Product Overview
The NVIDIA A100, released in 2020, is a landmark data center AI accelerator that introduced MIG (Multi-Instance GPU) and TF32 / FP16 / BF16 Tensor Cores. Although superseded by the H100, the A100 remains the most widely deployed AI training GPU, available in 40GB and 80GB HBM2e variants.
Core Specifications
| Parameter | 40GB Variant | 80GB Variant |
|---|
| Architecture | Ampere GA100 | Ampere GA100 |
| Process Node | TSMC 7nm | TSMC 7nm |
| Transistor Count | 54 billion | 54 billion |
| Memory | 40 GB HBM2e | 80 GB HBM2e |
| Memory Bandwidth | 1,555 GB/s | 1,935 GB/s |
| CUDA Cores | 6,912 | 6,912 |
| Tensor Cores | 432 (3rd Gen) | 432 (3rd Gen) |
| FP32 | 19.5 TFLOPS | 19.5 TFLOPS |
| FP64 | 9.7 TFLOPS | 9.7 TFLOPS |
| TF32 Tensor Core | 156 TFLOPS | 156 TFLOPS |
| FP16/BF16 Tensor Core | 312 TFLOPS | 312 TFLOPS |
| INT8 Tensor Core | 624 TOPS | 624 TOPS |
| TDP | 250 W / 400 W | 300 W / 400 W |
| NVLink | 600 GB/s | 600 GB/s |
| MIG | Up to 7 instances | Up to 7 instances |
Software & Drivers
Key Features
- 3rd Gen Tensor Cores: Support TF32, FP16, BF16, INT8
- MIG (Multi-Instance GPU): Partition a single GPU into up to 7 independent instances
- Structured Sparsity: Hardware-level 2:4 sparsity acceleration
- NVLink 3.0: 600 GB/s interconnect bandwidth
Use Cases
- LLM training (7B–70B models)
- Inference deployment
- HPC scientific computing
- Recommendation systems