AMD GPU Generational Performance Part 1 @ [H]

Vega struck me, since its release, as a stretch for AMD. They really had to milk it to get 1080-level performance out of it. Where Ryzen exceeded many performance targets, Vega only just barely made it.

Seeing these benchmarks, Vega isn't bad. But it's not what AMD really needed, either. It's somewhere in between. Neither an utter failure or an unqualified success. Still, this is better than no high end presence at all.

I hope the next GPU out of AMD brings us more competition. Nvidia is killing it from a performance angle, but their pricing and anti-competitive practices are just screaming for real competition again.

And, lol, my old 7970 still continues to serve in my HTPC, and dual 7970s still run my old arcade/console emulator build. GCN forever, lolol.
I think AMD's divorce from GloFo is really going to help them. I am very anxious to see what they are going to be able to do on TSMC's 7nm.
 
I think AMD's divorce from GloFo is really going to help them. I am very anxious to see what they are going to be able to do on TSMC's 7nm.
To think that at some point in seemed GloFo was going to beat TSMC at 7nm
 
well at least you could buy 2 out of the 4 cards...

Calling amd generations is a joke, they just change the powersupply and ram, and maaaaaybe tweak the die a little, can call it a rev. They have had the same consumer core for what, 6-8 years, is the 580 much different vs what i used to mine litecoins back in the day?
 
Thanks for the article, great work with lots of useful information. It is appreciated.
I found it really interesting to use both series of articles to compare AMD with Nvidia cards that released at the same time (in same price bracket) and see how performance stacks today.

Small nitpick: Page 13 - Second block should read "High" instead of "Highest Game Settings"
 
well at least you could buy 2 out of the 4 cards...

Calling amd generations is a joke, they just change the powersupply and ram, and maaaaaybe tweak the die a little, can call it a rev. They have had the same consumer core for what, 6-8 years, is the 580 much different vs what i used to mine litecoins back in the day?


They say ignorance is bliss...

If you were to invest the time into reading about the advancements over the years on both sides youd see all the changes that were made. CUDA has been around since 2007? GCN is on its 5th revision and has as evidenced here, provided large gains in a 4 year span. Intel is still working off its Core uArch as well starting with Sandybridge...So many things build in an evolutionary way. Zen was a fresh start for AMD, but im sure they used some concepts learned from previous designs as well. I suggest you read up on some of the architectural deepdives served up around the web and educate yourself on the various iterations of GPU's from both camps.
 
great article..glad you guys didn't include all the rebrands we had to deal with from AMD ..I have Hope for AMD to someday again . ..well make a new ATI 9700 \9800 Pro type card that rules in all games ..back in the day that gen was a huge leap that actually held competition ..then the rebrands started coming .. but at the same time our games started changing , my feeling is... that gen pushed the limits ..and yes i think AMD can do it again

gotta have hope right ! ?
 
Nice article.

It's a shame to see AMD graphics division get such little funding. Vega is ok, but outside of its compute performance it doesn't bring much competition. Navi is going to be the next Polaris, great value but not aimed at high end. Wonder when we will see another high end card from AMD. Maybe some of that Ryzen cpu sale money can help put the graphics division back on track. Soon as the 20xx series launch from nvidia, those used 1080ti prices gonna be great, just sayin.
 
Was comparing the AMD report 1 with the Nvidia report 3, so as to look at the performances and make comparisons... noticed that on the Witcher 3, different settings were used. The AMD run has hairworks on, the nVidia had it off. We can't make apples to apples comparisons if different settings were used. The AMD Report does have a second run with Hairworks off, but it looks like other settings were different from the Nvidia runs.

Can you add 1 more graph to the AMD report with settings that match those used in the nvidia reports? (Or add 1 more graph to the nVidia report with hairworks on at settings that match the AMD report?)

Thanks! :)
 
Last edited:
Nice review as usual, very interesting.



although i understand the context of the review is windows gaming but just a very minor niggle.. Vulkan is available on windows too mentioning alongside DX12 would be appreciated at the end of the article. especially seen as you benchmark a Vulkan title.
 
