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

[Bacon1]

Did you know that Quadro P6000 is a little bit faster in games than GTX Titan X? I know its hard to accept, but Pro cards isnt really that much slower in gaming than Geforce.

Here is 3840 core Pascal P6000 vs Titan X 3584 core:

So you are comparing the 7% higher core count + 16GB HBM2 P6000 vs the original Titan XP and are surprised that its faster by 6% and 17%?

Clearly memory bandwidth was an issue since its impossible to have 17% better perf in Hitman with only 7% more cores.

Why not compare the non-HBM2 Pascal Pro cards vs their gaming equivalents? P6000 has 12 TFLOPs vs 11 TFLops on Titan XP... Who would have thought raw performance matters!

Finally Vega comes, trying to mimic nvidia´s tile based rasterization and adding the HBC thing, with both things adding maybe 5 or 6 billion more transistors??
If nvidia gets a new technology and AMD copies it again, it will need like 3x the transistors for the same performance, and maybe HBM 3 or 4...

Why are you claiming that tile based rasterization is a Nvidia technology?? Its been used since the 90s and Nvidia didn't use it til Maxwell in 2014.

Regarding the possible driver issues. Is it possible that AMD put the vast majority of its coder into the ROCm development (link to German article about it) so they get it ready for Radeon Instinct? So my theory is that they left the driver development alone when they had a working but unoptimized driver for Vega and focused all their resources into their deep learning framework.

Yep, for all the outcry and raging that people do on forums, ML and enterprises pay far more and cry far less than gamers.

https://rocm.github.io/dl.html

 

[tential]

So then you are responding to the post with completely different thing.

Actually he was responding to my post. I was saying that pro gpus aren't slower than consumer gpus from Nvidia.
When I posted the post, i saw a new chart and tossed it in there that wasn't posted here before that confused posters into thinking that there was no evidence that pro gpus are just as fast as consumer gpus from Nvidia since my chart didn't include pro gpus.

I then posted the corresponding chart that proves the point.
Either way, the point stands, you arent able to come up with any current data that shows pro gpus performing insanely behind consumer gpus in gaming while I have provided this benchmark here that proves my point

Do you have anything regarding pro performance of gpus in gaming cart consumers to respond with now that I've explained this to you multiple times? You should then pm me, since after this post it's beating a dead horse and I don't think anyone else is confused and wants to hear this again.

 

[tential]

So you are comparing the 7% higher core count + 16GB HBM2 P6000 vs the original Titan XP and are surprised that its faster by 6% and 17%?

Clearly memory bandwidth was an issue since its impossible to have 17% better perf in Hitman with only 7% more cores.

According to you, gaming drivers from the titan xp should easily make up that deficit unless you're now backtracking on Nvidia gaming cards having drivers that significantly improve gaming performance compared to pro drivers?

I'm saying... wait for the RX before final judgement on how it games . They've stated for a while now that the FE is not a gaming first GPU.
Amazing that so many people here can look at Nvidia's Pro vs Consumer GPUs and ignore the massive difference the drivers have but ignore it for AMD's side.

Which is it?...

 

[Bacon1]

The bandwidth is actually poor to be using HBM2. Vega FE got 480GB/s.
Take Volta V100. It got HBM2 too, but the bandwidth is 900GB/s.

Isn't Volta using 4 stacks @ 225 each? Vega is using 2 stacks @ 240 each.

Volta V100 also costs what ~$14k vs Vega FE @ $1k...

Bet Titan V which I guess will be out this year, will feature the same memory as Volta V100.

Titan Volta won't be out this year, and if it is it makes NVidia's whole top end terrible buy... They already released two Titan XPs and to have a third Titan within a 2 year span would really show how little they care for their end users.

1700X and 1800X are basically the same. 3.4GHz/3.8GHz (3.9 with XFR) vs 3.6GHz/4.0GHz (4.1GHz). That's only a 200MHz difference in clock speeds.

Memory makes a huge difference on Ryzen machines though, so w/o knowing if the memory was the same it could make a huge difference, why

The graph was posted simply because it was new information and nothing to do with the post.....

Its been post a few times on these forums actually...

Too bad the actual data begs to differ. You're wrong on this matter.

Already replied to that graph, why not compare P6000 16GB HBM2 with the full Titan Xp not the cut down lower TFLop Titan XP?

