GPU model | Mars | TU102 |
---|---|---|
GPU variant | Mars LE | TU102-875-A1 |
Architecture | GCN 1.0 | Turing |
Foundry | TSMC | TSMC |
Manufacturing process | 28 nm | 12 nm |
Number of transistors | 950 million | 18,600 million |
Die size | 77 mm² | 754 mm² |
Release date | Apr 1st, 2013 | — |
---|---|---|
Generation | Solar System (HD 8700M) | — |
Production | End-of-life | — |
Interface | PCIe 3.0 x8 | — |
Predecessor | London | — |
Successor | Crystal System | — |
Base clock | 650 MHz | 1065 MHz |
---|---|---|
Boost clock | 700 MHz | 1395 MHz |
Memory clock | 900 MHz 1800 Mbps effective | 1750 MHz 14 Gbps effective |
Memory size | 2 GB | 4 GB |
---|---|---|
Memory type | DDR3 | GDDR6 |
Memory bus width | 128 bit | 384 bit |
Bandwidth | 28.80 GB/s | 672.0 GB/s |
Shading units | 384 | 4608 |
---|---|---|
TMUs | 24 | 288 |
ROPs | 8 | 96 |
Compute units | 6 | — |
Cache L1 | 16 KB (per CU) | 64 KB (per SM) |
Cache L2 | 256 KB | 6 MB |
SM count | — | 72 |
Tensor cores | — | 576 |
RT cores | — | 72 |
Pixel rate | 5.600 GPixel/s | 133.9 GPixel/s |
---|---|---|
Texture rate | 16.80 GTexel/s | 401.8 GTexel/s |
FP32 (float) performance | 537.6 GFLOPS | 12.86 TFLOPS |
FP64 (double) performance | 33.60 GFLOPS (1:16) | 401.8 GFLOPS (1:32) |
FP16 (half) performance | — | 25.71 TFLOPS (2:1) |
TDP | unknown | 260 W |
---|---|---|
Outputs | No outputs | No outputs |
Slot width | — | Dual-slot |
Length | — | 267 mm 10.5 inches |
Suggested PSU | — | 600 W |
Power connectors | — | 1x 6-pin + 1x 8-pin |
Board number | — | PG150 SKU 215 |
DirectX | 12 (11_1) | 12 Ultimate (12_2) |
---|---|---|
OpenGL | 4.6 | 4.6 |
OpenCL | 1.2 | 3.0 |
Vulkan | 1.2 | 1.2 |
Shader model | 5.1 | 6.6 |
CUDA | — | 7.5 |
Release date | — | Unknown |
---|---|---|
Generation | — | GRID |
Production | — | Active |
Interface | — | PCIe 3.0 x16 |