Last edited:
Nice review as usual, very interesting.



although i understand the context of the review is windows gaming but just a very minor niggle.. Vulkan is available on windows too mentioning alongside DX12 would be appreciated at the end of the article. especially seen as you benchmark a Vulkan title.
We run the games on the API that we find those to work best on.
 
Great article! With high or medium settings 290(X) or 390(X) series are able to play anything on 1440P even newest games without problems and that is good news for gamers without big money in their pockets as RX470 is on the same league of performance and it has returned to normal price nowadays. Moreover, CGN is the longest arch ever for GPUs and it has helped the driver team to optimise all those GPUs from 2012 and on much easier.
 
Go ahead and ban me for discussing something your article discusses — features like Rapid Packed Math being implemented by developers and how that impacts gaming.

GameWorks is very similar to Rapid Packed Math. It is one company's tech vs another's (Rapid Packed Math). Clearly, Nvidia has been more successful than AMD at getting such game performance enhancement tech (for their cards) implemented by devs.

If you can prove what I wrote isn't on-topic, as it pertains directly to the quoted text from your article, then I will retract my posts.

You are comparing apples to oranges. There is nothing wrong with gameworks as long as it doesn't slow down AMD hardware which it has done in the past. Rapid Pack Math is just a feature on AMD GPUs, not sure what one has to do with the other. Gameworks is a middleware software suite.

So no this thread isn't about discussing Game works. One gives performance boost the other gives you in game effects, like hairworks and stuff. Like I said apples and oranges.
 
While complicated by the turbo behaviour of more recent cards, it would be really interesting to see a clock-for-clock comparison across the various GCN versions.
 
While complicated by the turbo behaviour of more recent cards, it would be really interesting to see a clock-for-clock comparison across the various GCN versions.

I'd like that too on both sides. IIRC Maxwell is faster than pascal clock for clock.
 
While complicated by the turbo behaviour of more recent cards, it would be really interesting to see a clock-for-clock comparison across the various GCN versions.
seems like they did this about a year ago, but was only comparing 2 generations
 
These comparison graphs really remind me of my prior concern that one of the greatest issues holding back Fury and especially Vega has been the limited memory bandwidth. Although anecdotal due to my small sample size, with all the prior generations of ATI/AMD (and my one not too, too ancient Nvidia) graphics cards I've overclocked when core clock and memory bandwidth increased by the same percentage then frame rate scaled linearly. However increasing core count or core speed without commensurate memory bandwidth would eventually lead to rapidly diminishing returns.

The 290/390 transition to Fury had a more than 40% increase in cores with only a 1/3 increase in memory bandwidth along with a real problem in only having 4GB of memory. Remember how before Vega came out many of us thought there would be a 40-50% plus improvement in frame rate based on the core speed increase and possible architecture improvements. And yet with Vega there's a roughly 40-50% general increase in core clock speed (depending on heat soak) with a REDUCTION in memory bandwidth leading to a mere 20-30ish percent overall gain.


Clock a 56 (or a rare golden 64) to 750/1100 and keep them cool (as mide do at 35C under load) and I can show you north of a 50% improvement over Fury, with lower power usage then a stock clocked air cooled Vega.

AMD really shit the bed with rushing the cards out and shipping them with an insane default Voltage and that low HBM clock. The cards fly when they are properly tuned and cooled. They do love their memory bandwidth increases, as 945 to 1100 is insane, even with a fixed core speed. Ramp that core to 1750 and you have ~1080 OC+ results.

If AMD masked and respun VEGA on the 12nm node and tuned their voltage curves then they could offer 1080+ performance on the CLC /beefy air cooled models. I'd love for them to do that with a bump to 16GB of ram (wishful thinking but my Mining would love the boost).
 
I wish you would test on pubg as it's one of the most unoptimized games making it very demanding.
 
If you bought a 290x at launch jesus christ the VALUE. My god.

Great article too btw!
 
If you bought a 290x at launch jesus christ the VALUE. My god.

Great article too btw!

They were a great value. I bought a 290 at launch and slapped a full cover block on it. At 1.3Ghz/1.4Ghz mem it was a beast. I later added a second for Xfire and used them up until VEGA's launch.
 
