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

The answers weren't especially good. But then neither were the questions.

Ultimately I'm not surprised that Raja stuck to the corporate messaging here (least Legal lynch him). It's a hard line to walk in not promising that RX Vega will be out in H1, while at the same time keeping potential customers fully interested and engaged in the eventual launch. Just like in relationships, no one likes to come second.:p
 
It would not need to be a dual GPU though, Frontier performance is perfect for it.
I guess it is probable a dual GPU will be available down the line but when; I would say MI25 is more important than that as a chassis node-cluster solution and larger contracts-projects.
Regarding exclusivity; yeah it could had been a special order contract (like the version Nvidia has I think for Volta and 12n), but I cannot imagine how much AMD would need to pay SK Hynix for them to hold back also 1.6Gbps as they will be losing out on a lot of potential and playing catchup in lucrative areas such as network processors/backplanes/etc that Samsung just recently seems to be able to take a step into.
Cheers
It wouldn't need to, but in theory they could do the same thing as Naples with Vega. Ideal clocks on these things are probably below 1GHz for best perf/watt. At that point they need density beyond what all the PCIE lanes of Naples would provide. Nothing saying they have to, but could be an interesting possibility if entirely separate parts or the traditional dual GPU didn't work sufficiently.
 
The answers weren't especially good. But then neither were the questions.

Ultimately I'm not surprised that Raja stuck to the corporate messaging here (least Legal lynch him). It's a hard line to walk in not promising that RX Vega will be out in H1, while at the same time keeping potential customers fully interested and engaged in the eventual launch. Just like in relationships, no one likes to come second.:p

We at least got a confirmation that Vega FE has 2x8-pin power connectors and Raja's physical card was just an ES card.

iPtRDEA.png


Also, shots fired. I hope the Volta-digging doesn't come back to haunt AMD...

wJW68dy.png


EDIT Whoops, Simby already mentioned the power connector thing an hour ago.
 
Last edited:
So he mentions Vega doing better with geometry throughput without any developer interactions. The performance in the leaked time spy and firestrike tests doesn't look any better than what an overclocked Fury X would get though.

In fact, the LN2 run of a Fury card at 1.4Ghz with 1Ghz of HBM performed around 10% better than the 1080 which looks to be better than what Vega would get at 1.4Ghz from those leaked results. So the bandwidth or improved latency of memory makes a significant difference in firestrike or the sample was not running at the 1.2Ghz boost clocks.

The same sample had been shown earlier in the year equalling or besting the 1080, albeit in AMD favoring games, so that's another incongruency as well.
 
ES are probably running ES drivers, who knows what features were even enabled. Maybe it was even with debugging overhead. Who knows, there really isn't a point in looking at pre release benchmarks when drivers makes all the difference.
 
Yeah I forgot to mention the drivers, the sample perhaps was being tested mainly for compute applications and someone decided to give graphics a go. April seemed a little too late for driver optimizations but now that it's releasing after the pro. card, not that late.

Edit :

Q: Many argue that vega is just a refined polaris gpu, how would you respond to this ?
A: My software team wishes this was true:)
Vega is both a new GPU architecture and also completely new SOC architecture. It's our first InfinityFabric GPU as well
 
Last edited:
Some tidbits I've been getting from the AMA:

Was the card used in the gaming demos a Frontier Edition? And if so, was it the water cooled one or the air cooled version?
A: It was an air-cooled version

Is there a difference in performance/clock speed between the water and the air cooled version or is one just quieter/cooler?
A: There will be a slight difference in clock speeds, and therefore performance as well.


What I get from here is that the card used to get the results shown in the presentation was the air-cooled version, but the closed-loop card will be getting slightly higher performance.


Raja also keeps referring to Vega as a "SoC Architecture". I doubt there's an actual applications processor inside the chip, so I wonder if them calling SoC has to be with the chip being able to handle external storage directly and/or being able to access other I/Os.
 
They've used "SOC" term before too in this context, it just refers to the GPU chip being built from "separate IP blocks"
 
I think it means
I'm guessing it's because they modularized the GPU components, so they can easily build new SOCs with Vega technology. We already know Raven Ridge will feature Vega cores.

Cheers
I think its also probably, likely things like the command processor have had to change significantly to match the changes in ALU and ROP's.
 
My first real chip, 20 years ago, was called an SoC because it used a generic CPU to do some lightweight DSP processing and link layer management. The final CPU code was burned in a 32KB ROM.

There's no strict definition of SoC. As soon as a chip has some kind of CPU in addition to the regular application specific logic, it's called an SoC.
 
Back
Top