NVIDIA A100 Tensor Core GPU Architecture white paper (Ampere) - Table 4 Comparison of NVIDIA Data Center GPUs

Vendor primary architecture document, and the ONLY first-party NVIDIA specification that describes the 40 GB A100. The public A100 datasheet (nv-a100-datasheet) carries spec columns for 80GB PCIe and 80GB SXM only; the 40 GB cut is documented here instead. Two first-party statements fix the 40 GB part's memory exactly: (a) prose, 'The A100 GPU includes 40 GB of fast HBM2 DRAM memory on its SXM4-style circuit board. The memory is organized as five active HBM2 stacks with eight memory dies per stack. With a 1215 MHz (DDR) data rate the A100 HBM2 delivers 1555 GB/sec memory bandwidth'; and (b) Table 4 'Comparison of NVIDIA Data Center GPUs', rows 'Memory Interface 5120-bit HBM2', 'Memory Size 40 GB', 'Memory Data Rate 1215 MHz DDR', 'Memory Bandwidth 1555 GB/sec', 'GPU Board Form Factor SXM4', 'GPU Boost Clock 1410 MHz', 'TDP 400 Watts', 'SMs 108', 'FP32 Cores / GPU 6912', 'FP64 Cores / GPU (excl. Tensor) 3456', 'Peak FP64 TFLOPS (non-Tensor) 9.7', 'Peak FP64 Tensor TFLOPS 19.5', 'Peak FP32 TFLOPS (non-Tensor) 19.5', 'Peak FP16 Tensor TFLOPS 312/624', 'Peak BF16 Tensor TFLOPS 312/624', 'Peak TF32 Tensor TFLOPS 156/312', 'Peak INT8 Tensor TOPS 624/1248', 'Peak INT4 Tensor TOPS 1248/2496', 'Transistors 54.2 billion', 'TSMC Manufacturing Process 7 nm N7'. Table 4 is a single 'NVIDIA A100' column, so every figure in it describes the 40 GB SXM4 die as originally shipped. The slash notation in the tensor rows is dense/sparse, matching the A100 datasheet's '* With sparsity' footnote, so the DENSE value is the first number of each pair. URL verified live and served as application/pdf on 2026-10-03.