NVIDIA GRID A100B vs NVIDIA Tesla K40m
Processeur graphique
| Nom du GPU | GA100 | GK110B |
|---|---|---|
| Architecture | Ampere | Kepler |
| Fonderie | TSMC | TSMC |
| Taille du processus | 7 nm | 28 nm |
| Transistors | 54,200 million | 7,080 million |
| Taille des matrices | 826 mm² | 561 mm² |
Carte graphique
| Date de publication | May 14th, 2020 | Nov 22nd, 2013 |
|---|---|---|
| Génération | GRID | Tesla |
| Production | Active | End-of-life |
| Bus interface | PCIe 4.0 x16 | PCIe 3.0 x16 |
| Prix de lancement | — | 7,699 USD |
Vitesse clock
| Base clock | 900 MHz | 745 MHz |
|---|---|---|
| Boost clock | 1005 MHz | 876 MHz |
| Mémoire clock | 1215 MHz 2.4 Gbps effective | 1502 MHz 6 Gbps effective |
Mémoire
| Taille de la mémoire | 48 GB | 12 GB |
|---|---|---|
| Type de mémoire | HBM2E | GDDR5 |
| Bus mémoire | 6144 bit | 384 bit |
| Bande passante | 1,866 GB/s | 288.4 GB/s |
Render Config
| Shading units | 6912 | 2880 |
|---|---|---|
| TMUs | 432 | 240 |
| ROPs | 192 | 48 |
| Comptage SM | 108 | — |
| Tensor cores | 432 | — |
| Cache L1 | 192 KB (per SM) | 16 KB (per SMX) |
| Cache L2 | 48 MB | 1536 KB |
| Comptage SMX | — | 15 |
Performance théorique
| Taux de pixel | 193.0 GPixel/s | 52.56 GPixel/s |
|---|---|---|
| Taux de texture | 434.2 GTexel/s | 210.2 GTexel/s |
| FP16 (half) performance | 55.57 TFLOPS (4:1) | — |
| FP32 (float) performance | 13.89 TFLOPS | 5.046 TFLOPS |
| FP64 (double) performance | 6.947 TFLOPS (1:2) | 1.682 TFLOPS (1:3) |
Board Design
| Taille Slot | IGP | Dual-slot |
|---|---|---|
| TDP | 400 W | 245 W |
| Suggestions pour le PSU | 800 W | 550 W |
| Sorties | No outputs | No outputs |
| Connecteurs de puissance | None | — |
| Longueur | — | 267 mm 10.5 inches |
Caractéristiques graphiques
| DirectX | — | 12 (11_1) |
|---|---|---|
| OpenGL | — | 4.6 |
| OpenCL | 3.0 | 3.0 |
| Vulkan | — | 1.1 |
| CUDA | 8.0 | 3.5 |
| Shader model | — | 5.1 |
