GPU model | GA102 | TU104 |
---|---|---|
GPU variant | GA102-890-A1 | TU104-895-A1 |
Architecture | Ampere | Turing |
Foundry | Samsung | TSMC |
Manufacturing process | 8 nm | 12 nm |
Number of transistors | 28,300 million | 13,600 million |
Die size | 628 mm² | 545 mm² |
Release date | Apr 12th, 2021 | Sep 13th, 2018 |
---|---|---|
gpu.details.availability | 2021 | — |
Generation | Tesla | Tesla |
Production | Active | Active |
Interface | PCIe 4.0 x16 | PCIe 3.0 x16 |
Base clock | 885 MHz | 585 MHz |
---|---|---|
Boost clock | 1695 MHz | 1590 MHz |
Memory clock | 1563 MHz 12.5 Gbps effective | 1250 MHz 10 Gbps effective |
Memory size | 24 GB | 16 GB |
---|---|---|
Memory type | GDDR6 | GDDR6 |
Memory bus width | 384 bit | 256 bit |
Bandwidth | 600.2 GB/s | 320.0 GB/s |
Shading units | 9216 | 2560 |
---|---|---|
TMUs | 288 | 160 |
ROPs | 96 | 64 |
SM count | 72 | 40 |
Tensor cores | 288 | 320 |
RT cores | 72 | 40 |
Cache L1 | 128 KB (per SM) | 64 KB (per SM) |
Cache L2 | 6 MB | 4 MB |
Pixel rate | 162.7 GPixel/s | 101.8 GPixel/s |
---|---|---|
Texture rate | 488.2 GTexel/s | 254.4 GTexel/s |
FP16 (half) performance | 31.24 TFLOPS (1:1) | 65.13 TFLOPS (8:1) |
FP32 (float) performance | 31.24 TFLOPS | 8.141 TFLOPS |
FP64 (double) performance | 976.3 GFLOPS (1:32) | 254.4 GFLOPS (1:32) |
Slot width | Single-slot | Single-slot |
---|---|---|
Length | 267 mm 10.5 inches | 168 mm 6.6 inches |
Width | 112 mm 4.4 inches | — |
TDP | 150 W | 70 W |
Suggested PSU | 450 W | 250 W |
Outputs | No outputs | No outputs |
Power connectors | 8-pin EPS | None |
Board number | PG133 SKU 215 | PG183 SKU 200 |
DirectX | 12 Ultimate (12_2) | 12 Ultimate (12_2) |
---|---|---|
OpenGL | 4.6 | 4.6 |
OpenCL | 3.0 | 3.0 |
Vulkan | 1.2 | 1.2 |
CUDA | 8.6 | 7.5 |
Shader model | 6.6 | 6.6 |