We’re not at all surprised that NVIDIA’s Pascal cards have been chosen as the processor of choice for hurricane tracking, its confirmed specs are out of this world.
As reported by VR World, at the Japanese equivalent of NVIDIA GTC (GPU Technology Conference), NVIDIA unveiled the specifications for their upcoming GPU family, Pascal.
Making use of TSMC’s (Taiwan Semiconductor Manufacturing Company) 16nm FinFET process, and the 3D transistors it allows for as a result, Pascal GPUs allegedly feature as many as 17 billion transistors – for reference sake, that’s more than the transistors of the combined transistor count of NVIDIA’s GeForce GTX Titan X (GM200) and AMD’s Radeon R9 Fury X (Fiji XT), 8 billion and 8.9 billion transistors respectively.
Alongside the significant boost in computing power for the likes of the next GeForce and Tesla ranges, it also equites to noticeable power efficiency gains and, as a result, far less excess heat (heat pollution).
Rather unfortunately, however, NVIDIA confirmed that Pascal GPUs will only launch with a middling 16GB of HBM2 memory, and what we really mean when we say middling is “fantastic” – for the gaming-orientated GP100 GPUs at least.
16GB is a lot of VRAM, far more than 4k resolutions will ever require, especially when you take the sheer bandwidth of HBM2 into account.
The GDDR5, in the GTX Titan X, for example, can manage (at best) 336.5GB/s thanks to its 384-bit interface and 7 GHz memory clock. And AMD’s Radeon R9 390X, even with its larger 512-bit bus, still can’t break 400GB/s, sporting a bandwidth of 384GB/s.
So HBM2’s bandwidth of 1TB/s is rather insane, to say the least. And, depending on memory manufacturers says NVIDIA, 32GB of HBM2 is a possibility, though that will unlikely be available for gaming cards. Besides, we can’t think of a single gaming scenario where 16GB of VRAM capable of a 1TB/s bandwidth is not sufficient.
It’s still worth considering, however, that dual-GPU Tesla cards powered by Pascal would mean 32GB HBM2 and 2TB/s of bandwidth – we’re pretty sure the USS Enterprise isn’t working with 2TB/s of bandwidth.
Pascal will mean far higher TFLOPS performance figures than any GPU prior, so expect an equally shocking price tag. So we’re willing to bet that the rest of the Pascal family sports GDDR5X memory instead.
We can’t wait to see how AMD’s Greenland GPUs compare.