Little is known at the moment of the Vega 20 card that AMD presented at the Taipei Computer Show and that as we confirmed will come in a lithography of 7 nm thanks to TSMC and with 32 GB of VRAM HBM2 in its version for Radeon Instinct. From the presentation by Lisa Su and the filtered data of TSMC, a series of conclusions have been drawn up to the point that Vega 20 could be up to 65% faster than Vega 64.
Although Radeon Instinct goes in another way totally different from what the normal user looks for and needs, the TCS left brushstrokes of what Vega 20 is going to offer.
The sending of samples takes months between the beta testers but it is not necessary that in principle nothing leaks since Lisa Su already showed the size of the complete die with its HBM and its interposer, as always all encapsulated in the same PCB.
Given that TSMC figures the improvement of transistors in 2.8 times the capacity compared to 14 nm, how big is Vega 20? Starting from several bases such as the size of Vega 64 is 510 mm and that the HBM2 is 7.75 × 11.87 mm, it can be concluded that with a gain of 70% of the area the size of Vega 20 it would be approximately 15.0 × 23.9 mm = 358.5 mm2, which rounding off could be 360 mm2.
This data is necessary to be able to calculate what TSMC calls ” standard cell libraries ” and that in its 14 nm it was 7.5 tracks compared to the 6 tracks for 7 nm. This is important, since AMD said it did not use the libraries for its chips made by Global Foundries, but its own, which TSMC seems to be going to happen again as it improves the design of its chips at the cost of losing density in the process.
Specifically, it seems that the final density of TSMC wafers will be around twice the density of 14 nm vs 7 nm, or what is the same, if Vega 20 had been manufactured in 14 nm its size would be around 720 mm2, a barbarity of chip.
Then, knowing this, how many CU could host Vega 20 in 7 nm? Easy, the theory says that the double having twice the “space” (128 CUs), but being more realistic, we must assume that they would be able to introduce between 88 CUs and 96 CUs according to the limitations of their architecture.
Assuming they add 88 CUs and that the GPU manages to maintain the same frequency as Vega 64 we would be talking about approximately a linear improvement of 37.5%.
To this pure performance data (maintaining the same architecture without any improvement) we must add another one from TSMC. In the 2017 IEDM it affirmed that it could achieve an improvement of 55% of the consumption at the same frequency or instead, increase the yield by 40% at the same consumption as the previous generation.
The data is very illuminating, although it is necessary to start from Vega 64 to 14 nm consume 295 watts, we could estimate that with the 88 CUs and the increase of the channels of HBM2 ( doubling performance ) together with a reduction of the consumption of 40 % would give frequencies with boost close to 1.7 GHz being the base around 1.4 GHz. If this improvement in consumption does not occur, we would be talking about Vega 20 could reach 400 watts in 14 nm, something unacceptable.
The consumption is not at all clear, in the presentation that was made it was possible to see a PowerColor card that could well be the equivalent in Vega 20 to the Vega 56 ITX that was presented more than a year ago, but this time it has been seen as instead of a single 8-pin PCI-e connector they have had to include 8 + 6 pins. Perhaps it is an indication that consumption will rise?
Going back to the above, all this would raise the performance around 65% -67% on Vega 64 with quite impressive figures of about 20.9 TFlops in SP, if finally the consumption does not trigger and AMD does not take the path of NVIDIA with exclusive units for RT.
If instead opts for the path marked by NVIDIA and ignores Microsoft, it is clear that the performance in SP will plummet and that the differences will not be reduced as expected.
This is evidently theory, what we know for sure is that the manufacturing costs of 7 nm LPP of GF compared to 14 nm are more than doubled, around $ 10,000 per wafer and it is estimated that TSMC has some costs lower but also above 14 nm (quite a lot).
This would mean that it is very possible that the new Vega 20 have a higher price than their versions in Vega 10, so that the final price for the consumer can be increased.