WYP
News Guru
Futuremark's latest VR benchmarking Utility is here, but how well does today's hardware cope with the strain?
Read more on VRMARK Cyan Room.
Read more on VRMARK Cyan Room.
VRMARK Cyan Room DX12 Benchmark - AMD VS Nvidia
- Thread starter WYP
- Start date
Dicehunter
Resident Newb
I managed a smidge over 8100 with average FPS of 176 with my Vega 64 
WYP
News Guru
I managed a smidge over 8100 with my Vega 64
Overclocked I assume?
Dicehunter
Resident Newb
AngryGoldfish
Old N Gold
This is the kind of positioning I wanted Vega to be in games. 
Dicehunter
Resident Newb
This is the kind of positioning I wanted Vega to be in games.
To be fair Vega does do well and nips at the heels of the 1080 Ti when games take advantage of it's hardware i.e Forza 7.
NeverBackDown
AMD Enthusiast
Agreed but that's the problem. When games take advantage of it. And not many do. Most people would rather just throw in a gpu and not worry about it.
Not when it comes to drivers imo AMDs is a better quality. I have had less issues, but Nvidia does still have the release advantage. Whether it's better quality or not is up to debate
Wraith
Bettyswollocks
I managed a smidge over 8100 with average FPS of 176 with my Vega 64
8960, averaging 195.33fps on the 1080Ti at stock.
https://www.3dmark.com/vrm/23532852?
Dicehunter
Resident Newb
Nice ! That's the extra £200 over V64 in action ^_^
I can't even run this thing. It just kinda quits after the first sequence. No error, no crash, even at stock speeds. Damn!
EDIT: apparently it doesn't crash, just seems to run in the background! I managed to barely break 10,000, but I doubt that result is legit. Maybe this thing doesn't like me running at 4K or something.
EDIT: apparently it doesn't crash, just seems to run in the background! I managed to barely break 10,000, but I doubt that result is legit. Maybe this thing doesn't like me running at 4K or something.
Last edited:
AngryGoldfish
Old N Gold
Yeah, but it's not to the level of the RX 480 or Fury X. The RX 480 in games that favour Nvidia still comes close. Same for the Fury X against the 980Ti and 1070. The issue with Vega is it's behind the GTX 1080/1070/170Ti in games that favour Nvidia, not the 1080Ti. There are huge disparities. It's not a consistent performer at all.
here is my score on Vega 64 LC with 50% power target and HBM2 @ 1120mhz
https://www.3dmark.com/vrpcr/5545
Agost
New member
This test shows (again) that even in best case scenario, Vega has a serious internal bottleneck and practically no architectural gains from Fiji, at least in gaming oriented applications. In professional workload it showed better clock for clock performance than Fiji though, this still confuses me.
AngryGoldfish
Old N Gold
Where are you seeing that in this new VR test? I'm not doubting or anything; I'm really curious.
Agost
New member
The poor performance scaling of shader number was clear even with Fiji, but Vega takes it to the next level, as showed by gamers nexus tests.
V64, at stock, performs about 30% better than a Fury X in Cyan Room while having an average clockspeed of about 40% higher; Vega was marketed as a completely different beast compared to Fiji (better geometry engine performance, big ipc gains, better rendering tech etc.) but still performs just like an OC'd Fury X, despite having 3.6 billion more transistors. Same for V56, which performs practically the same at the same clocks.
Heck, probably even Polaris has better perf/clock than Vega (about 15% more than Hawaii). I'm pretty sure Vega isn't particularly bandwidth starved as some say, since memory OC has a rather low impact on performance compared to core OC; the latter also speeds up geometry engines and ROPs, which I suspect are a major bottleneck for GCN in general.
Link to GN testing Fiji vs Vega: here
Link to GN testing clock for clock: here and here
p.s.: don't get me wrong, I'm critical about Vega because I had high hopes for this architecture.
p.p.s.: I'm not a software/electronic engineer, I'm just collecting data from people who can test this stuff.
Last edited:
AngryGoldfish
Old N Gold
Thanks for that, mate.
I was really more referring to this new VR test. I know that clock-for-clock Vega is not very much faster than Fury X—which also very much disappointed me as AMD kept rehashing how much more advanced Vega was. But I haven't seen a direct Fury X versus Vega 64 VRMARK test.
That said, the 56 vs 64 test I don't remember seeing. I had no idea that at the same clock speeds Vega 56 was so close to the 64. That's insane. Unless there was some error in GM's testing or if drivers needed maturing, Vega 56 with an overclock is going to be a mighty little card and much better value than the 64.
I was really more referring to this new VR test. I know that clock-for-clock Vega is not very much faster than Fury X—which also very much disappointed me as AMD kept rehashing how much more advanced Vega was. But I haven't seen a direct Fury X versus Vega 64 VRMARK test.
That said, the 56 vs 64 test I don't remember seeing. I had no idea that at the same clock speeds Vega 56 was so close to the 64. That's insane. Unless there was some error in GM's testing or if drivers needed maturing, Vega 56 with an overclock is going to be a mighty little card and much better value than the 64.
Agost
New member
You're welcome
The poor shader scaling is evident in every GCN card with more than ~2560 cores, it's from Hawaii times, it got serious in Fiji and now reaches its peak with Vega. 64 ROPs and 4 geometry engines are clearly not enough and I don't see the "11 pixel per clock with 4 engines" stuff AMD told us working at all.
IIRC nvidia geometry pipelines are called PolyMorph engines and GPUs from Kepler on have 1 per SM; the 1080 has 64 ROPs and 20 geometry engines, the 1080Ti has 96 ROPs and 30 engines (same ratio). Meanwhile both Vega cards have 64 ROPs and 4 geometry engines, just like Fiji and Hawaii. This could also explain the small difference in performance between Fiji and Hawaii too.
Vega and Fiji were probably never meant for gamers, but mostly compute (geometry engines and ROPs are useless for compute AFAIK) and that's because of AMD's lack of funds. I hope they will solve it with Navi, by sticking together more smaller GPUs (higher ROPs count) and hopefully increase the number of geometry engines per die.
AngryGoldfish
Old N Gold
I guess if games took advantage of things like Vega's Draw Stream Binning Rasterizer and Rapid Packed Math, IPC might be higher compared to Fiji?
Agost
New member
As you can see from GN tests, performance per clock is already much higher than Fiji, but that's for compute. DSBR looks like non existent and RPM is only a feature which uses FP16 for certain tasks, thus needing active developement for games and programs.
If Vega 11 will ever see light and will have something like 2048 cores and 32 ROPs, I'm pretty sure its gaming performance would be higher than 50% of V64 at the same clocks (assuming a single 1GHz HBM2 module is used). Maybe it will get even closer to 60 or 65% the performance of V64, given how the 512 SP difference in Vega 10 have almost zero impact on frame rate.
NeverBackDown
AMD Enthusiast
RPM is less useful.
DSBR is very useful. Does it require active development? I'm not sure. I am 90% sure Nvidias DSBR is always active regardless of game development. If this is the case then AMD needs more complex driver software to determine what needs culling and what doesn't
