Radeon Vega Architecture Thread Of Shame

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.
Status
Not open for further replies.

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
Guys, I know this is a bit stretched out comparison, however. The Vega GPU averaged well over 60 FPS in 4K doom right? We can say that it more likely averaged around 70 FPS in this resolution.

I wanted to check how it compares to GTX 1080 Zotac AMP! Extreme version. The GPU that draws 222W on average in games:
power_average.png

What was the score?
doom_3840_2160.png


58 FPS on average in Doom 4K Vulkan, on a platform with 6700K OC'ed to 4.5 GHz.
AMD demo platform was using Ryzen(?) CPU.
 

IllogicalGlory

Senior member
Mar 8, 2013
934
346
136
58 FPS on average in Doom 4K Vulkan, on a platform with 6700K OC'ed to 4.5 GHz.
AMD demo platform was using Ryzen(?) CPU.
You can only really compare benchmarks done in the same area. Same principle as was being pointed out RE: the Battlefront Vega demo. We don't know where TPU does their bench run.
 

PhonakV30

Senior member
Oct 26, 2009
987
378
136
Yes.It was AMD Ryzen.Just look at MB.It's Red and I can see "Golemit".it's engineering sample.
 

Glo.

Diamond Member
Apr 25, 2015
5,661
4,419
136
You can only really compare benchmarks done in the same area. Same principle as was being pointed out RE: the Battlefront Vega demo. We don't know where TPU does their bench run.
It does not have to be TPU even. Most reviews post GTX 1080 Doom Vulkan 4K performance in 50 FPS range(average), for founders edition, which is in line with TPU number. I have just checked that. TomsHardware review for example posts for Titan X around 77 FPS in the game and 56 FPS for 1080 Founders Edition, whereas (I think) GamersNexus posts 80 FPS average for the same game and same GPU(TitanX).

So at least there is some logic in the numbers we see.
 

JDG1980

Golden Member
Jul 18, 2013
1,663
570
136
What's everyone's opinion? I am going to start with 45-50% faster than Fury X at 4K for $599-649 as I noted above. Would this be disappointing to you guys?

That would be an extremely disappointing showing. And if the performance is that bad, it won't sell at that price.

Fury X only has about 70% of the gaming performance of stock GTX 1080. (Source) If Vega 10 was only 50% better than that, it would be a mere 5% better than GTX 1080. (0.70 x 1.50 = 1.05)

GTX 1080 has a 314mm^2 chip, uses GDDR5X, and has a TDP of ~185W. (Forget about the AIB cards, we're comparing stock to stock.) If AMD needs a 450-500mm^2 chip with expensive HBM2 and a 225W TDP to simply equal GP104, that's an EPIC fail. They would have to price it below $499, and at that point they'll be losing money. (And if Nvidia cuts the GTX 1080 to $499 like they did with the 980 before it, then AMD will still not get the sales.)

Vega 10 HAS to match GP102 or it's not worth releasing as a consumer gaming part.
 
  • Like
Reactions: xpea

HurleyBird

Platinum Member
Apr 22, 2003
2,670
1,250
136
I can't believe no one has figured out which area the demo was running at and benchmarked a 1080 or Titan XP in the same location yet. I'm sure some forum user eventually will do so, but how not a single tech site can be bothered to pluck this low hanging fruit is beyond me.
 

sirmo

Golden Member
Oct 10, 2011
1,012
384
136
That would be an extremely disappointing showing. And if the performance is that bad, it won't sell at that price.

Fury X only has about 70% of the gaming performance of stock GTX 1080. (Source) If Vega 10 was only 50% better than that, it would be a mere 5% better than GTX 1080. (0.70 x 1.50 = 1.05)

GTX 1080 has a 314mm^2 chip, uses GDDR5X, and has a TDP of ~185W. (Forget about the AIB cards, we're comparing stock to stock.) If AMD needs a 450-500mm^2 chip with expensive HBM2 and a 225W TDP to simply equal GP104, that's an EPIC fail. They would have to price it below $499, and at that point they'll be losing money. (And if Nvidia cuts the GTX 1080 to $499 like they did with the 980 before it, then AMD will still not get the sales.)

Vega 10 HAS to match GP102 or it's not worth releasing as a consumer gaming part.
1080 wasn't more than 50% faster than 980ti, was it? I don't think it would be extremely disappointing personally. I don't know why AMD has to be held to a different standard. It wouldn't be ground breaking, but it would still be a solid improvement, and some form of a competition at the high end.
 

sirmo

Golden Member
Oct 10, 2011
1,012
384
136
According to the latest December Steam HW survey, it is 4:1.
It is possible, although I am not sure Steam is reporting numbers correctly. For instance I never remember being asked to do a hardware survey but like once 2 GPUs ago. I had to trigger it manually. Does anyone know what the criteria is for Steam asking you to do a survey?

If it's only on brand new installs, when you first install it, then it might be skewed in favor of new systems. And we know Nvidia has a big lead in laptops so that might be skewing the numbers compared to folks just upgrading their GPUs. That's at least one of my theories. Because marketshare reports don't quite match the Steam numbers.
 

