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

[.vodka]

Seems like no miner will run on the card. No mining results so far, thankfully

Very boring to watch. His actions or lack of make me believe he knows something isn't right about the results.

Yeah, I get that vibe too.


Oh well, RX Vega will clear any doubts in a month.

 

[JDG1980]

Seems to me that there are two possibilities at this point.

Pessimist/Realist: Vega is really as bad as it looks from the Founders Edition results; none of the bottlenecks holding back Fiji were meaningfully fixed, so Vega basically acts like an overclocked Fury X. The driver development team is working frantically to try to boost Vega's performance another 10% or so in the next month, because otherwise, RX Vega will have to be sold so cheap that AMD will take a loss.

Optimist: The FE driver is a 'safe' driver that is stable, but doesn't support any of the new features (at least not on existing DX11 titles): no Draw Stream Binning Rasterizer, no Primitive Discard Accelerator, maybe not even the improved load balancing. It's currently acting like an overclocked Fury X because it's running with basically the Fiji (or at best Polaris) feature set. The current state of the driver that supports the new features is still unsuitable for release - maybe it crashes every 5 minutes, or has major visual bugs, or something else bad enough that even on a bleeding-edge card it would be unacceptable. Therefore, the driver development team is working frantically to get the 'real' Vega driver in an acceptable state by the time RX Vega's release window hits.

I'd like the Optimist argument to be true, but I'm skeptical. This is because there have been too many such unfulfilled claims in the past. Remember when Bulldozer just needed a new Windows scheduler to shine? Or when we couldn't believe the leaks for Fury, or Polaris, because these weren't with release drivers and surely they would go up another 10%-20% on launch day?

It's hard to believe that AMD would really come up with a product as bad as Vega appears to be from the FE benchmarks. But Bulldozer proves they can screw things up bad enough that it's an actual regression. Let's hope they didn't do so, but be prepared for the strong possibility that they did.

 

[Malogeek]

Seems like no miner will run on the card. No mining results so far, thankfully

The random YouTube guy last night was getting 30mh on his FE, basically equivalent of a 580?

 

[piesquared]

Really? Because my quoted part was in the same sentence. How is that not ignoring it? Your quote was specifically misleading.

I think you are projecting. Assuming you can read, look at the summary. It claims a lower clock while in gaming mode. And just above that says the performance is the same. Comprende?

 

[raghu78]

Seems to me that there are two possibilities at this point.

Pessimist/Realist: Vega is really as bad as it looks from the Founders Edition results; none of the bottlenecks holding back Fiji were meaningfully fixed, so Vega basically acts like an overclocked Fury X. The driver development team is working frantically to try to boost Vega's performance another 10% or so in the next month, because otherwise, RX Vega will have to be sold so cheap that AMD will take a loss.

Optimist: The FE driver is a 'safe' driver that is stable, but doesn't support any of the new features (at least not on existing DX11 titles): no Draw Stream Binning Rasterizer, no Primitive Discard Accelerator, maybe not even the improved load balancing. It's currently acting like an overclocked Fury X because it's running with basically the Fiji (or at best Polaris) feature set. The current state of the driver that supports the new features is still unsuitable for release - maybe it crashes every 5 minutes, or has major visual bugs, or something else bad enough that even on a bleeding-edge card it would be unacceptable. Therefore, the driver development team is working frantically to get the 'real' Vega driver in an acceptable state by the time RX Vega's release window hits.

I'd like the Optimist argument to be true, but I'm skeptical. This is because there have been too many such unfulfilled claims in the past. Remember when Bulldozer just needed a new Windows scheduler to shine? Or when we couldn't believe the leaks for Fury, or Polaris, because these weren't with release drivers and surely they would go up another 10%-20% on launch day?

It's hard to believe that AMD would really come up with a product as bad as Vega appears to be from the FE benchmarks. But Bulldozer proves they can screw things up bad enough that it's an actual regression. Let's hope they didn't do so, but be prepared for the strong possibility that they did.

The realist is neither an optimist or pessimist. That person is going to not bother with any leaks before actual press reviews of RX Vega . The realist kind of person is not emotional unlike the other two who do tend to sway according to their emotion and bias.

 

[.vodka]

Seems like tiled rendering is confirmed, this isn't good old GCN anymore.

Makes it all even more weirder than before.


edit: tried trianglebin on my 290. It behaves completely different, Vega is definitely tiled rendering based.

