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

I used a Vega 64 on a Mac with the eGPU thingy, didn't perform half bad, seems like AMD is on point with their drivers for the iMac Pros

 

Just send a (broken) Vega to this fellow. He has already covered Polaris and Fiji. Also it seems he keeps improving his technique - Zen.

 

The people complaining about RX VEGA's unused features, somehow won't be complaining about them, when later this year Volta ends up supporting those game features too, right..? (Not really hard to see what a 7nm Vega w/2.4ghz HBM2 will do, either.)

Why be political about it ?


Many AAA Developers have said that as Windows10 platform matures, they are making use of every technique offered. Lets not forget that the Xbox One X is part of the radeon ecosystem.

 

Exactly which unused Vega features that are also supported but unused by Volta did you have in mind?

Reread your first sentence. Yes, the one with a heave dose of whataboutism. Do you notice the irony?

 

I'm sure there are other good Vega 10 die shots in the wild by now, but in case it's helpful, here is a version of that PDF with the obscuring logo removed and below is the image file for that page 2 die shot. It only appears to be 1091x818, but maybe that's still useful.

 

As an aside, it appears that the PS4 Pro is next up. The current shot is pretty blurry, although presumably a more straight-on shot is on its way. It seems like it was more difficult to etch down cleanly with a chip on this process (comments indicating GV100 is a goal?).

Perhaps a more clear shot would be of interest in the console forums. It's an interesting example of the layout games that can be played. There are at least three CU variations from what I can see so far. There's been a bit of speculation as to the ROP count for the Pro, although I don't know if this shot would be sufficient to settle it given the willingness to vary blocks for area or other reasons, and some custom hardware. There are some borrowed elements from Polaris and Vega somewhere in there.
The chip.

 

Hay, gusy...

In GPU-Z, you have a "SOC Clock" graph for Vega GPUs; what exactly is this, is it known? I'm thinking like, infinity fabric-related, but I'm not sure - I have no way of knowing after all, just being some guy on the internets.

Most of the time it sits at just over 1GHz (1028MHz) on one of my boards, and a hair over 1.1GHz (1107MHz) on the other, while running constant Folding@Home (because of the winter chill ), but I've witnessed small variations. That is with HBM running at 1GHz btw, in case it matters...

 

Should be Infinity, but AMD never clarified what they were doing with it. Somehow overbuilt for server and only connecting HBM to the core currently. Best guesses are extra PCIe lanes, pro duo, or Ripper APU with Snowy Owl. It should have been clocked slightly higher than the memory as I recall.

 

It seems to have some relation to core clock. If core is >1500MHz, SOC clock is 1100MHz. If core is <1500, SOC clock drops down to 1028, sometimes less. Memory clock = 1000, at all times.

 

Oh, but they did. Check page 10:
https://www.hotchips.org/wp-content...b/HC29.21.120-Radeon-Vega10-Mantor-AMD-f1.pdf

 

Has anyone re-done those tests that showed worse performance on Vega than on Fiji? "Now that the drivers have improved"...

 

Is external memory bandwidth a reliable measure of performance considering that there are differences between the architecture's internal bandwidths (does that question make any sense? )

 

IIRC that has been remedied. Was a bit buried during the RX Vega launch test, but there:
http://www.pcgameshardware.de/Radeo...3/Tests/Benchmark-Preis-Release-1235445/3/#a3

 

Yes, that page shows what AMD is doing with infinity fabric, but it doesn't nail down that GPU-Z's "SOC clock" is infinity fabric clock.

 

From the rest of Anarchist4000's posting it seemed pretty clear to me, that with „overbuilt for server“ etc., it was pretty clear that he did not mean the GPU-z value when he said „but AMD never clarified what they were doing with it.“.

 

For those not familiar with this Vega presentation, it contains interesting benchmark showing gaming DBSR on/off perf. and power gains:

The average gains are not anywhere near "the unicorn-level".

This makes one think about the Vega project goals. Given Fiji vs Vega10 clock-to-clock being the same within an error margin, they must have ramped-up the frequency. Vega10@1.3GHz surely has a nice power profile. The current Vega10@1.6GHz not so much... But hey, they had to beat their previous (28nm, pioneering and bottlenecked) flagship by a meaningful margin. What were they thinking?

 

Thats not really the way vega operates, its more around how you "ask" it to consume power. Its "average" clock is very dynamic. but the difference in clock rate between ~175watt and 275watt normally isn't many mhz's but that can occur at any clock range depending on game/location.

my "default" vega OC is
-1% power
P6 1562mhz, 975mv
P7 1650mhz, 1075mv
memory 1075mhz , 925mv

on my card this increases performance, but the difference between this and changing power to +50 and P7 to high 1700's + more mv is normally not much more then 100 mhz.

edit: just did a quick check in superposition difference between -1 and + 50 with the above settings and it is about 100mhz (1500 vs 1600) and about 80 watts ( 275 vs 190)

 

Nowhere in there could I find an explanation for Raja saying Infinity is larger than necessary for gaming devices though. The implication would be it connects something that only shows up in server or compute more robustly. SSG, Infinity mesh for MI25 with Epyc, etc as shown in that presentation.

Well if the software side still isn't fully hashed out, I doubt anyone would reasonably expect to measure huge gains in actual testing when those measurements were taken in 2016 I believe.

 

Isn't it established that IF is the internal interconnect in vega 10, rather than the old system of dedicated buses/crossbars. So it'd be used for all kinds of internal data transfers, not just server related stuff. Which would be funky btw, since there is hardly any vegas used that way right now...

 

Yeah, but the question is what unknown capability made it larger than necessary for gaming, worth mentioning by Raja, but applicable only to server as different chips weren't an option? It would seem to be providing an interconnect for some pro feature or routing traffic differently for a standard chip in compute configuration.

Virtualization, storage/SAN adapters, APU, or perhaps ECC/RAID would seem the only apparent pro features that could significantly scale the network.