AMD Vega (FE and RX) Benchmarks [Updated Aug 10 - RX Vega 64 Unboxing]

[beginner99]

the 14LPP process is average at best

I actually disagree here. Ryzen at 2-3 Ghz is extremely efficient, better than intel in fact.

 

[OatisCampbell]

My bet is the Vega FE cards have early silicon with a bad stepping, some critical flaw that was patched in BIOS/Drivers with a large performance penalty, thus delaying the retail RX Vega cards for a new stepping to build enough quantity to ship to AIBs, whilst the older flawed stepping stock was quickly gobbled up internally by AMD to create an emergency "Prosumer" card which wouldn't be hurt as much by the performance penalty. Just a thought but didn't some industry insiders tweet AMD intentionally delayed Vega and that it was a bold yet worthwhile move? This could be it.

I said this in the rumor thread, that these might be a flawed revision and the RX model a re-spin.
Definitely one of the more mysterious launches I've ever seen, but I still think the RX will bring 1080 level performance for less money. (And that this is all it needs to do)
People need to stop thinking of AMD as an underdog fable where a ragtag group of plucky engineers will somehow save the day on 1/3 the R&D funds. Creating a GPU that is as fast as a 1080 is light years beyond what any of us could do. This isn't a "they had longer time and the same job as NV guys, it must be faster" equation.

 

[french toast]

I actually disagree here. Ryzen at 2-3 Ghz is extremely efficient, better than intel in fact.

Ryzen is great, very efficient design yes I agree there, 14nm lpp is very good at low frequency targets, goes drastically down hill past 3.5ghz and falls off of a cliff at 4ghz with ryzen CPU.
Last year nvidia launched the 1050 which was using the 14nm lpp and it got lower frequencies if my memory serves me correctly when compared to 16nm tsmc, it's unusual to see smaller GPUs of the same uarch clock lower than their bigger die cousins.

Then you you have the "3rd gen Lpp" Rx 580 debacle which is actually LESS efficient than its one year earlier Rx 480 "second gen" 14nm, it got really bad power consumption for just 5% perf improvement with a full year of revision.
That shows me that it's tapped out and has hit its limit. It is very good process at what it is designed for = low power, but falls off of a cliff past a certain frequency threshold when compared to processes that a designed for HP silicon, such as tsmc and Intel which have a more robust history of fabbing high powered chips over Samsung which has until 14nm LPU at least been designing low power processes for its mobile soc customers (including itself).
14nm LPP although not leading edge is probably quite competitive with Intel 14nm+ and tsmc 16nmFF+ at low frequency and lower power states, (perhaps favourably) but it's clearly not suitable for what AMD is using it for.

Im not flaming amd for using that process, it had no choice and 14nm lpp is pretty decent, it's just not ideal for these use cases.

EDIT; Just to add, I strongly feel if Amd wasn't tied to global foundries and could have secured 16nm FF+ we would be seeing 10-15℅ higher frequencies across all of AMDs >100w products, just my 2 pence worth.

 

[Bouowmx]

Disqus user #define is back at it again: https://disqus.com/home/discussion/...mance_benchmarks_detailed/#comment-3391224760
7126 Time Spy graphics score

Spoiler: Image

 

[SpaceBeer]

Last year nvidia launched the 1050 which was using the 14nm lpp and it got lower frequencies if my memory serves me correctly when compared to 16nm tsmc, it's unusual to see smaller GPUs of the same uarch clock lower than their bigger die cousins.

Then you you have the "3rd gen Lpp" Rx 580 debacle which is actually LESS efficient than its one year earlier Rx 480 "second gen" 14nm, it got really bad power consumption for just 5% perf improvement with a full year of revision.

EDIT; Just to add, I strongly feel if Amd wasn't tied to global foundries and could have secured 16nm FF+ we would be seeing 10-15℅ higher frequencies across all of AMDs >100w products, just my 2 pence worth.

GTX 1050 Ti can go to ~1900 MHz, while TSMC chips go to ~2100 MHz, right? So it is ~10% difference. It is important, but not that much. It means with TSMC AMD could have 10% less CUs (smaller chip size) for the same performance. If GF is 10% cheaper than TSMC, than it doesn't matter. And I believe that is the case, since AMD is probably GF's biggest customer, while TSMC would give priority to Apple and nVidia and maybe even some other companies. It might be AMD got more than 10% better price form GF than from TSMC

As I know, RX 580 is more efficient than RX 480. It can reach same clocks with lower voltage. But AMD and AIB's pushed them to the limit in order to get 5% better performance. Imagine if AMD sells full P10 at 900-950 mV with 4 GB VRAM - it would be better and more efficient than GTX 1060 3GB. Similar to Fury Nano - same perf/watt as GTX 980, even we all know Maxwell is more efficient than GCN (but at much higher price)

So when it comes to the largest chip you have (~600 mm^2), process (clocks) matters since you have no place for improvements. But in lower segments, you can alway make larger chip than competition, especially if you can get better price/mm^2 from foundry

I would assume 56 CUs RX Vega could be sold for $500, if not even less, and be at least in GTX 1080 range

 

[Glo.]