I only play what interest me which I like big worlds and FPS combined and as Hawaii was made for Eyefinity as 4K really didn't exist much pass 30Hz back then , Hell I play 4K 60Hz with a Ryzen 1400 and 16Gb of memory via club 3D dp to HDMI 2.0 with a dam usb wifi stick just to show u AMD got game.



Sorry my phone lens had dust on it but that is the HHgregg 42" um Seiki 4K 60Hz that cost $249 display model as it has been firmware upgraded.

 
I'm kind of disappointed you didn't include gcn 1.0 card (6970)

Mainly because 290x reference designs were so horrendous you were looking at a 10% performance increase only from 6970.

So add 10% over another 2 years and you'll really see the disappointment in the entire gcn lineup.
 
I'm kind of disappointed you didn't include gcn 1.0 card (6970)

Mainly because 290x reference designs were so horrendous you were looking at a 10% performance increase only from 6970.

So add 10% over another 2 years and you'll really see the disappointment in the entire gcn lineup.

you should re-check your knowledge base.. 6970 wasn't GCN, the first GCN GPU was the HD7970. which was by itself a huge jump over cayman's HD 6970 in the order of the 45%.. nowadays it should be in the 60% - 70% with the overhead fixes and driver performance jumps.. and the 290X over the HD 7970 should be another 30% - 40% extra performance, so I don't really know what are you speaking off... HD 7970 aka R9 280X (rebranded) and R9 290X are probably the two best GPU ever made with a LARGE longevity.
 
Mainly because 290x reference designs were so horrendous you were looking at a 10% performance increase only from 6970.

I actually went from a 6970 to a 290 (non-X) and the difference was huge (probably 50-100%).
 
These comparison graphs really remind me of my prior concern that one of the greatest issues holding back Fury and especially Vega has been the limited memory bandwidth. Although anecdotal due to my small sample size, with all the prior generations of ATI/AMD (and my one not too, too ancient Nvidia) graphics cards I've overclocked when core clock and memory bandwidth increased by the same percentage then frame rate scaled linearly. However increasing core count or core speed without commensurate memory bandwidth would eventually lead to rapidly diminishing returns.

The 290/390 transition to Fury had a more than 40% increase in cores with only a 1/3 increase in memory bandwidth along with a real problem in only having 4GB of memory. Remember how before Vega came out many of us thought there would be a 40-50% plus improvement in frame rate based on the core speed increase and possible architecture improvements. And yet with Vega there's a roughly 40-50% general increase in core clock speed (depending on heat soak) with a REDUCTION in memory bandwidth leading to a mere 20-30ish percent overall gain.

The reason why Vega has less memory bandwidth is AMD improved their DCC algorithms from Fiji. Hawaii didn't even have DCC. Though you can always bump memory clock back up to 1000MHz to get the same 512GB/s as Fiji. Also because SK Hynix dropped the ball on 2Gbps HBM2, forcing AMD to use OC'd Samsung 1.6Gbps chips running at 1.9Gbps/1.356v.

The biggest issue is that Vega is an upclocked Fiji with some new tricks. GP102 has 6 GPCs with 6 raster engines and high ROP count (88-96). Vega is held back by its 4 shader engine design with 4 rasterizer+geometry engines and 64 ROPs. An easy 30% increase in rendering performance would come from simply adding 2 more shader engines and rebalancing architecture; AMD opted not to do that as GCN has physical limitations (4 shader engines). Vega could have matched or beat the 1080Ti with 6 shader engines, 4608 shaders, 6 raster+geometry, and 96 ROPs.

Nvidia is very aggressive with DCC, as those bandwidth savings can be considered as extra memory bandwidth. Same for AMD.

484GB/s is equivalent to the 1080Ti.
 
read it, liked it, kinda disappointed in AMD performance...but given all of the factors involved, it is expected. can't wait for round two...
 
I think AMD's FineWine thing is also biting them back, cause the 290X is way ahead of its competitor, the GTX 780 Ti and the GTX 980 Ti aged better than Kepler. AMD's fault of its own success regarding performance scalability between their own GPUs.
 
Back
Top