Thala

Golden Member
Nov 12, 2014
1,355
653
136
Partially because Polaris's power draw is such that scaling it to competitive performance levels would break the bank in terms of the power budget for a big consumer GPU.

I disagree with this statement. If we take RX480 and GTX1060 as reference, Pascal has a 20% perf/watt advantage over Polaris. This is not a showstopper for a big consumer GPU so why would it break the power budget?
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Bigger chips/cards always have better perf/watt ratio than smaller ones :)
Always? :) In TPU's charts at launch the 7870 had 28% better Perf/W than a 7970 at 1920x1200 and 22% better at 2560x1600. The 260X was basically tied with the 290X at 2560x1600 and was 15% better at 1080p. The GTX 980 was 9% more efficient than Titan X at 4k and 14% at 1080p. Big chip Titan XP finally got a win and was 6% ahead of GTX1080 at 4k, but GTX1080 is 5% ahead at 1080p.

Not saying the bigger chips aren't sometimes more efficient, but that is quite often not the case. The biggest chip having a 30% lead over the next in line is pretty rare, at least since the 28nm transition happened.
 

Bacon1

Diamond Member
Feb 14, 2016
3,430
1,018
91
Always? :) In TPU's charts at launch the 7870 had 28% better Perf/W than a 7970 at 1920x1200 and 22% better at 2560x1600. The 260X was basically tied with the 290X at 2560x1600 and was 15% better at 1080p. The GTX 980 was 9% more efficient than Titan X at 4k and 14% at 1080p. Big chip Titan XP finally got a win and was 6% ahead of GTX1080 at 4k, but GTX1080 is 5% ahead at 1080p.

Not saying the bigger chips aren't sometimes more efficient, but that is quite often not the case. The biggest chip having a 30% lead over the next in line is pretty rare, at least since the 28nm transition happened.

Well in general yes, but it all depends on the clocks as well.

Bigger chips = more "cores" which you can run at lower speeds = lower power draw while having more performance. Now if you instead clock them all really high you end up with a huge power draw. OC'd 980 Tis end up using more power than Fury X for example, while stock 980 ti uses much less.
 
  • Like
Reactions: prtskg

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Well in general yes, but it all depends on the clocks as well.

Bigger chips = more "cores" which you can run at lower speeds = lower power draw while having more performance. Now if you instead clock them all really high you end up with a huge power draw. OC'd 980 Tis end up using more power than Fury X for example, while stock 980 ti uses much less.
I wouldn't necessarily compare a Fury X to a 980 Ti (especially on this sub), but clocks for the pairs I chose have the smaller card being clocked higher but are still more efficient. For GCN1, the 7870 was clocked faster at 1GHz vs 925MHz for the 7970 and was still more efficient. At stock the 260X is clocked at 1.1GHz vs 1GHz for the 290X. GM204 and GP104 were both clocked higher than their corresponding big chips as well, and still were just as if not more efficient.

The only case where the bigger slower clocked chip was more efficient was Fury X.
 

Bacon1

Diamond Member
Feb 14, 2016
3,430
1,018
91
The only case where the bigger slower clocked chip was more efficient was Fury X.

Huh? You can't compare cross-architectures clocks vs power efficiency. There is a great graphic that shows as clock speed goes up the power reqs skyrocket. Thats why a 20% OC will end up using 40%+ more power. Look at Nano vs X. Both are exact same chip except the clocks are lower and the Nano has a huge reduction in power because of it.
 

MrTeal

Diamond Member
Dec 7, 2003
3,554
1,658
136
Huh? You can't compare cross-architectures clocks vs power efficiency. There is a great graphic that shows as clock speed goes up the power reqs skyrocket. Thats why a 20% OC will end up using 40%+ more power. Look at Nano vs X. Both are exact same chip except the clocks are lower and the Nano has a huge reduction in power because of it.
Of course, no one is disputing that. The pairs of chips I mentioned in my post were all the same generation; GCN1, GCN2, GCN3, Maxwell and Pascal. You'd made the point that bigger chips could be clocked lower and thus have better efficiency. I was just pointing out that within each of those generations, in addition to being more efficient the smaller chip was actually clocked faster than the larger chip. In most of those cases, if you clocked the two chips in the same generation at the same speed you'd either need to OC the big chip or underclock the small one, and the perf/w lead of the smaller chip would increase.

Either way, this is getting pretty off topic for something that was relatively off-the-cuff and meant to show that Fiji had extremely good perf/w relative to Tonga, and it's not improbable a large part of that is HBM.
 
  • Like
Reactions: crisium

raghu78

Diamond Member
Aug 23, 2012
4,093
1,475
136
Wow that's a really nice comparison. I was wondering when someone was going to do this. No wonder polaris was such a disappointment. It was basically just a die shrink with only minor improvements, most probably coming from the increased buffer sizes and L2 cache. AMD has been putting out these tiny incremental improvements for going on 6 years now. Sooner or later they need to make a significant leap in performance. I hope that's what Vega is. But if that's what Vega is, then why is it so big? Why wouldnt they make such a radical change starting with a smaller die? It really makes no sense. AMD has this knack for always making me shake my head in confusion...