No difference in performance between Gaming and Pro drivers.

It is the same story as is with Ryzen. New architecture - software is not ready.

 

[crisium]

Reboot RTG back to ATi in Canada, please.

 

[IEC]

IIRC the earliest demos had Vega in some sort of Fiji compatibility mode with drivers (i.e. no/few Vega-specific features enabled).

It is possible that Vega FE still lacks proper support for Vega's gaming-oriented features. At least, I hope for AMD's sake that is the case. Performance close to Fiji level clock per clock would support that hypothesis, at least. To AMD's credit, they did repeatedly emphasize that gamers should skip the Frontier card.

I'm waiting until RX Vega to see where the "gamer" card lands, but AMD is going to have to pull a rabbit out of a hat if the "gaming" mode on Vega Frontier represents an "optimized" gaming experience. Because it clearly has zero optimization versus Fiji.

 

[Head1985]

Disqus user #define is back at it again: https://disqus.com/home/discussion/...mance_benchmarks_detailed/#comment-3391224760
7126 Time Spy graphics score

Spoiler: Image

Here is mine GTX1070 OC
http://www.3dmark.com/3dm/13200413

GPU score:Vega is 6% faster.

 

[Karnak]

Disqus user #define is back at it again: https://disqus.com/home/discussion/...mance_benchmarks_detailed/#comment-3391224760
7126 Time Spy graphics score

Spoiler: Image

That's roughly 2x470 CF. There has to be something wrong.

 

[french toast]

GTX 1050 Ti can go to ~1900 MHz, while TSMC chips go to ~2100 MHz, right? So it is ~10% difference. It is important, but not that much. It means with TSMC AMD could have 10% less CUs (smaller chip size) for the same performance. If GF is 10% cheaper than TSMC, than it doesn't matter. And I believe that is the case, since AMD is probably GF's biggest customer, while TSMC would give priority to Apple and nVidia and maybe even some other companies. It might be AMD got more than 10% better price form GF than from TSMC

As I know, RX 580 is more efficient than RX 480. It can reach same clocks with lower voltage. But AMD and AIB's pushed them to the limit in order to get 5% better performance. Imagine if AMD sells full P10 at 900-950 mV with 4 GB VRAM - it would be better and more efficient than GTX 1060 3GB. Similar to Fury Nano - same perf/watt as GTX 980, even we all know Maxwell is more efficient than GCN (but at much higher price)

So when it comes to the largest chip you have (~600 mm^2), process (clocks) matters since you have no place for improvements. But in lower segments, you can alway make larger chip than competition, especially if you can get better price/mm^2 from foundry

I would assume 56 CUs RX Vega could be sold for $500, if not even less, and be at least in GTX 1080 range

We also don't know if Pascal architecture has a hard 2100mhz limit, meaning the difference maybe slightly more than 10%, but that's purely conjecture on my part so we will take the 10℅, which is in my range of 10-15℅ difference.

In your scenarios you missed my point about increasing performance by 10℅ ( tsmc 16nm FF+. Vs 14nm Lpp) as opposed to increasing cost savings which is what you alluded to, no getting around it AMD would have had measurably increased clock speeds had they been using tsmc, by at least 10℅.

Also when judging a product you have to look at its final performance, in this case Rx 480 vs Rx 580 at launch intended clocks, in this case as I pointed out Rx 580 is significantly less efficient than its predecessor despite only gaining ~5℅ performance, that's clearly abysmal for what was touted as effectively a new process generation (3rd gen), of course at some undervolted frequency 3rd gen 14nm lpp will be more efficient but that's beside the point ( if anything underlines mine), which is 14nm Lpp is not suited to high frequency designs no matter what revision it is in, any improvements only really effect low/ moderate frequencies and don't do much to extend max frequency without paying a nasty price in efficiency well above what your would expect with a foundry/process that specialises in HP silicon (tsmc,intel), as shown by Rx 580.

 

[SpaceBeer]

But if power efficient GP107 is able to reach 1900MHz, it's not only the process that holds Polaris at ~1400 MHz. We know that, since Vega is able to reach ~1600 MHz, while P10/P11 can't. Move Polaris/Vega production to TSMC and they still won't be much better. You might get RX 580 performance at GTX 1060 power level. Maybe not even that. Not a big deal

I would say it is AMD who has to work harder to get more performance from 14 LPP.

