The next high-end silicon architecture from NVIDIA Pascal under codename GP100 will be a beast in terms of the numerical calculation. According to the leaked slide from NVIDIA research fellow, the company has revealed that the chip is able to offer double precision floating point decimals (DPFP) of no less than four TFLOPs, increasing at three times the performance thrown by a Tesla K20 (1.31 TFLOPs) based on “Kepler” GK110 silicon.
NVIDIA Pascal GP100 silicon to Feature 4 TFLOPs DPFP
The same slide reveals performance decimal floating point (SPFP) of 12 TFLOPs, four times compared to GK110 silicon and almost double compared to Maxwell GM200. We can also see that GP100 silicon may unite HBM2 memory and GDDR5X memory. The 1024 GB / s bandwidth mentioned in the slide belongs to the HBM2 memory.
This would be surprising in any case, pointing to a re-construction of the other shader cluster since the last Maxwell architecture out, possibly again the addition of dedicated Double Precision units. Likewise 12 teraflops compute power is for a professional graphics card, which usually does not compete with fully extended clock rates, point to a real monster chip. The GPU based on this chip would not exceed the clock rate of 1000 MHz with 6144 shader units – and thus a smooth doubling of what the GM200 chip has to offer (3072 shader units).