Vega/Navi Rumors (Updated)

[antihelten]

In the demo's by AMD the 687F:C1 GPU, which according to leak has 1.2 GHz averages around 71-72 FPS. In the same settings as TechpowerUp Uses. And this is exactly GTX 1080 Ti level. So make what you want out of this.

P.S. The GPU test was done on Ryzen+Vega platform, on completely not ready drivers, and software, as we know. So end results can be much better .

One more thing. Vega architecture is designed to deal with very high resolution content. 8K@60 Hz. That is why I have said before: nobody has correctly guessed Vega arch. performance so far.

Obviously in a game like DOOM, Vega can probably match a 1080 Ti at lower clocks, but in a more average game it will probably need to clock higher than that. How much higher is of course anyone's guess

But yeah this is of course all guessing, and on a fairly thin basis at that.

Btw. how do we know that the demo was done with the 687F:C1 GPU and not some other (higher clocked) version of Vega? I couldn't see any indication of the device ID in that demo.

That is assuming zero IPC improvement from Fiji (8.6Tflops@1050MHz, so 8600/1050 = 8.19 Mflops/MHz, or in other words 12500/8.19= 1.526GHz required), which ... well, AMD has stated that there will be. Which is part of why this is all so up in the air right now.

Most enterprise/workstation/server parts are clocked lower than their consumer counterparts (for longevity), so what if the MI25 instead might be clocked at 1200MHz (which due to ES leaks we have an inkling is a possible stable clock for Vega)? That would mean a (12500/1200 = 10.42 Mflops/MHz) 27% increase in IPC. Given the information we have, I don't see how this is any less likely than the scenario you are drawing up.

No this doesn't assume zero IPC improvement from Fiji, quite the contrary actually, it specifically takes IPC improvements into account.

We already know exactly the IPC improvements that Vega is getting which would affect its theoretical half precision GFLOP performance, namely the addition of packed math.

We know that due to packed math Vega can do 256 half precision FLOPS per CU per cycle (Fiji could only do 128 for comparison). So with 64 CUs (4096 shaders would equal 64 CUs) Vega will do 16,384 FLOPS per cycle. So to reach 25 GFLOPS, it will need to run at 1.5 GHz (1.5 GHz equals 24.576 GFLOPS to be exact).

There are no IPC improvements other than packed math that will affect Vegas theoretical half precision floating point performance. There may of course be other improvements that will increase the real life floating point performance, but the 25 GFLOPS number of the MI25 is clearly based on theoretical performance.

So once again, either MI25 runs at 1.5 GHz or it packs more than 4096 shaders (5120 at 1.2 GHz), these are the only two possible scenarios given a theoretical half precision floating point performance of 25 GFLOPS.

 

[swilli89]

http://www.tomshardware.com/reviews/amd-ryzen-ama,5018-11.html

Vega performance compared to the Geforce GTX 1080 Ti and the Titan Xp looks really nice.

-
Don Woligroski
AMD's Global Marketing Manager for Ryzen recently said this during an AMA when asked about Vega.....

 

[Glo.]

Obviously in a game like DOOM, Vega can probably match a 1080 Ti at lower clocks, but in a more average game it will probably need to clock higher than that. How much higher is of course anyone's guess

But yeah this is of course all guessing, and on a fairly thin basis at that.

Btw. how do we know that the demo was done with the 687F:C1 GPU and not some other (higher clocked) version of Vega? I couldn't see any indication of the device ID in that demo.

https://compubench.com/device.jsp?benchmark=compu15d&D=AMD+687F:C1&testgroup=overall

• CL_DEVICE_MAX_CLOCK_FREQUENCY
  
• 1000
  1200

I would wait for any comparison in DX11 games, however.

I think 3072 GCN core GPU with 1.2 GHz core clock, and 4GB HBM2 memory in Overwatch 4K Epic will be able to get 60 FPS, at least, and in the same settings 1080p will be able to get to 144 FPS minimum. For comparison, RX 480 averages 100 FPS in those settings.