 

[tential]

The amount of excuses for Vega at this point over this past year has been unparalleled.

 

[ryzenmaster]

While I'm personally perfectly fine with 1080 level performance, I'm disappointed to see how much power Vega FE needs to just barely reach that level. Lets be intellectually honest for a moment here: we know the consumer variant will not be that much different, and while it is likely to be slightly above 1080 in terms of real world gaming performance, it falls short from what is needed for restoring faith in their GPU department.

For the rx Vega to challenge gtx 1080ti or in fact the upcoming Volta, it would be facing some of the same problems nuclear fusion reactors do.. thermals are not manageable. Again personally 1080 ish performance fits my needs perfectly if it comes at reasonable price, but after all this waiting I'm more frustrated than I am optimistic and I can't blame people if they go for Pascal instead.

Navi needs to be something truly out of this world unless they somehow manage to pull a miracle with rx Vega yet.

 

[Glo.]

It appears we have more answers.

The GPU-Z is recognizing the driver as 17.1.1 because the gaming driver for this GPU may actually be from that period.

Check out what AMD rep has said about Game driver:
https://forum.beyond3d.com/posts/1989522/

That also tells why there is no difference between Game and Pro mode when running gaming apps, and might tell why the GPU is recognized as Greenland.

What has been rumored at the time of Vega presentation(January, the driver 17.1.1 is also from January) that in that moment it was ONLY Fury X drivers, tuned only to work "just" with Vega architecture.

This may answer few questions.

 

[raghu78]

The amount of excuses for Vega at this point over this past year has been unparalleled.

There are no excuses here. imo AMD were better off by not launching Radeon Vega FE now and launching it alongside RX Vega in early August. I do not like this attempt to sell a card to prosumers without fully optimized gaming drivers. Prosumers also game along with doing content creation work on their PCs. So AMD were horribly wrong in trying to sell a product 1 to 1.5 month before its ready. I strongly disagree with AMD's approach. As far as Vega performance I will wait for RX Vega launch day. But what AMD have done with Radeon Vega FE launch is totally unacceptable.

 

[stahlhart]

Thread cleaned -- keep the SJW garbage out of this discussion and stay on topic.
-- stahlhart

 

[Mopetar]

It's time for Raja to go. No longer can we make the excuse that he's just inheriting someone else's problems. Vega is his legacy, and it looks like it's going to cripple RTG for years to come. Enough. Get rid of Raja, get rid of the cheap Chinese design team, and reboot RTG in America where it belongs.

Raja might be a good enough engineer, but it's clear he shouldn't be managing or leading RTG. So many of the problems that they're having would be far less of an issue if they had someone who was better at scheduling, understanding marketing, and managing expectations. Also, for what it's worth I think Raja is an awful presenter and they should get someone with an actual stage presence in front of the crowd. I'm pretty sure they can sift through the engineering team to find at least a few people who have the technical knowledge to speak about the product and answer questions while also having some experience or aptitude for public speaking.

 

[T1beriu]

Thread cleaned -- keep the SJW garbage out of this discussion and stay on topic.
-- stahlhart

SJW garbage? Wow. I would refrain from hypocrite comments like that, especially when you delete defending comments but keep the accusing ones.

Infraction issued for moderator callout.
-- stahlhart

 

[Dribble]

There are no excuses here. imo AMD were better off by not launching Radeon Vega FE now and launching it alongside RX Vega in early August.

They probably had no choice - it had been announced as 1H of the year, it had to come out then or their might have been legal consequences. Tbh if I were AMD right now I'd be focusing on Ryzen, then trying to fight a loosing battle vs Nvidia in high end gpu's. Intel seems to be the easier target.

 

[piesquared]

No difference in performance between Gaming and Pro drivers.

It is the same story as is with Ryzen. New architecture - software is not ready.

No difference in performance, but the card was running at 1348-1528 in gaming mode, while running at 1650 in pro mode.

 

[Phynaz]

They probably had no choice - it had been announced as 1H of the year, it had to come out then or their might have been legal consequences. Tbh if I were AMD right now I'd be focusing on Ryzen, then trying to fight a loosing battle vs Nvidia in high end gpu's. Intel seems to be the easier target.

The problem being Ryzen is in a shrinking market. Makes no sense to focus there.

 

[thilanliyan]

Get rid of Raja, get rid of the cheap Chinese design team, and reboot RTG in America where it belongs.

Wasnt the graphics side in Canada (former ATI)?

 

[EXCellR8]

Very
Energy-consuming
Graphics
Adapter

well i'll be damned

 

[Malogeek]

No difference in performance, but the card was running at 1348-1528 in gaming mode, while running at 1650 in pro mode.

Did you decide to ignore this?

then after re-testing it was back to 1650Mhz max