While we just recently reported on alleged AMD Big Navi specifications, it seems to have merit as yet another well-timed leak seems to unveil ridiculously powerful Nvidia GPU's in the works.
On Twitter, both _rogame and w_at_ar_u have been able to reveal not only specs, but also in-depth numbers and Geekbench benchmarks of a number of upcoming electronics, including the Ampere GPUs from Nvidia.
While Ampere in itself is nothing new, no one had imagined the massive specs that look too good to be true to be revealed - it is a massive leap in specs and performance from previous GPU architectures.
As each Streaming multiprocessor has 64 CUDA Cores, and there are 124 Streaming Multiprocessors, we can specify 7936 CUDA Cores. It makes sense as it is a 55% increase from the Tesla platform and its Tesla V100 with 5120 Cores, so still a realistic specification with a clock speed of 1.1GHz, this should result in 18 TFLOP of computing power.
The Geekbench 5 OpenCL score of 222377 is achieved with 32 GB HBM2e memory at 1200 MHz, using a 4096 bit bus interface like the Big Navi. A whopping 32 MB L2 cache, more than five times that of the Volta GV100.
Another version of Ampere comes with 118 Streaming Multiprocessors, 7552 CUDA Cores, 24 MB L2 cache, and 24 GB of HBM2e memory while using a 3072but bis at 1200 MHz clock speed, giving it 16.7 TFLOPs, and resulting in a 184.096 OpenGL Geekbench score.
A third variant is interesting, sporting 46.8 GB HBM2e memory, but with only 108 Streaming Multiprocessor units (6912 CUDA cores) and 1.01 GHz clock speed. Geekbench score this time is only 141.654.
Nvidia has a large keynote at GTC - so perhaps we will soon know more?
<bild></bild>