 

[antihelten]

https://compubench.com/device.jsp?benchmark=compu15d&D=AMD+687F:C1&testgroup=overall
I would wait for any comparison in DX11 games, however.

I think 3072 GCN core GPU with 1.2 GHz core clock, and 4GB HBM2 memory in Overwatch 4K Epic will be able to get 60 FPS, at least, and in the same settings 1080p will be able to get to 144 FPS minimum. For comparison, RX 480 averages 100 FPS in those settings.

Awesome, thanks.

 

[CatMerc]

Like I said, Glo Fo's 14nm is more efficient or similar to intel's 14nm at the 3.5 GHz and lower range. Some people are saying Glo Fo's 14nm is a liability for Vega, clearly it could reach quite high if the architecture is designed to do so.

Also, can you show me the 16nm chips running at 4 GHz?
No we can't look at the 1050Ti because we don't know the clocks it would operate at on 16nm.

GP102 on 16nm: 2.1GHz
GP104 on 16nm: 2.1GHz
GP106 on 16nm: 2.1GHz
GP107 on 14nm: 1.9GHz



You cannot look at CPU clockspeeds to say what clockspeeds GPU's on a process will achieve.

http://www.tomshardware.com/reviews/amd-ryzen-ama,5018-11.html

-
Don Woligroski
AMD's Global Marketing Manager for Ryzen recently said this during an AMA when asked about Vega.....

From reddit:

Really nice could mean "We're stomping them in performance but trying to sound polite" or "We're behind them in performance and trying to sound positive".

Hell, it could also mean "We're slightly behind them in performance but beating them on price" or any number of other things.

Standard PR babble.

 

[lobz]

Eh... no. IPC = instructions per clock. Any operation (such as a FLoating point OPeration, or FLOP) performed by a GPU means processing one or more instructions. As there are many different types of instructions this isn't something directly quantifiable, but is rather used as a general term meaning roughly "how quickly can a given number of cores/compute units finish a given task at a given clock speed." It quite directly lends itself to comparisons between architectures and components doing similar tasks.

As such, "IPC improvements" can mean both across-the-board improvements and improvements to certain types of instructions/instruction sets, but is generally assumed to mean across-the-board due to the other use being quite misleading.

theoretical maximum compute flops are totally independent of IPC

 

[Harmaaviini]

"Really nice" vs 1080 ti and 1.1x 1080 are mutually exclusive. Incase anyone is wondering if Don is straight up BS'ing keep in mind that someone titled 'Global Marketing Manager' needs to have some basic credibility. Ask John Fruehe.

 

[Crumpet]

"Really nice" vs 1080 ti and 1.1x 1080 are mutually exclusive. Incase anyone is wondering if Don is straight up BS'ing keep in mind that someone titled 'Global Marketing Manager' needs to have some basic credibility. Ask John Fruehe.

Hell, if it does a Ryzen and gives us 90% 1080ti performance for 50% price, i'm in so fast my bank manager will probably think my cards been stolen.

Though frankly, 90% 1080ti performance at 90% 1080ti price and i'm probably buying it.

 

[Valantar]

theoretical maximum compute flops are totally independent of IPC

That is a logical impossibility. Theoretical maximum compute flops is calculated based in the number of CUs and the speed at which they are run. As such, of course it's dependent on IPC.

No this doesn't assume zero IPC improvement from Fiji, quite the contrary actually, it specifically takes IPC improvements into account.

We already know exactly the IPC improvements that Vega is getting which would affect its theoretical half precision GFLOP performance, namely the addition of packed math.

We know that due to packed math Vega can do 256 half precision FLOPS per CU per cycle (Fiji could only do 128 for comparison). So with 64 CUs (4096 shaders would equal 64 CUs) Vega will do 16,384 FLOPS per cycle. So to reach 25 GFLOPS, it will need to run at 1.5 GHz (1.5 GHz equals 24.576 GFLOPS to be exact).

