Quadro customers have been getting a lot of love over the last few years. The current generation of Pascal cards such as the P5000 and P6000 are industry-leading products.
The thing is, when it comes to professional GPU applications raw performance isnít all that matters.
This new upcoming card from Nvidia doesnít actually provide more performance than a P6000. It doesnít significantly alter the core Pascal architecture or really innovate in big ways. What it does do is bring much-needed features to science and engineering pros who need to do very specialized work that a P6000 canít.
A Numbers Game Before we get into what those features are, first let's looks at the technical specifications for the GP100. This is a dual-slot card, which means you wonít need a special case for it or anything like that. Itís also just as power-efficient as other Tesla parts, maxing out at a frugal 235W. There are 3584 CUDA cores, which is a massive number, but par for the course in a Quadro Pascal card. RAM stands at 16GB, also nothing to write home about anymore.
The Big Difference So whatís the fuss about? First, that RAM is not GDDR as weíd find on a P6000. Itís the new HBM or High-bandwidth Memory we first saw on AMD cards. Thereís no more 4GB limit here and you get an incredible (approximate) 720GB/s of memory bandwidth. Thatís not far from double what GP102 cards have.
Of course, some cards can now have 24GB of GDDR5X, so if your applications needs more than the 16GB that HBM2 provides the GP100 may not be for you. You can however hook multiple GP100s together with NVlink connections, perfect for CUDA-heavy applications. It just wonít help with RAM capacity.
The most important feature has to be the support for FP64 instructions at 5.2TFLOPS.
This has been achieved through direct architectural changes and the FP64 performance absolutely destroys what the P6000 can do. Itís also going to be much faster at FP16 operations thanks to the double-speed stacked instructions.
The Time has Come So depending on the specific nature of your compute project, the GP100 may turn out to be an absolute bargain. It launches in March 2017 and we donít know exactly how much it will be, but somewhere North of six grand is likely.
This is definitely not a card for everyone but we already have a good idea which of our customers could do with a GP100 or two in their future machines. So you can look forward to another epic Titan build when these monsters hit the market later this year.