Better yet, compare to parts that use the same cores and memory?

The point of the post is that a p6000 is keeps up with a titan xp in gaming despite being a pro card.

Wrong Titan Xp. Is there a reason you need to post the same graph multiple times when its wrong?

 

[Cloudfire777]

According to you, gaming drivers from the titan xp should easily make up that deficit unless you're now backtracking on Nvidia gaming cards having drivers that significantly improve gaming performance compared to pro drivers?




Which is it?...

Exactly.

If Pro cards had drivers that would make it slower....it would be slower. Which it isnt, which the charts showed.

 

[Cloudfire777]

Isn't Volta using 4 stacks @ 225 each? Vega is using 2 stacks @ 240 each.

Volta V100 also costs what ~$14k vs Vega FE @ $1k...



Titan Volta won't be out this year, and if it is it makes NVidia's whole top end terrible buy... They already released two Titan XPs and to have a third Titan within a 2 year span would really show how little they care for their end users.



Memory makes a huge difference on Ryzen machines though, so w/o knowing if the memory was the same it could make a huge difference, why



Its been post a few times on these forums actually...



Already replied to that graph, why not compare P6000 16GB HBM2 with the full Titan Xp not the cut down lower TFLop Titan XP?

Better yet, compare to parts that use the same cores and memory?


Wrong Titan Xp. Is there a reason you need to post the same graph multiple times when its wrong?

What the heck are you talking about?
There is only one Titan Xp.

And a new Titan is probably coming very soon. Need to satisfy their investors you know. People that buy Nvidia cards know that their cards will not be on top for long (notice how I said not replaced, as in the previous cards will still work if newer cards come along )

Yes Volta cost a lot of momey, hence my point that if Vega FE beats Titan Xp in some compute benchmarks, Nvidia have a solid reason to release Vega V.

 

[Bacon1]

Exactly.

If Pro cards had drivers that would make it slower....it would be slower. Which it isnt, which the charts showed.

You are comparing apples and oranges. 16GB HBM2 more core P6000 vs the original Titan XP not the newer Titan Xp.

Compare equal core + memory.

 

[Bacon1]

What the heck are you talking about?
There is only one Titan Xp.

https://www.nvidia.com/en-us/geforce/products/10series/titan-x-pascal/

https://www.nvidia.com/en-us/geforce/products/10series/titan-xp/

Nvidia seems to differ....

 

[tential]

You are comparing apples and oranges. 16GB HBM2 more core P6000 vs the original Titan XP not the newer Titan Xp.

Compare equal core + memory.

And look now you're completely ignoring the gaming drivers the titan has over the quadro. Why when they make a massive difference in your OWN words?

 

[Cloudfire777]

https://www.nvidia.com/en-us/geforce/products/10series/titan-x-pascal/

https://www.nvidia.com/en-us/geforce/products/10series/titan-xp/

Nvidia seems to differ....

Thats Titan X and Titan Xp

Pretty confusing naming, I agree

 

[Bouowmx]

NVIDIA Quadro P6000 does not have 16 GB HBM. It has 24 GB, 384 bit, 9 GT/s, 432 GB/s, GDDR5X, less bandwidth than the original TITAN X (Pascal).

Getting so side-tracked.

 

[Cloudfire777]

You are comparing apples and oranges. 16GB HBM2 more core P6000 vs the original Titan XP not the newer Titan Xp.

Compare equal core + memory.

Look, it really isnt that difficult:

300 cores more on the Quadro M6000 perform pretty much like any Geforce card with 300 cores would over the Titan X.

If drivers on the M6000 was sucky in gaming as some people suggests for Pro cards (for Vega FE as people use as argument for poor Vega performance), it wouldnt perform like that in the result I posted

 

[Bacon1]

And look now you're completely ignoring the gaming drivers the titan has over the quadro. Why when they make a massive difference in your OWN words?

No... I'm saying remove all the other variables because its pointless to talk about drivers when you have two completely different GPUs. Use the same GPU but driver as Gaming vs Pro.

Compare P5000 vs 1080 which afaik is the closest we can:

Cores: 2560 each
Memory: 16GB GDDR5X vs 8GB GDDR5X
Core clocks appear to be the same at around 1740 each.

https://hothardware.com/reviews/nvidia-quadro-p6000-and-p5000-workstation-gpu-reviews?page=6

https://hothardware.com/ContentImages/Article/2562/content/time-spy.png

1080 is 7% faster while having half the VRAM.

