NVIDIA A16 16GB

The NVIDIA A16 is a quad-GPU board built for graphics-rich virtual desktops, and it is in this database because it turns up on GPU marketplaces as four independently rentable accelerators rather than as one card. Each of its four GA107 GPUs carries 1,280 Ampere CUDA cores, 40 third-generation Tensor Cores and 16 GB of GDDR6 at 200 GB/s, which NVIDIA rates at 4.5 TFLOPS FP32, 9 TFLOPS dense TF32, 17.9 TFLOPS dense FP16 and 35.9 dense INT8 TOPS. That is a modest accelerator by datacenter standards and it is not trying to be anything else: the A16 exists to put a real GPU and a decent frame buffer in front of as many concurrent users as possible, up to 64 per board, with four hardware encoders and eight decoders to feed them. For inference work the useful comparison is the ratio rather than the absolute figure, since its Tensor Cores sit in the professional Ampere bucket, running TF32 at twice the shader rate and FP16 at four times it, with structured sparsity doubling both again. The whole board draws 250 W, so four of these GPUs cost about what a single mid-range datacenter card does.