While we did not have a chance to even see the Pascal GPU at CES 2016 show earlier in January, it appears that the compute performance of the GPU has leaked in a new academic paper.

Nvidia has promised a lot when it comes to Pascal GPU and although it has been rumored that the GPU is actually in its testing phase, we did not had a chance to see it during CES 2016 show but now it appears that we at least can get an idea regarding the compute performance of the GPU thanks to the leaked academic paper.

Originally spotted by 3DCenter.org, the academic paper, created by Manual Ujaldon, a Spanish university professor and CUDA Fellow, is talking about the benefits of new memory technologies, comparing HBM, GDDR5 and DDR3, and their use in compute dependent environments. While this is not that important, the paper actually shows the theoretical double- and single-precision performance for Nvidia's flagship Pascal GPU, the GP100.

According to the slide, the GP100 could provide up with up to 4 TFLOPs of double-precision computing performance and 12 TFLOPs of single-precision compute performance, which adds up to a 1/3 DP to SP ratio, which is similar to the Kepler architecture. This is quite a performance boost compared to the Kelper GK100-based Tesla K20X GPU which peaked at 1.31 TFLOPs of double-precision and 3.95 TFLOPs of single-precision compute performance.

The slide also reveals the memory bandwidth details set at 1TB/s, which suggest that the GP100 Pascal GPU will use HBM2 memory.

Hopefully, we will hear more about Nvidia's GP100 GPU soon and while these compute performance numbers are high, bear in mind that these will be reserved for Nvidia's professional lineup.






Source: via Techpowerup.com.