Like I said, its still not complete apples to apples because we don't know what the actual clocks were during the test and the VRAM is half as much on the 1080, but at least its way closer than the P6000 vs Original Titan XP you are trying to compare.

Wait until RX drivers are out and we'll have an easy way to tell... Vega FE Gaming mode vs Pro mode.

Thats Titan X and Titan Xp

Pretty confusing naming, I agree

Yes, and the one tested in that chart was the original Titan XP, not the newer higher core count Titan Xp.

 

[Bacon1]

NVIDIA Quadro P6000 does not have 16 GB HBM. It has 24 GB, 384 bit, 9 GT/s, 432 GB/s, GDDR5X, less bandwidth than the original TITAN X (Pascal).

Getting so side-tracked.

You are right, I was thinking P6000 was the big one, that's the GP100 which has 16GB HBM2.

Either way, comparing P6000 vs Titan X(P/p) doesn't make sense as they are widely different.

 

[tential]

No... I'm saying remove all the other variables because its pointless to talk about drivers when you have two completely different GPUs. Use the same GPU but driver as Gaming vs Pro.

Compare P5000 vs 1080 which afaik is the closest we can:

Cores: 2560 each
Memory: 16GB GDDR5X vs 8GB GDDR5X
Core clocks appear to be the same at around 1740 each.

https://hothardware.com/reviews/nvidia-quadro-p6000-and-p5000-workstation-gpu-reviews?page=6

https://hothardware.com/ContentImages/Article/2562/content/time-spy.png

1080 is 7% faster while having half the VRAM.

Like I said, its still not complete apples to apples because we don't know what the actual clocks were during the test and the VRAM is half as much on the 1080, but at least its way closer than the P6000 vs Original Titan XP you are trying to compare.

Your post is not remotely relevant to gaming. Please address the subject.

Either way since when was this 7% difference massive? Because this 7% massive difference won't help Vega.

 

[Bacon1]

Your post is not remotely relevant to gaming. Please address the subject.

How is it not relevant? Its the same benchmark you guys were showing but flipped and comparing the P6000 vs Titan XP instead of the much closer in hardware P5000 vs 1080! Its a TimeSpy benchmark which last I checked, people seemed to think was relevant to gaming performance... Test against games that need many driver optimizations and you'll find it's a bigger difference.

Why not wait for the Vega RX to see how it games compared to the Prosumer Vega FE? Its less than a month away.

 

[tential]

Its less than a month away.

Who said that? Did amd confirm reviews of rx Vega coming out in less than a month? Or do you have insider knowledge?
We know it will launch. Not availability or launch availability.

If drivers are really this pathetically incomplete then this is only your own speculation that we will know how rx Vega will perform. If we want to play it by pure facts like you suggest, then we don't have a clue when rx Vega will be reviewed by reviewers or available for purchase. It could be September for all we know.

We only currently know it will be launched, whatever that means is anyone's best guess.

Edit :
How was Volta remotely supposed to be afraid of vega again?

 

[Bacon1]

Who said that? Did amd confirm reviews of rx Vega coming out in less than a month? Or do you have insider knowledge?
We know it will launch. Not availability or launch availability.

https://twitter.com/Radeon/status/881233662821486593

We made a promise to launch RX Vega during SIGGRAPH, Ali. In fact, our CEO made that promise. We’re going to keep it

https://twitter.com/killyourfm/status/881204308863180800

And you know Lisa Su said over a month ago they said they'd launch Radeon RX Vega at SIGGRAPH:

https://youtu.be/oWMGUdWVw2Y?t=53m23s

Not going to bother responding to the rest of your trolling, don't feel like getting yet another infraction from responding to your comments.

 

[Kenmitch]

The bickering back and forth is both amusing and pointless.

Waiting for the official launch of Vega RX seems like the sane thing to do.

 

[Magee_MC]

While that is somewhat true, it is also true that they have been working on drivers ever since they started with the simulations.
This is how they get their baseline performance level, and then they decide what needs to be optimized, rinse & repeat.
Then, after first silicon is back, they keep optimizing the drivers, run it through test suites, and compare that to the baseline they had in the simulations.
If they don't match, they see what went wrong, and depending on the problem, could result in a respin of the silicon.

