHz |
| Mémoire clock | 1593 MHz 3.2 Gbps effective | 1250 MHz 10 Gbps effective |
Mémoire
| Taille de la mémoire | 80 GB | 16 GB |
|---|---|---|
| Type de mémoire | HBM2e | GDDR6 |
| Bus mémoire | 5120 bit | 256 bit |
| Bande passante | 2,039 GB/s | 320.0 GB/s |
Render Config
| Shading units | 6912 | 2560 |
|---|---|---|
| TMUs | 432 | 160 |
| ROPs | 160 | 64 |
| Comptage SM | 108 | 40 |
| Tensor cores | 432 | 320 |
| Cache L1 | 192 KB (per SM) | 64 KB (per SM) |
| Cache L2 | 80 MB | 4 MB |
| RT cores | — | 40 |
Performance théorique
| Taux de pixel | 225.6 GPixel/s | 101.8 GPixel/s |
|---|---|---|
| Taux de texture | 609.1 GTexel/s | 254.4 GTexel/s |
| FP16 (half) performance | 77.97 TFLOPS (4:1) | 65.13 TFLOPS (8:1) |
| FP32 (float) performance | 19.49 TFLOPS | 8.141 TFLOPS |
| FP64 (double) performance | 9.746 TFLOPS (1:2) | 254.4 GFLOPS (1:32) |
Board Design
| Taille Slot | Dual-slot | Single-slot |
|---|---|---|
| Longueur | 267 mm 10.5 inches | 168 mm 6.6 inches |
| TDP | 250 W | 70 W |
| Suggestions pour le PSU | 600 W | 250 W |
| Sorties | No outputs | No outputs |
| Connecteurs de puissance | 8-pin EPS | None |
| Board number | P1001 SKU 200 | PG183 SKU 200 |
Caractéristiques graphiques
| DirectX | — | 12 Ultimate (12_2) |
|---|---|---|
| OpenGL | — | 4.6 |
| OpenCL | 3.0 | 3.0 |
| Vulkan | — | 1.2 |
| CUDA | 8.0 | 7.5 |
| Shader model | — | 6.6 |