AMD have Polaris 10 and Polaris 11 to cater to the USD 100 - USD 300 GPU space. What they need now is Vega to start addressing the high end segments. AMD is already very late against Pascal GP102 and GP104. It makes sense to launch the flagship Vega 10 (roughly 500 sq mm) first and then Vega 11 (My guess is 350 sq mm die size). Finally AMD can replace Polaris 10 and Polaris 11 with smaller Vega chips. I am waiting to see how big of an improvement Vega 's NCU is to the original CU. The rumoured variable length SIMDs would go a long way in improving shader utilization. Vega with 4096 sp and 2 HBM2 stacks is around 500 sq mm on 14nm FINFET. Fiji with 4096 sp was 600 sq mm on 28nm. AMD seem to have targetted their improvements on shader utilization, higher IPC and higher clocks. A dumb shrink of Fiji to 14nm would have been around 330-350 sq mm. The fact that Vega 10 is almost 50% larger with same sp count seems to hint there is major changes in the architecture. AMD have to keep in mind that Volta is due for a H1 2018 launch. So they need to get Vega out the door as quickly as possible. AMD needs Vega refreshes in H2 2018 to compete against Volta. I doubt 7nm is going to be ready atleast for anybody other than Apple in that timeframe. So AMD have to look at tweaking Vega to achieve further higher clocks, IPC and performance to compete against Volta. The earliest Navi on 7nm looks likely to launch is mid 2019
 

.vodka

Golden Member
Dec 5, 2014
1,203
1,537
136
Of course they are, and you misunderstood my post, or your sarcasm isn't needed. The beauty of the interposer + GPU + HBM paradigm debuted in Fiji is that it's a self contained small package. Provide the infrastructure required in a small, simple PCB and you've got the video card ready. The classic GPU + n memory chips like AMD's Hawaii cards or nV's GXxxx have to be supported by a huge, complex multi layered PCB for all those memory chips and their traces... this doesn't happen with Fiji, Vega, or GP100. It's elegant. Costly, but elegant.

On top of all that, Vega is rated for 225w TDP according to the leaked slides, Fiji was a 300w TDP part. A TDP limited part like the Nano would be more than feasible here, with increased performance on top.
 
Last edited:

Bacon1

Diamond Member
Feb 14, 2016
3,430
1,018
91
On top of all that, Vega is rated for 225w TDP according to the leaked slides, Fiji was a 300w TDP part. A TDP limited part like the Nano would be more than feasible here, with increased performance on top.

Yep, they can easily make one. The only question is "will someone"? Who knows how profitable the Nano was but I'm guessing that it didn't sell that well. SFF builds aren't that popular and its hard enough for AMD to sell top end cards due to mindshare. The boards will be small and we might even see a low profile version if someone thinks it will sell well.

I'm guessing we'll see the larger air cooled and another AIO or WC block but not AIO so just the hookups like this:

https://www.amazon.com/XSPC-Razor-Fury-Water-Block/dp/B0161SBVHK

Will be interesting to see how small the PCB will be though.

Just reminder for Fury X vs 980 Ti PCBS:

http://www.hardwaresecrets.com/wp-content/uploads/GTX980Ti-vs-R9FURYX-h08-1024x768.jpg

http://www.hardwaresecrets.com/wp-content/uploads/GTX980Ti-vs-R9FURYX-h16-1024x768.jpg

from http://www.hardwaresecrets.com/video-card-review-geforce-gtx-980-ti-vs-radeon-r9-fury-x/5/
 

SpaceBeer

Senior member
Apr 2, 2016
307
100
116
One thing came to my mind
http://cdn.videocardz.com/1/2017/01/AMD-VEGA-10-specifications-1000x546.jpg

If Vega for HPC has 64 CUs, and 12 TFlops compute power, it means its (boost) clock is around 1450 MHz. And we know previous HPC cards had 100-150 MHz lower clock compared to gaming GPUs. Which should mean that gaming Vega will have (boost) clock at 1500-1600 MHz?

On the other side, as discussed before, in ~530mm^2 chip, made in 14nm, you should be able to fit at least 80 CUs. I mean, P10 is 232mm^2 and has 36 CUs. As someone mentioned, it might be due to NCU changes. But 72 CUs chip (2xP10) should be within 400mm^2, probably less. This die is just too large in my opinion. When we compare Hawaii -> Fiji, increase in CUs is similar to increase in number of transistors and die. I suppose Fiji is more complicated due to HBM controller. And even die shrink of Fiji shouldn't be larger than 400mm^2. Also Tahiti -> Tonga, they have put 700 million transistors more in just a little bit larger chip (7mm^2). So does this changes really require ~30% more space? Or is there Vega with more than 64 CUs?

At this moment they can pack more cores per mm^2 than nVidia, which helps them overcome clock disadvantages. Are they taking the same way now and plan to make GPUs with smaller amount of faster cores? It's hard to compare their designs since, there are no chips with same configuration and clocks, but the difference is not that big. In that case, 64 NCU (96 ROPs?) Vega @1500-1600 MHz should really be competitor to Titan X Pascal
 
Status
Not open for further replies.