AMD RX Vega

Just an update: The insta crash that happened with 2.8 Eevee’s shadow cube size above 512px is fixed with the 8f01623ef31 version, thanks to Clément Foucault.

My guess is that with the limitation there is not enough vram, so it starts using the system ram, which is much slower!

1 Like

Followup on the RX Vega
11 days of 24x7 rendering
Zero crashes

Scene average render times per frame (1080p)
Dual 1080 system 7min
Quad RX Vega 64 system 1min

1080s take ~ 25s to prep the scene
Vegas take ~ 35s to prep the scene

25s to finish the frame after 35s prep is stunning
Big change on animation was activating the “Persistent Images” option under Render.Performance settings

Denoising is way faster on the Vega than 1080
Prep is slower
Loading kernel can be insanely slow (up to 5min) but thats only for the first time you render

All in all I would now strongly back anyone getting RX Vega cards as a work horse for Blender.
They are outperforming the 1080ti cards in some of my contractors PCs
They cost about $350 less than a 1080TI
At stock they are using ~ 50w less than a stock 1080ti

Only thing I hate is the long kernel load times.

@Snowpz I have not played around with the Vram… didnt even know you could. Not had ANY ram problems though, not even with the monster scene I have been animating the past few weeks

3 Likes

Cool, thanks for the benchmarks, the fast HBM2 VRAM of the Vega might help here, now if the Vega 64 had more VRAM, it would be fantastic, cant wait to see if AMD really comes up with affordable 32GB cards before the end of the year, this should be interesting!

1 Like

Thanks for doing more tests. That’s what i was writing about… :slight_smile:

BTW
Had also mentioned, compilation times are longer on Vega just because of OCL.
For fair comparison also use Cycles OCL with NVidia cards (or as you did before test with RadeonPro, Indigo & LuxCore - all are OCL+CPU engines) - then you’ll feel even better :wink:

PS
Vega automatically uses OoC memory as it hits the limit.