AMD Vega Hardware Reviews

CaptainGinger said:
Looks like the Nvidia cards are underperforming rather than any Vega magic, the RX 580 has a higher 99 percentile result than the 1080 Ti @ 1080p. Maybe some NV DX12 driver hiccups?
Click to expand...
Possibly, however:
The ranking in Forza 7 is very unusual. Nvidia has confirmed ComputerBase, however, that the results are so correct, so there is no problem with the system in the editorial regarding GeForce.
Click to expand...
Forza 7 was developed by Turn 10 Studios and is based on the in-house ForzaTech engine. The game uses only the DirectX 12 API, but only the feature level 11_0 and, like all other DirectX 12 games so far, no new hardware features of the API. Whether Async Compute is used, is unknown, but appears in view of the test results as probable.
Click to expand...
 
CaptainGinger said:
Looks like the Nvidia cards are underperforming rather than any Vega magic, the RX 580 has a higher 99 percentile result than the 1080 Ti @ 1080p. Maybe some NV DX12 driver hiccups?
Click to expand...

Those frametime numbers can't be right because all the higher performing cards are getting much higher frametimes than the low-end cards. I think that's just minimum framerate, not frametime.

Regardless, the comparison to the nvidia cards don't explain how the Vega 64 is getting 60-65% more performance than the Fury X, despite clocking only ~45% higher and having lower theoretical bandwidth. At 4K the Fury X tanks hard probably because it gets short on VRAM.
Up until now, best case scenario for the Vega 64 in games has been a Fury X + 40% (which corresponds to the higher clocks). Something is indeed happening here.
 
ToTTenTranz said:
Those frametime numbers can't be right because all the higher performing cards are getting much higher frametimes than the low-end cards. I think that's just minimum framerate, not frametime.

Regardless, the comparison to the nvidia cards don't explain how the Vega 64 is getting 60-65% more performance than the Fury X, despite clocking only ~45% higher and having lower theoretical bandwidth. At 4K the Fury X tanks hard probably because it gets short on VRAM.
Up until now, best case scenario for the Vega 64 in games has been a Fury X + 40% (which corresponds to the higher clocks). Something is indeed happening here.
Click to expand...

Maybe it's 99th percentile frametime converted to FPS? I guess we'll have to wait and see if the devs comment on why Vega seems to be doing so well.

Agreed about the Fury X 4K numbers probably tanking due to limited framebuffer.
 
CaptainGinger said:
Looks like the Nvidia cards are underperforming rather than any Vega magic, the RX 580 has a higher 99 percentile result than the 1080 Ti @ 1080p. Maybe some NV DX12 driver hiccups?
Click to expand...
Also 1080Ti is only 8% faster than 1080 @1080p and 1440p, GTX 1070 only 15% faster than 1060 ..etc. Even the RX 580 is outputting better min fps than 1080Ti. I also noticed a GPU underutilization of about 80% on my 1070 even @4K Ultra, so something is definitely holding the fps back on NVIDIA's GPUs in Forza 7's demo, since this doesn't happen in Forza 6 or Forza Horizon 3.


I believe this discussion should be transferred to the DX12 thread though?
 
Last edited:
Forza 7 was developed by Turn 10 Studios and is based on the in-house ForzaTech engine. The game uses only the DirectX 12 API, but only the feature level 11_0 and, like all other DirectX 12 games so far, no new hardware features of the API. Whether Async Compute is used, is unknown, but appears in view of the test results as probable.
Click to expand...

So what do we call this, DirectX 11.5?
 
Look at that, some performance numbers with varying levels of MSAA. Pity it's only at one resolution and only for two cards: 1080 FE and Vega 64.

Vega MSAA hit is certainly looking worse.
 
Grall said:
Makes one wonder, where in the GPU would this hotspot be located? What is it doing that's causing such massive local heat buildup...?
Click to expand...

I'm speculating, but with the more recent chips AMD's indicated there are dozens of power/temp monitors throughout the chip.
There may not be any single area every time, but a chip readout of the hottest measured/calculated sector of the chip.

Silicon's thermal conductivity is not that great, which limits how much heat can be conducted horizontally through the chip. There is some marginal benefit, as seen when whole CPU cores are gated off and Ryzen's turbo gives a few more watts to a neighboring core, but that's less likely with a fully active GPU.

Chips can already take local global dissipation needs beyond spec, and given the non-linear behavior and areas known to be low-demand it makes local effects even more challenging. Without more highly-reactive power/temp sensors physically distributed throughout, the global clock and power guard bands need to be wider to account for how long it takes before the chip knows that a given area has spiked.
I'm curious if the hot-spot is a physical measurement from a local thermal diode near some active silicon, or if it's possibly a mixture of temperature and calculated power/temp values based on physical characterization and small proxy units that perform placeholder calculations and serve as a conservative representation of what the actual units are doing.

Pushing the clock/voltage envelope enough that the chip generally strives to stay in steeper regions of the power curve can potentially make things worse in scenarios where much of the chip is less active but the most active regions are drawing enough power to overwhelm their ability to dissipate.
It also raises a question as to whether a distinction needs to be made between a reactive DVFS implementation and a twitchy one.

Given some of the counter-intuitive behaviors with clock and voltage adjustments for getting performance, it may be that Vega is driving itself into regions of low ROI in terms of power/temp, and then needing to throttle back until power delivery or thermal dissipation can bring things back below the red line, at which point it might quickly ramp into the counterproductive range.
AMD's fast-reacting calculated DVFS values are also dependent on the physical characterization of the silicon, which various reviews/tweaks show might not be that precise and has been a theme for multiple generations of first-run AMD products.
 
DavidGraham said:
Also 1080Ti is only 8% faster than 1080 @1080p and 1440p, GTX 1070 only 15% faster than 1060 ..etc. Even the RX 580 is outputting better min fps than 1080Ti. I also noticed a GPU underutilization of about 80% on my 1070 even @4K Ultra, so something is definitely holding the fps back on NVIDIA's GPUs in Forza 7's demo, since this doesn't happen in Forza 6 or Forza Horizon 3.
Click to expand...
Sounds like driver overhead.
 
eastmen said:
Got my vega 56 for $360 and installed it tonight. I've been playing the SW battlefront 2 demo and my 290 couldn't play in 1080p ultra. The 56 vega does so with ease . Its also running cooler dispite having just the standard cooler on it while I had put an corsair h60 in a push pull config on the 290
Click to expand...

Fantastic price $360. Can I ask where did you get the Vega 56 for $360.

Today the current prices are $470 or higher.
http://www.nowinstock.net/computers/videocards/amd/rxvega56
 
You must log in or register to reply here.

Similar threads

D
AMD RX 7900XTX and RX 7900XT Reviews
17 18 19
Replies
374
Views
57K
DegustatoR
D
D
Power consumption for GPUs
2 3
Replies
53
Views
8K
Man from Atlantis
Man from Atlantis
Shortbread
Phil Spencer Interview: Redfall Reviews, Activision Deal - Kinda Funny Xcast Ep. 137
7 8 9
Replies
175
Views
12K
trinibwoy
T
Kaotik
AMD Radeon Software Driver Announcement thread
2
Replies
29
Views
13K
DegustatoR
D
Rys
Reopening Architecture and Products + rules and guidelines changes
3 4 5
Replies
90
Views
16K
Shifty Geezer
Shifty Geezer
Back
Top