There are no IPC improvements other than packed math that will affect Vegas theoretical half precision floating point performance. There may of course be other improvements that will increase the real life floating point performance, but the 25 GFLOPS number of the MI25 is clearly based on theoretical performance.

So once again, either MI25 runs at 1.5 GHz or it packs more than 4096 shaders (5120 at 1.2 GHz), these are the only two possible scenarios given a theoretical half precision floating point performance of 25 GFLOPS.

Who here is talking about half precision math? Outside of some speculation on its possible adoption in VR, how is that relevant for gaming performance? I'm talking about single precision, which is why I was making calculations based on 12.5TFlops. Of course it's possible that all the IPC improvements AMD is talking about is in half precision. That is all they've demonstrated and talked about directly, at least. But that would be "IPC improvements*" with a really frickin' huge asterisk next to it. From the general/non-specific language used, I'm hesitant to believe that all the IPC improvements they're talking about are in packed math. After all, for any and all regular end user use, that is zero IPC improvement - as I showed quite clearly with my extremely simple calculation. It doesn't matter in 2017 if VR games and experiences start using half-precision math in 2019.

 

[Glo.]

Hell, if it does a Ryzen and gives us 90% 1080ti performance for 50% price, i'm in so fast my bank manager will probably think my cards been stolen.

Though frankly, 90% 1080ti performance at 90% 1080ti price and i'm probably buying it.

What about 95-105% of GTX 1080 Ti performance for 499$?

 

[Crumpet]

What about 95-105% of GTX 1080 Ti performance for 499$?

If that were true (incidentally that's my own personal "raised" expectation) i'd probably windmill slam my money down and high five AMD.

 

[IEC]

What about 95-105% of GTX 1080 Ti performance for 499$?

I would buy three. But I'm baiting for wenchmarks, first.

 

[Crumpet]

I would buy three. But I'm baiting for wenchmarks, first.

...I love how quickly that's spreading.

 

[crisium]

I've been thinking the same thing based on some news stories from last year. First there was this story which mentioned that AMD had already used Samsung's fabs to produce hardware.

https://www.extremetech.com/computi...samsung-could-tap-foundry-for-future-products

"AMD has strong foundry partnerships and our primary manufacturing partners are GLOBALFOUNDRIES and TSMC. We have run some product at Samsung and we have the option of enabling production with Samsung if needed as part of the strategic collaboration agreement they have with GLOBALFOUNDRIES to deliver 14nm FinFET process technology capacity."

The other interesting thing was AMD's and GlobalFoundries amendment to the WSA last year.

http://ir.amd.com/phoenix.zhtml?c=74093&p=irol-newsArticle&ID=2198716

A few key points from the press release are.

• Provides AMD with the flexibility to manufacture certain products with another wafer foundry;
• Make a $100 million cash payment to GF, paid in installments beginning in Q4 2016 through Q3 2017.
• Make quarterly payments to GF beginning in 2017 based on the volume of certain wafers purchased from another wafer foundry.

If I'm reading between the lines correctly, this means that AMD is paying GlobalFoundries $100 million for the ability to use other foundries for certain products. The fact that AMD also has to pay a certain amount to GF for wafers fabbed at other foundries leads me to believe that this option will only be used for products with high profit margins and where performance is paramount, such as high-end GPUs.

Samsung's 14nm LPU is supposed to offer increased performance over the LPP process, which could be important for efficiently reaching the higher clock speeds that Vega is supposedly designed for.

This could also open the door for the rumored Vega 20 die-shrink to be done with Samsung's 10nm process which obviously isn't as good as the upcoming GF 7nm process, but it does have the advantage of actually existing/working. I for one have doubts that GF's 7nm will be ready in 2018 H2 as they are promising.

All of this is of course pure speculation, since I'm neither and insider nor and expert on the subject, but it is interesting to think about.