 

[IllogicalGlory]

Tiled rendering is active, HBCC is active. Looks like the drivers are working as intended.

I guess Vega's arch improvements really are in the margin of error range, maybe even less of an improvement than Polaris?

Ugh, what has AMD been doing? They've spent two years developing a 1400MHz Fury X?

 

[Elfear]

GPU of course under full load. It's a 300w GPU as per AMD's specs.

Strictly GPU. The power meters are hooked up to the GPU. About 20-25w are drawn from the PCIe slot, the rest from the PSU.

Thanks gents.

That is crazy high power consumption.

I still say something has to be very wrong. So far indications point to 1080 performance. Consider this:

A 1080 is ~35% faster than a Fury X and Vega consumes about 36% more power than Fury X (300W vs 220W). I realize power draw and performance don't have a linear relationship but it's still mind-boggling to think that there seems to be no improvements in moving down a full node nor in the architectural improvements NCU is supposed to bring.

 

[Mopetar]

The random YouTube guy last night was getting 30mh on his FE, basically equivalent of a 580?

I wonder if AMD would intentionally release a driver that gimps mining performance, at least while the card is new and supply is limited.

 

[Malogeek]

I wonder if AMD would intentionally release a driver that gimps mining performance, at least while the card is new and supply is limited.

Well it's about identical to a Fury X for ethereum mining performance and they both have the same shader structure.

 

[swilli89]

Yeah, gaming optimizations. But we don't know if AMD is still using a Fiji driver like months ago. That would explain why we're seeing only Fury X performance just with higher clocks.

Trust me.. I hope RX Vega is much faster than this for everyone's sake, consumers all around.. but think about what you are postulating.. That AMD released a Vega card with some hobbled together Fury driver? They've had working silicon back for almost a year now and they were showing it running game code almost 7 months ago.. These are "Vega" graphics drivers. Really the only hope we can have is that AMD, while not intentionally crippling Vega FE, are going to be unlocking some features of Vega that will vastly speed up game performance. After all, Vega FE merely needs to run games to test and debug, not win benchmarks.

 

[.vodka]

These are the ones I could catch. He'll retest and post a review tomorrow. Power consumption was 280-300W at the card for the performance seen here.

There's power throttling going on unless the fan is cranked all the way to 100%, justifying the water cooled version that will probably run 1600MHz all the time, the air cooled version tested here runs around 1400MHz at 83-84°C...

We'll see what RX Vega brings to the table.


He's testing specviewperf up next.

 

[notT1beriu]

RX Vega will have to be sold so cheap that AMD will take a loss.

No it won't. Why would anyone sell something that loses them money?! AMD is not Bugatti.

Even if it's going for $450-500, in the worst case scenario, it will bring money to at least cover the R&D expenses, but not make a profit. The card won't cost $400 to manufacture, the chip will be somewhere in the $100-250 range that they sell to AIBs.

 

[Karnak]

These are the ones I could catch. He'll retest and post a review tomorrow. Power consumption was 280-300W at the card for the performance seen here.

There's power throttling going on unless the fan is cranked all the way to 100%, justifying the water cooled version that will probably run 1600MHz all the time, the air cooled version tested here runs around 1400MHz at 83-84°C...

We'll see what RX Vega brings to the table.


He's testing specviewperf up next.

RX Vega has to be atleast 40% faster than the Frontier Edition. Good luck with that AMD.

 

[Head1985]

Seems like tiled rendering is confirmed, this isn't good old GCN anymore.

Makes it all even more weirder than before.


edit: tried trianglebin on my 290. It behaves completely different, Vega is definitely tiled rendering based.

So no magic driver month later.Vega runs like it should.
I think vega is just memory bandwidth bottleneck.Thats why it sucks in games.Less bandwidth than furyx+no new delta color compresion(only polaris one whitch is far worse than pascal one)

 

[RoarTiger]

RX Vega has to be atleast 40% faster than the Frontier Edition. Good luck with that AMD.

Why? So many seem to think AMD must beat a 1080ti to accomplish anything. AMD just need something to sell in the sub $500, 1070/1080 and upcoming 2070/2080 segment. I certainly dont understand how their flagship can appear to be so far behind NVIDIA top card after all these delays but it really doesnt matter in the long term. AMD now has Ryzen to prop them up as they make improvements. They just need actual products in certain high selling market segements. 1080 performance for $400 will last them for another year if they can deliver that.

