AMD Vega Hardware Reviews

Truthfully, if I was filthy rich this would be the card I'd want. Or a couple of those even. How cool would it be to virtually end all loading times in games and streaming hiccups in free-roaming games?

For be honest, i more interested on the Vega 16GB FE or 8 GB, but not for gaming, but raytracing.. 2 of them.. ( 26Tflops of FP32 )...
 
Does the SSG-storage actually show as HDD? At least they have some sort of API for it, currently supported (in beta) by Adobe Premiere & After Effects only
Probably, but AMD was using it as a block device so there wouldn't be a filesystem to mount. Probably want to tweak the page size for HBCC with all that, but I'd imagine they fill it like a standard graphics/compute resource.

For be honest, i more interested on the Vega 16GB FE or 8 GB, but not for gaming, but raytracing.. 2 of them.. ( 26Tflops of FP32 )...
With Threadripper you can probably get away with 1-2 more. Guess it depends on your dataset.
 
Probably, but AMD was using it as a block device so there wouldn't be a filesystem to mount. Probably want to tweak the page size for HBCC with all that, but I'd imagine they fill it like a standard graphics/compute resource.


With Threadripper you can probably get away with 1-2 more. Guess it depends on your dataset.

Well it is more a question of money ( as it is for my home system ).
 
http://www.tweaktown.com/news/58635...-leaked-benchmarks-gtx-1070-killer/index.html

Code:
Radeon RX Vega 56 benchmark results:
Battlefield 1: 95.4FPS (GTX 1070: 72.2FPS)
Civilization 6: 85.1FPS (GTX 1070: 72.2FPS)
DOOM: 101.2FPS (GTX 1070: 84.6FPS)
COD:IW: 99.9FPS (GTX 1070: 92.1FPS)

Classify this as a rumor for now.

Now if only power consumption is no more than the 210w tdp... Of course, a 1070 can likely be overclocked 20% and have a similar performance and power consumption. I do wonder if that is a reference 1070 or aftermarket overclocked one. I guess we will find out how Vega does soon enough, but Vega 56 could interest me if power consumption is in check and I can get one easily with a cooler I am happy with.
 
Now if only power consumption is no more than the 210w tdp... Of course, a 1070 can likely be overclocked 20% and have a similar performance and power consumption. I do wonder if that is a reference 1070 or aftermarket overclocked one
You have to keep in mind that Vega 56 is 14% behind Vega 64 in resources, which -assuming perfect scaling- means it's 10~15% behind Vega 64 in performance. As such it can easily be OC'ed to reach Vega 64 level, just like Fury to FuryX before, this makes it a very attractive option, considering the price.

EDIT:
Reviewing the numbers they are probably min fps, Doom numbers are on the low side for a 1070, considering I get 90fps+ @1440p with a 3770K, BF1 @1440p gives me 85+fps. Nothing in my system is OC'ed.
 
Last edited:
You have to keep in mind that Vega 56 is 14% behind Vega 64 in resources, which -assuming perfect scaling- means it's 10~15% behind Vega 64 in performance. As such it can easily be OC'ed to reach Vega 64 level, just like Fury to FuryX before, this makes it a very attractive option, considering the price.
Vega 56 also has clock disadvantage over Vega 64. Total FLOP disadvantage is 20.5% and memory bandwidth disadvantage is 18.0% against the cheaper Vega 64 model. But clock difference between Vega 56 and 64 is only 5%(boost)-7%(base). If the bottleneck is geometry based or ROP based, Vega 56 might actually get quite close to Vega 64. And we don't know how thermals play out in this comparison. Vega 56 has lower clocks and some disabled CUs. This might actually help it.
 
https://www.techpowerup.com/reviews/AMD/Radeon_RX_Vega_Preview/5.html

TPU posted full slide deck. Couple of interesting notes:
▪HBCC increased FPS in Heaven by 7% without aritifical VRAM limit.

▪Calculating from TDP's, being 295W for balanced and 210W for power saving, we get in BF1 4k:
295W for 75.8FPS~
210W for 72.6FPS~

This is quite extreme power scaling.

You have to keep in mind that Vega 56 is 14% behind Vega 64 in resources, which -assuming perfect scaling- means it's 10~15% behind Vega 64 in performance. As such it can easily be OC'ed to reach Vega 64 level, just like Fury to FuryX before, this makes it a very attractive option, considering the price.

EDIT:
Reviewing the numbers they are probably min fps, Doom numbers are on the low side for a 1070, considering I get 90fps+ @1440p with a 3770K, BF1 @1440p gives me 85+fps. Nothing in my system is OC'ed.
DOOM is highly variable depending on where you test it.

https://www.techpowerup.com/reviews/Zotac/GeForce_GTX_1080_Mini/12.html

A 1070 gets even lower than those numbers in TPU's testing.
 
This is quite extreme power scaling.

