Pascal Isn’t Done, Not By a Long Shot

GTX 1080 is awesome, but Nvidia is holding back

WE’VE KNOWN ABOUT Pascal for what seems like ages, and after five years of 28nm, GPUs finally have a new process node. Nvidia has already launched its GTX 1080 (reviewed on pg. 74), with GTX 1070 coming soon, both using the company’s GP104 architecture. The GTX 1080 is a beast that has devoured all contenders—just look at the bones! But while we’re suitably impressed, this is definitely not the biggest GPU that Nvidia could make.

That title currently goes to GP100, which is available in the Tesla P100. Except it’s not really available as a stand-alone card; if you want GP100, you’ll have to buy one of Nvidia’s $129,000 DGX-1 servers. At least that will net you eight Tesla P100s, which means they’re around $12,500 a pop.

As fast as GTX 1080 is, all you have to do is look at some of the specs of GP100 to guess where Nvidia will go next. GP104 has 7.2 billion transistors, a step down from the 8 billion transistors in the Titan X’s GM200, but 16nm FinFET has brought higher clocks, a smaller chip, and lower prices. GP100, in contrast, goes big, with 15.3 billion transistors and a 610mm2 die that’s nearly twice the size of the GP104. It also packs 16GB of HMB2 memory, capable of 720GB/s, and FP64 runs at half the rate of FP32 (compared to 1/32 on GP104).

Will we see GP100 in a Titan or 1080 Ti card in the future? Perhaps. Or Nvidia could do a consumer card without the FP64 and HBM, but with more FP32 cores and 12GB GDDR5X memory. Whatever the company builds, though, it will certainly trounce our current darling, the GTX 1080.