Nvidia Pascal Whitepaper Released with More Info on GP100 GPU
When Nvidia first announced its Pascal architecture it gave details for what to expect. Now, the company has released Whitepaper for its next gen architecture.
It is important to mention that when Nvidia and AMD release those TFLOP numbers on the Internet for their un-released graphics card, fans need to understand that those numbers are theoretical. For a graphics card to achieve that theoretical performance, its architecture needs to be very efficient. Nvidia GP100 GPU is no exception.
We already know that Nvidia has cut the Stream Multiprocessor on the GP100 from 128 FP32 to 64, which allows the graphics card to better distribute its processing power to perform tasks. Since each SM has the same amount of registers and other hardware, throughput is also increased.
Nvidia has also made some improvements to SMs so that datapaths are more streamlined and sharing of information through them takes less energy. Scheduler has also been improved so that SM has always a task to perform.
Cache sizes for Nvidia Pascal GP100 GPU has also been upgraded from 3 MB to 4 MB. A dedicated shared memory space of 64 KB has also been added to per Stream Multiprocessor, however this is lower than Maxwell which had 96 KB per SM.
If you consider that Nvidia Pascal has double the amount of Stream Multiprocessor on-board compared to Maxwell chips then it’s an increase of 16 KB per Stream Multiprocessor. Nvidia also detailed that Pascal will have HBM2 memory and also provided its layout.
What do you think of this Pascal architecture? Will Pascal help Nvidia to keep its dominance in the graphics card market?