AMD Vega 10, Vega 11, Vega 12 and Vega 20 Rumors and Discussion

[Deleted member 13524]

The GPUBoss lists both M395X Mac and M395X TDPs at 250W which are obviously both wrong. We know that PC-version of M395X has 125W TDP.

Where does this GPU Boss website even take their clock values from? They don't show any references to anything.

AFAIK there's hasn't ever been any Windows PC (desktop, laptop or AiO) with a M395X. Those "PC" scores are probably just the 2015 iMac booting windows. The GFXBench comparison that popped up in this thread literally says "bootcamp edition" in the description of the so-called "PC version".

Tonga Pro is about 15% above 1050 Ti in our gaming-only parcours, with application performance included (OpenCL), is is around 20%. Tonga XT possibly in the 20-25 (25-30-ish including OpenCL) percent range.

Though the M395X is clocked ~6.3% lower than the 970MHz Tonga XT, so maybe I would shave some 5% off from those predictions.
Regardless, the question still stands if the ">3x faster" statement refers to the 56CU+8GB Vega baseline or the 64CU+16GB version that will surely be available only as an upgrade.

It boggles the mind.... does not help their credibility (WCCFTech not AMD).

OMG the guys at WCCFTech actually thought that picture made up of pretty green rectangles was an actual die shot.
And finding out it wasn't served as basis for making up a FUD-full article. Wow..

 

[Kaotik]

OMG the guys at WCCFTech actually thought that picture made up of pretty green rectangles was an actual die shot.
And finding out it wasn't served as basis for making up a FUD-full article. Wow..

And they completely missed the fact that there is actual dieshot underneath as the base for the artist, just like with Polaris 10 (in case of Vega it's visible in some areas not colored in)

 

[3dilettante]

And they completely missed the fact that there is actual dieshot underneath as the base for the artist, just like with Polaris 10 (in case of Vega it's visible in some areas not colored in)

Maybe? It's more compression artifacts than anything, and many regions are so regular that they could be tiling a few "representative" blocks.
I'm not fan of fake die shots, but it's odd that wccftech needed a tweet in 2017 to figure out what most GPUs have had done for them by marketing for several generations.

Either that, or it stood out because AMD didn't use highly similar artistic choices to Nvidia like it did with the Polaris fake shot.

 

[Kaotik]

Maybe? It's more compression artifacts than anything, and many regions are so regular that they could be tiling a few "representative" blocks.

I don't think you can put half of the right side of the chip as "compression artifacts" really

 

[3dilettante]

I don't think you can put half of the right side of the chip as "compression artifacts" really

If you mean the strip on the right edge, that's one of the few areas that wasn't overwritten with colored shapes. Being that far off to the side, and without IO or being close the CUs or memory, it might be miscellaneous logic.

The blocky background behind the shapes for the CUs, command processors, and everything that might be interesting is blocky and suspiciously grid-like. The stuff under the command processors and geometry sections doesn't really line up with what's on top of it, and it doesn't change at all when under the shader engine front ends and the command processor block, which I doubt is correct.

 

[Silent_Buddha]

Regardless, the question still stands if the ">3x faster" statement refers to the 56CU+8GB Vega baseline or the 64CU+16GB version that will surely be available only as an upgrade.

It has to apply to the lowest version of Vega available unless there is annotation connected to it at the bottom of the page that states otherwise.

If it applied to only the higher end version, Apple would be opening themselves up to a potential lawsuit about false advertising if someone bought the lower end Vega and it was less than 3x the performance of the reference model.

Regards,
SB

 

[Deleted member 13524]

Vega Frontier is the highest clocked Vega coming out (or most compute performance) that is straight from Raja's mouth.

Source for this "Frontier is highest clocked Vega straight from Raja's mouth" statement, please?

What I do know is this statement from a reddit AMA:

Consumer RX will be much better optimized for all the top gaming titles and flavors of RX Vega will actually be faster than Frontier version!

You might be confused by this other statement:

On the compute side of things..Vega FE will be the fastest single GPU solution (>12.5 TFlops FP32) when it's available and our NCU packs several additional optimizations, including Rapid-Packed-Math which delivers >25 TFLops of FP16

What he means here is at time of release, the Frontier Edition will be the fastest single GPU solution regarding FP32 throughput. Which is true, because the higher-clocked RX Vega will only release afterwards.

