For their successor to Maxell 2.0 microarchitecture, Pascal, NVIDIA is pulling out all of the stops, especially for their high-end and enthusiast-grade GPUs.
In fact, their GP100 chip, the GPU powering whatever they eventually name their Pascal family, sports 17 billion plus transistors, a significant amount of HBM2 VRAM and about every other piece of tech NVIDIA could shove onto the silicon.
And thanks to the 16nm FinFET process, it’s a lot more than was possible on the older 22nm process. But what about lower-end GPUs. Surely all of that tech will mean fairly expensive cards, way out of the price range of the average consumer.
Enter the GP104. According to rumours reported on by TweakTown, NVIDIA is testing their GP104 GPU alongside a successor to GDDR5 RAM, GDDR5X.
GDDR5X RAM will allow for an estimated bandwidth of 448GB/sec, faster than even the GTX Titan X, but it won’t be able to compete with AMD’s Fury range, so who is GDDR5X for?
The GP104 nomenclature suggests a GPU powering the Pascal equivalent of the GTX 980 and GTX 970, while the GP100 powers the more monstrous GTX Titan X and GTX 980Ti equivalents, but that’s a lot less bandwidth than we were expecting for what is supposed to be high-end graphics cards – the GTX 1080 and GTX 1070, or whatever they call them.
HBM 2 will more than likely mean a higher price tag, but is the reduced performance worth it, or is NVIDIA shifting gears and targeting these cards at the midrange to high-end while the GP100 cards will be more run-of-the-mill enthusiast cards.
If we think about it a little more, GDDR5X does make a lot of sense. With the greatly increased compute performance of the GP104 over the GM204, the cards it powers will likely reach the performance of the GTX Titan X, or just about, and a noticeable increase in bandwidth over the GTX Titan X should mean removing any bottlenecks.
The immense bandwidth provided by HBM2 will likely be wasted on these cards. And with 440+GB/sec of bandwidth, these cards should more than handle 4k resolutions as well as VR tech for the foreseeable future.
It will mean a much cheaper price tag than if NVIDIA used HBM2, so will it mean seeing cards that are a lot faster than AMD’s Fury X but at a much more affordable price. AMD is in trouble if that’s what happens.
|GPU Family||AMD Arctic Islands||NVIDIA Pascal|
|GPU Name||AMD Greenland||NVIDIA GP100|
|GPU Process||TSMC 16nm FinFET||TSMC 16nm FinFET|
|GPU Transistors||15-18 Billion||~17 Billion|
|HBM Memory (Consumers)||4-16 GB (SK Hynix) HBM2||2-16 GB (SK Hynix/Samsung)
|HBM Memory (Dual-Chip Professional/ HPC)||32 GB (SK Hynix) HBM2||32 GB (SK Hynix/Samsung) HBM2|
|HBM2 Bandwidth||1 TB/s (Peak)||1 TB/s (Peak)|
|Graphics Architecture||GCN 2.0? (New ISA)||Next-CUDA (Compute Oriented)|
|Successor of (GPU)||Fiji (Radeon 300/Fury)||GM200 (Maxwell)|
*Table courtesy of WCCF Tech.