 

[swilli89]

No it won't. Why would anyone sell something that loses them money?! AMD is not Bugatti.

Even if it's going for $450-500, in the worst case scenario, it will bring money to at least cover the R&D expenses, but not make a profit. The card won't cost $400 to manufacture, the chip will be somewhere in the $100-250 range that they sell to AIBs.

If they're selling a chip for $250 to an AIB as a worst case, they will be fine. 450mm2 yields on more than good enough on 14nm LP by now I am sure of that. Sure HBM2 adds to the cost but they are still probably at a 30-35% margin for that.

Keep in mind AMD's gross average profit margin.. including all those low level mobile's, deeply discounted FX processors, and very cheap Polaris cards over the last year, and their average margin was about 35%

 

[Karnak]

Why? I dont understand why so many seem to think AMD has to beat a 1080ti to accomplish anything. AMD just need something to sell in the sub $500 1070/1080 and upcoming 2070/2080 segment. I certainly dont understand how their flagship can appear to be so far behind NVIDIA top card after all these delays but it really doesnt matter in the long term. AMD now has Ryzen to prop them up as they make improvements. They just need actual products in certain high selling market segements. 1080 performance for $400 will last them for another year if they can deliver that.

They don't need to IMO, but even if RX Vega is 40% faster than the FE it would be still slower than the 1080ti looking at the pcper benchmarks. And since there are so many and significant changes it shouldn't act just like a Fury X with higher clock speeds. But that is what we are seeing at the moment. And that is bad, really bad.

 

[QualityTime]

Seems like tiled rendering is confirmed, this isn't good old GCN anymore.

Makes it all even more weirder than before.


edit: tried trianglebin on my 290. It behaves completely different, Vega is definitely tiled rendering based.

I was one of the people in chat asking for trianglebin and it didn't seem to be doing any tiling.
Standard GCN behavior, fully render one triangle, render next, etc.

 

[.vodka]

Interesting. I must have missed that specific moment in the stream.

What the hell..?

Still, my 290 / normal GCN doesn't do that. It flashes the whole upper triangle weirdly all at once.

 

[Head1985]

No that is tile rasterizer.Normal GCN looks different

 

[tential]

These are the ones I could catch. He'll retest and post a review tomorrow. Power consumption was 280-300W at the card for the performance seen here.

There's power throttling going on unless the fan is cranked all the way to 100%, justifying the water cooled version that will probably run 1600MHz all the time, the air cooled version tested here runs around 1400MHz at 83-84°C...

We'll see what RX Vega brings to the table.


He's testing specviewperf up next.

Remember when people said those who post/compile benches first control the narrative? This is one of the first posts, and I'm sure there are others on reddit/neogaf now.

Honestly, we can all agree on something.
No one can predict how badly AMD will screw up a GPU launch.

 

[.vodka]

I'm just posting verbatim what the guy said in the stream. I'm not manipulating anything. That is the state of Vega FE for gaming today 29/06/17. I suppose the results for the nV cards were taken on the same test bench (5960x + 1500w PSU + open air environment with a crapload of air for the cards to work their best)

He also said he went over the phone with AMD BEFORE showing that on the stream, and they were ok with it... and that he'd have it posted on his site by tomorrow.

AMD was OK with these results.

They know it is this bad.

 

[tential]

Are Nvidia workstation GPU drivers also good at gaming while AMD's aren't and you need the dedicated gaming card for gaming?

I don't think that's the case personally.
I don't think you need the "dedicated gaming card" to get a very close look at RX Vega's performance.

 

[piesquared]

These are the ones I could catch. He'll retest and post a review tomorrow. Power consumption was 280-300W at the card for the performance seen here.

There's power throttling going on unless the fan is cranked all the way to 100%, justifying the water cooled version that will probably run 1600MHz all the time, the air cooled version tested here runs around 1400MHz at 83-84°C...

We'll see what RX Vega brings to the table.


He's testing specviewperf up next.

That's not quite accurate. They didn't need to crank it to 100% to prevent throttling, that is just all they tried. They went from stock straight to 100%, then never tried anything in between. IIRC the GPU was running mid 60c at 100%, so it definitely doesn't need to be max to prevent throttling. I'd guess a small bump in the fan profile would do it.