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

For all the hardware limitations on the first outing in Vega, wouldn't it have made sense to keep a semblance of an implementation if they knew future iterations would give better support considering how the played the features up? At least to give developers a chance to familiarise themselves and program with them in mind? Development takes time. Getting tools and support out early should be a boon. Unless the hardware turned out so broken as to be incompatible even on refined silicon like Navi assumedly is. Which begs the question as to why they even played up the features to begin with at the Vega launch? Something so obviously broken ought to have gone unmentioned.

Its easy to second guess AMD from an arm-chair position of course. But it does seems like setting themselves up for failure from a marketing perspective no? Both towards devs as well as to consumers.
 
Whatever you can cull before it gets fed into the geometry pipeline does not add to it's load.
 
Maybe there's even more Vega parts yet to come:

PCWatch: Instinct is based on 7nm, so how about changing to the Navi architecture and performance?

Forrest Norrod: There's going to be some overlap between the two. I think Lisa eluded to this earlier, where GCN and Vega will stick around for some parts and some applications, but Navi is really our new gaming architecture so I don't want to go beyond that. You'll see us have parts for both gaming applications and non-gaming applications.


https://www.anandtech.com/show/14568/an-interview-with-amds-forrest-norrod-naples-rome-milan-genoa
 
What would the use case be? A portable MI solution perhaps? It would be kind of cool to see a Zen 2 APU chiplet plus HBM chiplet. Heck, that would make for an incredibly powerful APU solution.
The current 15" MacBook Pro has Vega Mobile as a BTO option.

Apple is rumored to release a 16"–16.5" MacBook Pro around September of this year. Since the 15" MacBook Pro was updated with 8 core CPUs last month, the 16" (if it is real) is not likely to outright replace the 15" at first. I expect the 16" MBP to start at a higher price than either of the base models for the 15" MBP (the general speculation is $2999 or higher, at least among those who believe the rumors).

I think it would be great if the 16" MBP has a ~24 CU Navi with HBM2. It would be another factor that differentiates the 16" from the 15", as back when Apple had 15" and 17" laptops I thought that the 17" was too similar to the high end 15" (other than the display size/resolution). Unlike the 15" which is available with either a GDDR5 GPU or a HBM2 GPU, the 16" could be positioned high enough so that it only uses a HBM2 GPU, so presumably Apple can use a slightly smaller logic board than otherwise?
 
The current 15" MacBook Pro has Vega Mobile as a BTO option.

Apple is rumored to release a 16"–16.5" MacBook Pro around September of this year. Since the 15" MacBook Pro was updated with 8 core CPUs last month, the 16" (if it is real) is not likely to outright replace the 15" at first. I expect the 16" MBP to start at a higher price than either of the base models for the 15" MBP (the general speculation is $2999 or higher, at least among those who believe the rumors).

I think it would be great if the 16" MBP has a ~24 CU Navi with HBM2. It would be another factor that differentiates the 16" from the 15", as back when Apple had 15" and 17" laptops I thought that the 17" was too similar to the high end 15" (other than the display size/resolution). Unlike the 15" which is available with either a GDDR5 GPU or a HBM2 GPU, the 16" could be positioned high enough so that it only uses a HBM2 GPU, so presumably Apple can use a slightly smaller logic board than otherwise?
That's only possible if Vega 12 is in fact 24 CU part, sadly I'm not aware of any die shots, heck, even die size to do estimations from
 
LLVM changes for a GFX9.08 were noted here: https://www.phoronix.com/scan.php?page=news_item&px=AMD-GFX908-LLVM-Activity
There are also a number of recent LLVM commits not mentioned there that are related.

The list of dot instructions that started with MI50 and continued into Vega 12/14 had an odd gap where MI50 supported dot1 and dot2, and the Navi variants supported dot5 and dot6.
GFX908 appears to cover the whole list from dot1 through dot6.
Additionally, there's a new class of wide vector registers (AGPR) and MFMA instructions. I think there was a reference to the A in AGPR meaning accelerator, possibly. The MFMA instructions appear to mean Matrix FMA, with various combinations matrix size. There's a few that line up with Nvidia's 4x4x4 tensor units, and a large range of other size combinations including 16x16x16(F16) or even 32x32x8(F16).

There seems to be some relatively free back and forth movement between VGPRs and AGPRs (the latter appears to be larger with moves considered multiples of the other), although the hardware execution seems to be mostly independent.
Some of the comments concerning power management indicate that while vector operations can be issued while the new longer-latency operations are running, it's recommended to not do so because of an increase in power consumption and the risk of throttling, and there's a line referencing disabling some of the GCN power scheduling out of concern of the MAI instructions spiking power.

This does have SRAM ECC, and there's a line referencing the name MI-100.
 
It's supposedly pure compute chip with anything related to 3D ripped out

I think that's unlikely. Do you have a source for that?
If they ripped out all 3D hardware (ROPs? TMUs?) then it probably wouldn't be GFX9 ISA compatible.
 
Back
Top