AMD Radeon Instinct MI250X vs NVIDIA A100 PCIe
Graphics Processor
| GPU model | Aldebaran | GA100 |
|---|---|---|
| GPU variant | Aldebaran XT | — |
| Architecture | CDNA 2.0 | Ampere |
| Foundry | TSMC | TSMC |
| Manufacturing process | 6 nm | 7 nm |
| Number of transistors | 58,200 million | 54,200 million |
| Die size | unknown | 826 mm² |
Graphics Card
| Release date | Nov 8th, 2021 | Jun 22nd, 2020 |
|---|---|---|
| Generation | Radeon Instinct | Tesla |
| Production | unknown | Active |
| Interface | PCIe 4.0 x16 | PCIe 4.0 x16 |
Clocks
| Base clock | 1000 MHz | 765 MHz |
|---|---|---|
| Boost clock | 1700 MHz | 1410 MHz |
| Memory clock | 1600 MHz 3.2 Gbps effective | 1215 MHz 2.4 Gbps effective |
| GPU clock | — | 1410 MHz |
Memory Configuration
| Memory size | 128 GB | 40 GB |
|---|---|---|
| Memory type | HBM2E | HBM2E |
| Memory bus width | 8192 bit | 5120 bit |
| Bandwidth | 3,277 GB/s | 1,555 GB/s |
Render Configuration
| Shading units | 14080 | 6912 |
|---|---|---|
| TMUs | 880 | 432 |
| ROPs | 0 | 160 |
| Compute units | 220 | — |
| Cache L1 | 16 KB (per CU) | 192 KB (per SM) |
| Cache L2 | 16 MB | 40 MB |
| gpu.details.mcm | 2 | — |
| SM count | — | 108 |
| Tensor cores | — | 432 |
Performance
| Pixel rate | 0 MPixel/s | 225.6 GPixel/s |
|---|---|---|
| Texture rate | 1,496 GTexel/s | 609.1 GTexel/s |
| FP16 (half) performance | 383.0 TFLOPS (8:1) | 77.97 TFLOPS (4:1) |
| FP32 (float) performance | 47.87 TFLOPS | 19.49 TFLOPS |
| FP64 (double) performance | 47.87 TFLOPS (1:1) | 9.746 TFLOPS (1:2) |
Dimensions & Outputs
| Slot width | Dual-slot | Dual-slot |
|---|---|---|
| Length | 267 mm 10.5 inches | 267 mm 10.5 inches |
| Width | 111 mm 4.4 inches | — |
| TDP | 500 W | 250 W |
| Suggested PSU | 900 W | 600 W |
| Outputs | No outputs | No outputs |
| Power connectors | 2x 8-pin | 8-pin EPS |
| Board number | — | P1001 SKU 200 |
API Support & Features
| DirectX | — | — |
|---|---|---|
| OpenGL | — | — |
| OpenCL | 3.0 | 3.0 |
| Vulkan | — | — |
| Shader model | — | — |
| CUDA | — | 8.0 |
