If you combine the transistor count of NVIDIA’s GeForce GTX Titan X (GM200) and AMD’s Radeon R9 Fury X (Fiji XT), that’s 8 billion and 8.9 billion transistors respectively, you’d get a total 16.9 billion transistors.
Even then, you still don’t quite have the absurd 17 billion transistors that NVIDIA’s Maxell 2 successor, Pascal, packs on its silicon. That’s according to a report by Fuzilla.
Going from a 28nm process, as used in the development of Maxwell 2, to a 16nm FinFET process has done wonders for the GPU and the sort of density NVIDIA has been able to accomplish.
As a result, Pascal and its GP100 beastly compute orientated microarchitecture will be used in the creation of new cards for NVIDIA’s GeForce, Quadro and even Tesla ranges. The Quadro and Tesla being their industrial-grade workhorse GPUs.
It’s all well and good increasing the compute performance of your GPU multiple times over, but if the GPU cannot deliver that information via the necessary bandwidth, then it’s really a rather pointless endeavour.
To accommodate the massive transistor count, Pascal is going to pack a lot of VRAM, and we mean a lot.
Thanks to high-bandwidth memory 2 (HBM2), the 4GB limitation of HBM1 (as seen on the AMD’s Fury range), is no longer an issue.
In fact, the increased density of HBM2 allows NVIDIA to stack enough VRAM on the GP100 that we’re expecting between 16GB and 32GB on their top-end cards.
Keep in mind that as large a pool of memory as that is, it’s also connected to a 4096-bit bus and a heck of a lot of bandwidth. According to WCCF Tech:
With 8Gb per DRAM die and 2 Gbps speed per pin, we get approximately 256 GB/s bandwidth per HBM2 stack. With four stacks in total, we will get 1 TB/s bandwidth on NVIDIA’s GP100 flagship Pascal which is twice compared to the 512 GB/s on AMD’s Fiji cards and three times that of the 980 Ti’s 334GB/s.”
A 1 TB/s bandwidth is nothing to scoff at. It should eat 4K and even 8K resolutions for breakfast.
The GP100 Pascal cards will also come equipped with NVLink, a new interconnect technology we describe in our Pascal piece last week.
It essentially offers 5 to 12x faster transfer speeds of vanilla PCIE 3.0. It only makes sense to improve the throughput of the GPU with the rest of the PC.
It makes little to no sense to create this graphical beast and then have it unable to send its visual and computational concoctions to the CPU and RAM at a sufficient speed to prevent severe bottlenecking.
And if this is what Pascal is capable, imagine what NVIDIA is cooking up for Pascal’s successor, Volta.
Good luck AMD, that’s all we have to say. You may have dibs when it comes to HBM2 chips, but you’d better make one hell of a GPU to boot.