It would be pretty hard to not know the ballpark performance level after a few iterations of the above, unless the simulator was incredibly wrong, and nobody would reasonably think that basic driver optimizations could get you a ton of extra performance.
Game specific optimizations are a different animal, there, you pretty much know what needs to be improved, and you can work on those things individually.

The question that I don't know the answer to is, if AMD had to do a significant change to the layout of the silicon in order to remove some kind of show stopping problem, could that require significant rework of the drivers that they already had. The silicon that they had might have been nowhere near what they were expecting and needed, so they had no choice but to delay Vega and fix the problem and take the hit on the delay.

If so, then I think that could explain why FE released on old functional, but not optimized drivers and why RX Vega is so delayed. They needed Vega to perform as expected and needed something out in 1H17 to meet what they had told investors. They were stuck between a rock and a hard place with no good choices, and took the one that gave them the hit for the short term, but allowed them to succeed in the long term. If that's not what happened, then I've got nothing else at the moment to explain this confusing morass of a release.

 

[Magee_MC]

The bickering back and forth is both amusing and pointless.

Waiting for the official launch of Vega RX seems like the sane thing to do.

You and your reasonable approach of waiting to see what is actually released.

It is amusing, and it does give us something to do while we wait another month.

 

[Azix]

what gets to me is I don't see actual investigations being done by these people. pcper, gamersnexus for starters. Did they compare dx12 to dx11 where in the past AMD saw gains? opengl doom to vulkan doom to see if the performance is better in vulkan?

there are ways they can go about seeing if the gpu is actually working as it should.

Vega FE and Vega RX will have the same cores, the same silicon.
AMD used a dual driver setup with Vega FE. One that utilize the professional side to do professional workloads. Like designing a game. The other one runs a gaming driver code, for the same developer to run his game to see how smooth it will run on a gaming card. Thats what AMD themselves state and the idea behind this is kinda cool.

Why would Vega RX run much different then? Vega FE is 10-30% below GTX 1080 in games in gaming mode. At best it sounds to me that it can get close to GTX 1080. But even that looks like a stretch.

PcPer say Vega RX most likely will get around 10% more than Vega FE.

pcper is as clueless as everybody.

gaming mode is there to enabling gaming features, not gaming performance. its sometimes even slower than pro mode. Its simply to allow developers to test things out

 

[jpiniero]

The question that I don't know the answer to is, if AMD had to do a significant change to the layout of the silicon in order to remove some kind of show stopping problem, could that require significant rework of the drivers that they already had. The silicon that they had might have been nowhere near what they were expecting and needed, so they had no choice but to delay Vega and fix the problem and take the hit on the delay.

Perhaps it's just as simple as a new stepping that allows higher clock speed. The additional speed might not be much but it might be the difference between consumer Vega getting cancelled or not. Miners aren't going to touch this so they have to be pretty confident that it'd sell without much of a 'discount' if the high cost rumors are true.

 

[casiofx]

Perhaps it's just as simple as a new stepping that allows higher clock speed. The additional speed might not be much but it might be the difference between consumer Vega getting cancelled or not. Miners aren't going to touch this so they have to be pretty confident that it'd sell without much of a 'discount' if the high cost rumors are true.

Just how much the new stepping can get? Look st polaris rx480 to rx580, slight increase in performance with the cost of higher power consumption. This vega fe already using 300watts, you're expecting the new stepping to consume 350-400 watts?

 

[beginner99]

Just how much the new stepping can get? Look st polaris rx480 to rx580, slight increase in performance with the cost of higher power consumption. This vega fe already using 300watts, you're expecting the new stepping to consume 350-400 watts?

Depends on what was bad in the old stepping. Another theory is, that initial hardware had some kind of bug preventing one or more of the new features from working correctly. Therefore also explains the old "Fiji" drivers.

Instead of a hardware issue maybe driver team made a mistake and that mistake can lead to crashes,corruption, inaccuracies,...in general a wrong output. Now they needed to rework the whole driver.

Or nothing at all: Driver team just ran out of time, miscalculation of time needed to create the drivers. However this one theory has problems to explain why Vega FE is using drivers from January. If there was not a major, basic issue in the driver or hardware, then Vega FE could just use say a stable driver from May.

So yeah I think there was a major defect either in hardware or the driver and that has to be fixed now.