GPU model | GA102 | GM20B |
---|---|---|
Architecture | Ampere | Maxwell 2.0 |
Foundry | Samsung | TSMC |
Manufacturing process | 8 nm | 20 nm |
Number of transistors | 28,300 million | 2,000 million |
Die size | 628 mm² | 118 mm² |
GPU variant | — | TM670D-A1 |
Release date | Oct 5th, 2020 | — |
---|---|---|
gpu.details.availability | 2021 | — |
Generation | Tesla | — |
Production | Active | — |
Interface | PCIe 4.0 x16 | — |
Base clock | 1305 MHz | — |
---|---|---|
Boost clock | 1740 MHz | — |
Memory clock | 1812 MHz 14.5 Gbps effective | System Shared |
GPU clock | — | 1000 MHz |
Memory size | 48 GB | System Shared |
---|---|---|
Memory type | GDDR6 | System Shared |
Memory bus width | 384 bit | System Shared |
Bandwidth | 695.8 GB/s | System Dependent |
Shading units | 10752 | 256 |
---|---|---|
TMUs | 336 | 16 |
ROPs | 112 | 16 |
SM count | 84 | — |
Tensor cores | 336 | — |
RT cores | 84 | — |
Cache L1 | 128 KB (per SM) | 48 KB (per SMM) |
Cache L2 | 6 MB | 256 KB |
SMM count | — | 2 |
Pixel rate | 194.9 GPixel/s | 16.00 GPixel/s |
---|---|---|
Texture rate | 584.6 GTexel/s | 16.00 GTexel/s |
FP16 (half) performance | 37.42 TFLOPS (1:1) | 1,024 GFLOPS (2:1) |
FP32 (float) performance | 37.42 TFLOPS | 512.0 GFLOPS |
FP64 (double) performance | 1,169 GFLOPS (1:32) | 16.00 GFLOPS (1:32) |
Slot width | Dual-slot | IGP |
---|---|---|
Length | 267 mm 10.5 inches | — |
Width | 112 mm 4.4 inches | — |
TDP | 300 W | 15 W |
Suggested PSU | 700 W | — |
Outputs | 3x DisplayPort | No outputs |
Power connectors | 8-pin EPS | — |
DirectX | 12 Ultimate (12_2) | 12 (12_1) |
---|---|---|
OpenGL | 4.6 | 4.6 |
OpenCL | 3.0 | 1.2 |
Vulkan | 1.2 | 1.1 |
CUDA | 8.6 | 5.3 |
Shader model | 6.6 | 6.4 |
Release date | — | Jan 4th, 2015 |
---|---|---|
Generation | — | Tegra |
Production | — | End-of-life |
Interface | — | IGP |