It has to apply to the lowest version of Vega available unless there is annotation connected to it at the bottom of the page that states otherwise.

If it applied to only the higher end version, Apple would be opening themselves up to a potential lawsuit about false advertising if someone bought the lower end Vega and it was less than 3x the performance of the reference model.

This is my interpretation as well.
Companies can be wildly creative when they use the term "up to X times faster", but they must be very careful if they say "over X times faster". Failing to sustain the first is perhaps deceptive/imoral advertising, failing the second is just false advertising.

I doubt there's even a single word in apple's website that doesn't go through heavy vetting by their legal department.

 

[Razor1]

Consumer RX will be much better optimized for all the top gaming titles

What do you optimize? Cards, or drivers?

First part of that sentence is very clear, because of driver optimizations, RX vega will be faster in computer games than the Frontier edition.

 

[Kaotik]

What do you optimize? Cards, or drivers?

First part of that sentence is very clear, because of driver optimizations, RX vega will be faster in computer games than the Frontier edition.

The wording suggests that some RX-model will be faster than Frontier Edition period, not just in games.
I suspect the "max clock 1600" we're seeing will actually be RX's max clocks, as the performance figures AMD gave for Frontier Edition suggest about 1550 MHz, with 1600 MHz they would have said 26 TFLOPS FP16 instead of 25.

 

[Kaotik]

If you mean the strip on the right edge, that's one of the few areas that wasn't overwritten with colored shapes. Being that far off to the side, and without IO or being close the CUs or memory, it might be miscellaneous logic.

The blocky background behind the shapes for the CUs, command processors, and everything that might be interesting is blocky and suspiciously grid-like. The stuff under the command processors and geometry sections doesn't really line up with what's on top of it, and it doesn't change at all when under the shader engine front ends and the command processor block, which I doubt is correct.

I'm not suggesting there's anything to read about the dieshot visible, just that it is there

 

[Razor1]

The wording suggests that some RX-model will be faster than Frontier Edition period, not just in games.
I suspect the "max clock 1600" we're seeing will actually be RX's max clocks, as the performance figures AMD gave for Frontier Edition suggest about 1550 MHz, with 1600 MHz they would have said 26 TFLOPS FP16 instead of 25.

Just have to disagree with ya, cause we have seen how AMD talked in the past about clocks and performance, this time, it looks like Raja didn't want to go down that path and was explicit when showing Vega and talking about it. He went out of his way to say certain things in a certain manner. The feeling he gave when taking about Vega and the way he talked about it in the AMD is competently different than previous product launch. It actual came out as "realistic" expectations, instead of over hyping.

 

[Kaotik]

Just have to disagree with ya, cause we have seen how AMD talked in the past about clocks and performance, this time, it looks like Raja didn't want to go down that path and was explicit when showing Vega and talking about it. He went out of his way to say certain things in a certain manner. The feeling he gave when taking about Vega and the way he talked about it in the AMD is competently different than previous product launch. It actual came out as "realistic" expectations, instead of over hyping.

Well, we'll seen soon enough if the 1600 is for watercooled Frontier Edition or not - if it's not, as the TFLOPS' suggest, then it pretty much has to be for RX

 

[3dilettante]

I'm not suggesting there's anything to read about the dieshot visible, just that it is there

To cover all the bases, while there may be something in that section to the side that isn't filler, whether that's real is to be determined.
The "background" of the Polaris marketing pic has very little that conforms to the actual die.

 

[Anarchist4000]

What do you optimize? Cards, or drivers?

First part of that sentence is very clear, because of driver optimizations, RX vega will be faster in computer games than the Frontier edition.

With how little dependency they have on drivers it could be the card. ACE/HWS can be reprogrammed, instruction sets scheduled differently, raw clocks for cooling, etc. No reason the card couldn't use consumer drivers if there is even a difference. All the linux drivers are a unified stack for compute and graphics that should have been sharing the Windows codebase. AMD has been slow to merge DAL because the abstraction layer is against Linux standards.

