AMD is preparing their brand new 7nm based Vega 20 GPUs which will be featured inside the next-gen Radeon Instinct products. AMD talked about the new GPU at Computex 2018, giving us a sneak peek at the upcoming server parts but a detailed analysis has been done by Ascii.jp which reveals it to be an HPC powerhouse in the making.
AMD 7nm Vega 20 GPUs May Break The 20 TFLOPs Barrier, Estimations Reveal – Require Lots of Power, Really Expensive To Make
AMD is the first and only GPU maker to announce a 7nm product which will be launching in the 2nd half of 2018, as confirmed by AMD themselves. The new 7nm process based Vega 20 GPU will come with a range of new features while offering twice the density compared to 14nm Vega 10 GPU since AMD is packing more stuff in a smaller die and will also be twice as much efficient due to enhanced architecture design and process maturation.
The Vega 20 GPUs are tailor-made and optimized for HPC, Automotive and AI markets. As such, they have been equipped with new Deep Learning OPs, HPC algorithms and outfitted with the latest I/O interface (PCIe 4.0). With all the new tech and process enhancements involved, Vega 20 can turn out to be a compute powerhouse, allowing RTG to tackle their competitor’s HPC solution, the Tesla V100, which has been in the market for over a year now.
Based on the estimation done by Ascii.jp, they are rounding the die size of the Vega 20 chip around 360mm2 while the Vega 10 chip measured at 510mm2. So that’s about a 70% reduction in the total die size. But here’s the thing, according to GloFo, 7nm LPP offers a 2.8 times increase in logic density which means AMD is using their own cell library rather than Global Foundries which comes with a 2x increase over the 2.8x increase as GloFo states.
If the Vega 20 GPU was to be made on a 14nm LPP process, it would have measured at 720mm2. If the same calculation is driven to form a relation between the GPU die and Compute units, we would get up to 88 CUs (up to 96 CUs) on such a chip. That’s a 37.5% jump from the current 64 CU design on the Vega 10.
It is being concluded that the extra space could mean a grouping of extra CUs on the chip. If that is the case, which seems unlikely since we have seen previous information and slides mentioning 64 CUs on the Vega 20 GPU, the other factor that would be considered is the frequency. Now the clock rate on 7nm LPP is a big trick, you either get 55% power savings from the same frequency as 14nm or you can get 40% faster clock rates from the same power as 14nm. Here’s the thing though, if you were to aim for 40% faster clocks at the same power package, you’d also have to feed power to the extra HBM2 VRAM (4 Stacks vs 2 Stacks), the new DNN instruction sets. The power package under such a scenario would easily exceed 400W.
So taking a more conservative approach, the site estimates a 20% jump in clocks, saving around 30-40% power (300-350W), and a performance jump of 65%. This would lead to a jump from 12.66 TF to 20.9 TF. A remarkable jump indeed which would blow past the Tesla V100 (15.7 TFLOPs at max) in a 300W package. This is all theory at the moment but even without a core bump, the clock speed estimates are very likely to happen so it would be a jump in performance for HPC regardless.
AMD Vega 20 GPU Not A Gaming GPU – Price and Production Cited as Main Constraints
The other part of this story is whether Vega 20 would be a gaming GPU or not. In previous rumors, we have seen that the possibility of Vega 20 arriving for gaming is just impossible. That is being re-affirmed by Ascii, who state that price and production would be the main constraint in delivering such a GPU to gamers.
Vega 20 was always meant to be an HPC product with Navi taking the helm of the next Radeon RX gaming part (more on that here). The thing is that Vega 20 is just super expensive to manufacturer. A single 7nm wafer costs north of $10,000 US and the prices keep on climbing. As if the price wasn’t enough of an issue, production of 7nm parts is also very complicated and won’t hit high volume until 2019.
So those hoping for an AMD 7nm Vega 20 based graphics card, don’t expect it at all and even if AMD somehow managed to make a product for gaming, expect your wallets to be really deep and ready.
Do you think AMD will be breaking the 20 TFLOPs barrier with the Vega 20 GPU?