Nvidia likes ‘em huge. The first GV100 Volta GPU Nvidia confirmed off is a mammoth 815mm2, however whereas the gaming Volta chips aren’t going to be fairly that enormous, they’ll nonetheless be the largest GPUs you’ll have ever seen on a gaming card.
What’s one of the best GPU for the time being? Well, that is our decide of the best graphics cards… assuming the crypto-miners haven’t nabbed all of them.
When CEO, Jen-Hsun Huang, began waving across the first Volta card, the Tesla V100, at their GPU Technology Conference again in May he was calling it a slice of silicon “at the limits of photolithography, meaning you can’t make a chip any bigger than this.” And, provided that even utilizing TSMC’s diminutive 12nm transistor know-how the GPU is that damned huge, I’m inclined to imagine him.
There are, in any case, 21 billion transistors contained in the V100 chip and an enormous variety of CUDA cores too. As detailed within the Volta whitepaper Nvidia launched this week, the full-fat GV100 GPU sports activities a complete of 5,376 CUDA cores, whereas the Tesla V100 accelerator is barely going to be accessible utilizing 5,120 of them.
Only 5,120 cores? Pfft, small fry…
If Volta had been to observe the identical course travelled by the Pascal social gathering we’d be taking a look at a Titan Xv being launched subsequent 12 months with 5,120 cores. But it’s simply not sensible to launch a graphics card with a chip the scale of the GV100, so the cut-down GV102, which can type the idea of Volta’s Titan playing cards and a possible GTX 2080 Ti, should be a lot smaller than the 815mm2 of the GV100.
With Pascal, the die measurement dropped from 610mm2 to 471mm2, going from the GP100 to GP102, and but nonetheless retained the identical core rely because the skilled chip as a result of it ditched each the HBM2 on-die reminiscence and the additional double precision cores.
Even taking that under consideration (we’re expecting Nvidia to use GDDR6 not HBM2 with the gaming playing cards) the consumer-facing Volta GPUs will have to be round 630mm2 in the event that they need to squeeze over 5,000 CUDA cores inside, making them the most important gaming chips Nvidia could have ever produced. And that’s additionally going to make the Volta Titan a mighty costly chip to provide, not to mention purchase.
That ‘biggest chip’ title was beforehand held by the GTX Titan X and GTX 980 Ti’s Maxwell-powered GM200 GPU. That topped out at 601mm2 with three,072 CUDA cores. On the purple facet of the graphics divide, AMD’s Fiji GPU was their largest in latest reminiscence at 596mm2 with four,096 of their GCN cores.
There have been rumours that AMD’s competing Vega architecture will launch with chips at over 500mm2, however they have not launched any official statements about the actual measurement of the preliminary Vega 10 GPU.
Of course the extra mainstream shopper Volta variants aren’t going to deal with a chip of that form of scale, however they’re nonetheless more likely to be fairly beefy. If Nvidia observe the Pascal trajectory then the GV104 GPU might nonetheless find yourself being between 420mm2 and 480mm2.
With the earlier technology, the three,584cores/28 SMs of the GP102 GPU translated into 2,560 cores/20 SMs for the GTX 1080’s GP104, so I’d guess the Volta equal could be one thing like a drop from 5,120 cores/40 SMs to four,096 cores/32 SMs for a GTX 2080 card.
Though, with Volta, Nvidia have modified from Pascal’s 5 streaming microprocessor (SM) per normal processing cluster (GPC) structure to 1 which has seven SM’s per GPC. If they maintain this for the extra mainstream chips then the GTX 2080 may very well find yourself at 28 SMs (4 Volta GPCs) which might give it the identical CUDA core rely because the GTX 1080 Ti at three,584. And that’s bought a pleasant symmetry about it.
But hey, that’s an entire lot of numerical hypothesis proper now, however what we will be certain of is that when Volta does arrive in a extra consumer-oriented type it’s going to be large.
Source