Yup. I guess the significantly lower power draw from HBM2 (and probably some aggressively low core clocks/vcore states) allows for large power savings.

jtppE7j.jpg


The endnote slide
claims it's all using the highest settings in every game, but it's all on 1080p though. It might still make Vega the most power efficient card for older and less demanding games.


▪HBCC increased FPS in Heaven by 7% without aritifical VRAM limit.
Another interesting fact is that you can apparently decide in the driver how much system RAM HBCC can take over. I wonder if this will make the game recognize the GPU's total VRAM as HBM2+system-alocated RAM.
If the HBCC is successful at seamlessly prioritizing what needs to be in the HBM2 and what can be streamed through PCIe, this could indeed be a game changer for lower-end GPUs with single HBM2 stacks (i.e. Vega 11 wouldn't really hurt by having only 4GB, for example).
 
Yup. I guess the significantly lower power draw from HBM2 (and probably some aggressively low core clocks/vcore states) allows for large power savings.

Another interesting fact is that you can apparently decide in the driver how much system RAM HBCC can take over. I wonder if this will make the game recognize the GPU's total VRAM as HBM2+system-alocated RAM.
If the HBCC is successful at seamlessly prioritizing what needs to be in the HBM2 and what can be streamed through PCIe, this could indeed be a game changer for lower-end GPUs with single HBM2 stacks (i.e. Vega 11 wouldn't really hurt by having only 4GB, for example).

T̶u̶r̶b̶o̶C̶a̶c̶h̶e̶ ̶(̶c̶a̶n̶t̶ ̶r̶e̶m̶e̶m̶b̶e̶r̶ ̶w̶h̶a̶t̶ ̶t̶h̶e̶ ̶A̶M̶D̶ ̶s̶o̶l̶u̶t̶i̶o̶n̶ ̶w̶a̶s̶ ̶c̶a̶l̶l̶e̶d̶)̶ ̶ HyperMemory says hello to his grandson!!!
 
Last edited:
Another very interesting tidbit is the "Power Save Mode" slide, showing ~20-40% savings in a number of high-profile games:

sBYaRiN.jpg


At first I thought this was a bit of a rehash of the Radeon Chill slide, but turns out these results are taken at 4K and highest settings.
And the actual performance difference between turbo and power saving modes is 2 and 3%.

V3lPq1u.jpg




T̶u̶r̶b̶o̶C̶a̶c̶h̶e̶ ̶(̶c̶a̶n̶t̶ ̶r̶e̶m̶e̶m̶b̶e̶r̶ ̶w̶h̶a̶t̶ ̶t̶h̶e̶ ̶A̶M̶D̶ ̶s̶o̶l̶u̶t̶i̶o̶n̶ ̶w̶a̶s̶ ̶c̶a̶l̶l̶e̶d̶)̶ ̶ HyperMemory says hello to his grandson!!!
And AGP Texturing from the late 90's says hello to those teenagers.
 
Last edited by a moderator:
And AGP Texturing from the late 90's says hello to those teenagers.

Was it actually used though? I seem to remember some tweaking guides where that was one of the AGP features recommended to be turned off. With PC's having such low amounts of memory at time, seemed quite counter productive.
 
Only Intel's i740 had proper functioning implementation of AGP texturing, to cut costs from the on-board memory. Since a lot of the user base at the time was using dodgy AGP chipsets with even dodgier driver support, most of the graphics IHVs didn't bother too much with the enhanced features of AGP. 3Dfx was a prime example of that and by the time they folded, the DRAM prices allowed for larger amounts of video memory, that negated further the usability of AGP texturing.
 
HBCC increased FPS in Heaven by 7% without aritifical VRAM limit.
That's rather interesting as I'd expect HBCC to only help when constrained by memory. On an 8GB card with a game needing half of that you would think it would preload all the resources. This would seem to indicate that from the onset it's paging only required data and increasing minimum framerates (decreasing stall time) as only a fraction of the original data would need moved.

I wonder if this will make the game recognize the GPU's total VRAM as HBM2+system-alocated RAM.
HBM should be acting like cache so primarily non-addressable. In that case the game would see whatever value of system memory is assigned. That said there are likely different pools making it a bit more complex.
 
That's rather interesting as I'd expect HBCC to only help when constrained by memory. On an 8GB card with a game needing half of that you would think it would preload all the resources. This would seem to indicate that from the onset it's paging only required data and increasing minimum framerates (decreasing stall time) as only a fraction of the original data would need moved.


HBM should be acting like cache so primarily non-addressable. In that case the game would see whatever value of system memory is assigned. That said there are likely different pools making it a bit more complex.
It is very interesting. Max FPS shouldn't go up under any circumstance I can think of, but frametime consistency could improve.
 
It is very interesting. Max FPS shouldn't go up under any circumstance I can think of, but frametime consistency could improve.
They would if constrained by memory, but I wouldn't think that is the case on this benchmark.

Another likely possibility is that the HBCC absorbed the resource management work from the CPU. Lower CPU load leading to higher max FPS.
 
Back
Top