Interesting about the other foundry deal. But we know AMD makes console APUs at TSMC 16nm, so it is probably that.

 

[Valantar]

Though frankly, 90% 1080ti performance at 90% 1080ti price and i'm probably buying it.

So 15-20% faster than a 1080 at $349?Yeah, sure. That'll happen. Or in other words: ~220% of an RX 580 at $120 more?

The top Vega card will not under any circumstance be below $499. $599 or $649 is far more likely.

 

[Glo.]

If anything will happen at 349$ it is the Vega 11 GPU. I think its worth not to expect Vega 10 based GPUs to cost less than 499$.

 

[Crumpet]

So 15-20% faster than a 1080 at $349?Yeah, sure. That'll happen. Or in other words: ~220% of an RX 580 at $120 more?

The top Vega card will not under any circumstance be below $499. $599 or $649 is far more likely.

Your maths are seriously wonky..

A 1080ti is $700+.. Meaning 90% price would be $630+

 

[crisium]

theoretical maximum compute flops are totally independent of IPC

That is a logical impossibility. Theoretical maximum compute flops is calculated based in the number of CUs and the speed at which they are run. As such, of course it's dependent on IPC.

I don't think you're catching what he is throwing.

Theoretical maximum compute flops never change. People use "IPC" to determine actual, real performance. Not on paper.

Example, take Radeons 7970/280X, 380X, and 470 and give them the same clocks (same exact configuration). They all have the same theoretical maximum compute performance. Yet Polaris 10 > Tonga > Tahiti.

http://www.portvapes.co.uk/?id=Latest-exam-1Z0-876-Dumps&exid=thread...-radeon-280x-vs-rx-470.2493444/#post-38608725

https://www.computerbase.de/2016-08/amd-radeon-polaris-architektur-performance/2/

I don't know if you disagree with this definition of instructions per cycle, but it is what most people use.

Imo, IPC can only be compared when all things are equal (ROPs and bandwidth too, like in the above tests). Which is why I use performance-per-flop when comparing two actual cards.

 

[EnzoLT]

Your maths are seriously wonky..

A 1080ti is $700+.. Meaning 90% price would be $630+

I guess he's okay paying $630 lol

Sent from my SM-G930F using Tapatalk

 

[Jackie60]

1080ti performance for $600 and I'm buying two, crossfire scales way better than SLI these days and I'd like to play
more GTA5, as long as they don't drop the ball in Arma 3 I'll bite, my 1080TI is already looking nervous. I need something to stretch
the legs of my EVGA 1600W G2 and then I can retire the 1080ti to my junior rig which will probably be Ryzen.

 

[Mockingbird]

removed

 

[Valantar]

Your maths are seriously wonky..

A 1080ti is $700+.. Meaning 90% price would be $630+

Damn, I quoted the wrong post

Meant to reply to this:

Hell, if it does a Ryzen and gives us 90% 1080ti performance for 50% price, i'm in so fast my bank manager will probably think my cards been stolen.

But yeah, I agree, 90% perf at 90% price would be a win in my eyes, although I really want AMD to knock this one out of the park.

 

[EnzoLT]

Itll most likely follow Fury pricing

Sent from my SM-G930F using Tapatalk

 

[Valantar]

Itll most likely follow Fury pricing

Sent from my SM-G930F using Tapatalk

That's my thinking too. A $649 water cooled version, a $599 air cooled and ever-so-slightly cut down version, and something (small Vega?) at $349 and $499. I'm just hoping the design is better balanced than Fiji so that scaling between parts is more or less linear along price/CU count.

 

[lobz]

That is a logical impossibility. Theoretical maximum compute flops is calculated based in the number of CUs and the speed at which they are run. As such, of course it's dependent on IPC.

Don't blame the messenger. I didn't come up with how the industry calculates the maximal theoretical computing flops of a processing unit. Also your 2 sentences are totally contradictory.