The wording suggests that some RX-model will be faster than Frontier Edition period, not just in games.
I suspect the "max clock 1600" we're seeing will actually be RX's max clocks, as the performance figures AMD gave for Frontier Edition suggest about 1550 MHz, with 1600 MHz they would have said 26 TFLOPS FP16 instead of 25.

The 25 markets more easily as it produces even numbers in a rack configuration.

 

[lanek]

It boggles the mind.... does not help their credibility (WCCFTech not AMD).

http://wccftech.com/amd-radeon-rx-vega-die-shot-fake/

At the risk, to be a bit harsh, do they have still any credibility left ? ... their comment section is a trolling jungle without any inteligent comments ( let alone no moderation at all ), and their line is: publish as many content you can (for clickbait), without checking any fact, and feed the comment section. It is a bit sad, as i think they could do way better with what they have in hand.

( for be a little bit nice, i think somewhat, they are doing, trying a bit better than let say 2 years ago )...
.........................

Back to the Apple "marketing term contest"..

The last imac have something like around 3.95 TFlops of FP32 compute, we know that FE is at around 12Tflops.. .. they write over 3x more performance.. dont need to read between the lines and start conjecture about metrics over it..

 

[ieldra]

At the risk, to be a bit harsh, do they have still any credibility left ? ... their comment section is a trolling jungle without any inteligent comments ( let alone no moderation at all ), and their line is: publish as many content you can (for clickbait), without checking any fact, and feed the comment section. It is a bit sad, as i think they could do way better with what they have in hand.

( for be a little bit nice, i think somewhat, they are doing, trying a bit better than let say 2 years ago )...
.........................

Back to the Apple "marketing term contest"..

The last imac have something like around 3.95 TFlops of FP32 compute, we know that FE is at around 12Tflops.. .. they write over 3x more performance.. dont need to read between the lines and start conjecture about metrics over it..

Of course they are talking about the compute throughput, however on the slides I saw 11tflops. That is not over 3x. That's slightly under, perhaps 11tflops was the cut die ?

What was the TDP of the 395x. TPU says 250w TDP, other websites says 125W. What is it ?

 

[lanek]

Of course they are talking about the compute throughput, however on the slides I saw 11tflops. That is not over 3x. That's slightly under, perhaps 11tflops was the cut die ?

What was the TDP of the 395x. TPU says 250w TDP, other websites says 125W. What is it ?

Look the timing, this will be really hard to read between the llines of marketing words... AMD could have reveal that the ImaCPro will use Vega during computex, tthey have not do it, Apple could have exhaustive information provided ( they have not do it ).... theres a lot of commercial aggrement between them.. and ImacPro will not be comercialized before the end of the year ....

We know the 395x compute power and we know that FE, instinct have around 12Tflops ,,,, why will you read something between this ? It is an all in one desktop...

I really doubt that Apple will use the max Vega can give in this case.or their design will look as the MacPro ( who is a dead market for Apple anyway, hence why they transfer it to the iMac system ( no more stand alone )..

 

[Kaotik]

What was the TDP of the 395x. TPU says 250w TDP, other websites says 125W. What is it ?

125W. Even desktop Tonga XT's TDP is under 200W

 

[ieldra]

125W. Even desktop Tonga XT's TDP is under 200W

So it is unreasonable to assume they will equip this new iMac with a 250w card. I imagine an mxm card (tonga) with 8gb memory will have limited pcb space for beefy vrm, at the expense of efficiency, compounded with inevitably high temperatures I'd expect this to clock far lower than desktop tonga, so 930 mhz seems unlikely. TPU page states clocks of 730 mhz or something

 

[Deleted member 13524]

So it is unreasonable to assume they will equip this new iMac with a 250w card.

The new imac will get Polaris cards. Vega is going into imac pro only.

I imagine an mxm card (tonga) with 8gb memory will have limited pcb space for beefy vrm, at the expense of efficiency, compounded with inevitably high temperatures I'd expect this to clock far lower than desktop tonga, so 930 mhz seems unlikely. TPU page states clocks of 730 mhz or something

Notebookcheck and wikipedia claim 909MHz on the M395X, though that's probably boost clock values.