A quick dive into a positive indicator for the Biggest Navi Reveal next month, as the launch has minimal leaks prior to its reveal next month.
A lot of people have a negative opinion on Radeon’s ability to get CU scaling over 60%. Due to a performance illusion created by the release of salvaged GPU dies with each generation (RX 5700 at 36 CUs, RX Vega at 56 CUs) that show below 60% CU scaling.
This performance illusion is created because engineers are motivated to reduce the architectures infrastructure and shared resources to bare minimum for the CUs, which lets them squeeze out some extra GPU dies per silicon wafer. This does, frequently, cause the full product (RX 5700XT at 40 CUs, RX Vega at 64 CUs) to operate a few percentage points below their full capabilities due to the complete CU package slightly lacking infrastructure and shared resources.
Here is example, from my January 2020 23 game FPS testing showing average performance difference on the RX 5700 Series.
Sapphire RX 5700 36 CUs averaging 1.715Ghz = 100%
Reference 5700XT 40 CUs averaging 1.85Ghz =
110.79% Here is a link to screenshot of those FPS results.
https://imgur.com/a/V52fWUd Normalising the clocks (my 21 game overclocking results showed 3.37% extra FPS on the 5700XT per 100mhz. This can be used in this calculation is: 135mhz times 3.37 divided by 100 minus 10.79 equals 6.24%. Finally, 6.24 times 0.91 times 10 equals 56.8%, which is CU scaling of the extra 4 CUs on RX 5700 Series.
However, this
56.8% scaling on those extra 4 CUs is actually a performance illusion, because in practice removing those 4 CUs gave the remaining 36 CUs more infrastructure and shared resources, which let the salvaged 36 CU product produce a tiny amount of extra performance per CU, which eroded the CU scaling gap between the 36 CU part and 40 CU part.
When architecture is scaled up by Radeon, they will begin with perfect infrastructure and shared resources for all the CUs and from that point of maximum die size, the engineers will spend time working out ways to reduce that infrastructure and shared resources for the CUs and this will be used to raise the number of GPU dies per silicon wafer, which with Radeon invariable leads to better pricing than Nvidia for us; PC gamers.
Therefore, using scaling from salvaged GPU dies, that have this performance illusion, can lead people to have an unduly pessimistic outlook on what Radeon’s engineers can achieve given ample time and ample funding.
Radeon has done a rapid refresh and scaling up of CUs in the same year on the process node not so long ago! In 2017, AMD did a Polaris Refresh (RX 580) on GlobalFoundries 14nm node and later that year they did a scaling up of GCN with RX Vega launch (RX Vega 64). Therefore, there is no need to guess about what kind of CU scaling the Radeon Division has been told to achieve!
The RX 580 base clock (1257mhz) is 10mhz faster than RX Vega 64 base clock 1247mhz, but the boost clock was around 200mhz higher for RX Vega 64. RX Vega CUs power efficiency was remarkably like the Polaris CUs as well. Therefore, RX Vega did look like a bigger version of Polaris, but redesigned to be extremely attractive for datacentre buyers.
Using older benchmarks from the website “The Guru of 3D” (carried out in 2018 for RX 590 launch across 13 games) gives us an average performance difference between the RX 580 versus RX Vega 64 at the 2560x1440p of +60.92%. Those averages can be used to calculate the CU scaling for Polaris versus RX Vega.
RX 580 36 CUs = 100%
RX Vega 64 CUs = 160.92%
100 divided by 36 CUs times 64 CUs = 177.77%
Net difference = 16.85%
CU Scaling =
79.70% Here is link to a screenshot of “The Guru of 3D” FPS result as a head-to-head comparison.
https://imgur.com/a/WZjAwkQ From the figures above, the last time it was commercially viable for AMD to release gaming GPUs at ultra highend on the same node as their lower performance GPUs they targeted scaling of 79.70% on the CU increase.
To conclude, it should be expected, that the Radeon Division will have achieved a 70% to 80% CU scaling for them to be proceeding with Biggest Navi gaming GPU launch later this year and general pessimistic sentiments are misplaced.