NVIDIA GeForce GTX TITAN X 12GB

The GeForce GTX TITAN X is the card a great deal of early deep learning was actually trained on, and it is in this database as the floor of the whole catalog: the last NVIDIA flagship before half precision, mixed-precision training and Tensor Cores existed at all. Its GM200 die carries 3,072 CUDA cores across 24 SMs at a 1,075 MHz boost, which works out at 6.60 TFLOPS of single precision, and NVIDIA paired it with 12 GB of GDDR5 on a 384-bit bus at 336.5 GB/s -- a frame buffer no other consumer card of 2015 came close to, which is precisely why researchers bought it. What it cannot do is the interesting part. NVIDIA's own CUDA documentation lists half-precision throughput on this compute capability as not available, so there is no FP16 path to fall back on, no integer inference path, and nothing resembling a tensor unit; everything runs in FP32 at 6.6 TFLOPS. Set against a modern accelerator quoting thousands of teraflops of FP8, that number is the measure of how much of the last decade's progress came from precision rather than from transistors.