AMD Vega 20 benchmark results spotted

[HEXUS]

Very early results aren’t great but are interesting to look over, with various specs revealed.

Read more.

[ajax2061]

Will be looking to buy a card this year as fury struggling in some games at 1600p (16:10 screen) could be a choice time will tell

[lumireleon]

....so they have developed their own version of TENSOR core?

[WildW_UK]

32GB? I thought there was a shortage of that stuff.

[Xlucine]

32GB? I thought there was a shortage of that stuff.

It's aiming to replace the FE vega, which had 16GB. Consumer GPUs will have less RAM

[tygrus]

If it's an early sample then they are often clocked far slower than the final plus uses unoptimised drivers that contain more debug code turned on in the background. I very much doubt the above numbers represent the performance of the final product.

[deksman2]

If the 7nm TSMC process (which Vega 20 uses) is designed for High performance (like Pascals 16nm was), then Vega 20 should be able to clock much higher in comparison to Vega 10 at much lower power draw... or at the very least, AMD should be able to clock Vega 20 to same clocks as Vega 10 at much lower power draw.

Remember the reasons why Vega 10 needed a lot of power was because of:
1. 14nmLPP which was not designed for high clocks (and was efficient for AMD GPU's running at around 1100Mhz).
2. 1000 more compute units/stream processors than Pascal (a difference of 40%).
3. Lower yields on 14nmLPP resulting in higher voltages for Vega.

If this 7nm TSMC process is indeed for high performance, I would expect higher core and HBM clocks (increasing HBM clocks on Vega 10 resulted in most performance gain for minimal power draw - this new process should allow AMD to clock the HBM and core higher).