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

Yup, it's more realistic to think that 2/3 of the "staff" might be working on console related (or just all semi-custom) projects. Not just on a Sony project. And even that seems too high.

Regards,
SB
 
Or, it could be simply that the RTG seem to consistently miss milestones/make wrong guesses under the now-departed leadership. The bet on importance of HMB to high-end performance has backfired, due to either technical issues with the development, unexpected progress with the conventional memory, or both and cost them substantially. Coming off Hawaii, which was an extremely competitive high-end part that was inherited and following it up with Fiji and Vega probably gave the top company leadership some food for thought as far as the direction of the group that was demanding more and more autonomy with questionable results to show for it. The sequence of events that had Raja being left out of the executive bonus pool, followed by his extended leave of absence, change in role, and ultimately a departure to Intel, of all places, seems pretty telling as far as some of the internal dynamic is concerned.
 
Fired vs "Well, we'll take 2/3 of your staff to work on Sony-Navi".

It was really weird, I hope someday somebody will talk about the whole Vega situation (even if I don't think it will happen, I'm still waiting the R600 explanation...)

R600 situation was pretty thoroughly explained with RV670...
 
Seems awfully naive to me (or maybe convenient) to ignore AMD's anemic R&D expenditure in the ~3 years that preceded the release of Polaris + Zen + Vega, and the obvious impact it had on RTG's execution.

Most of us will agree that AMD's (specifically RTG's) marketing was terrible in the unnecessary hype they couldn't live up to (e.g. saying Polaris was coming "several months ahead" of nVidia's first FinFet GPUs, Vega being awfully late, saying Vega would clock above 1.7GHz, forcing Vega 56 down reviewers' throats at the last minute, and many others).
Regardless, AMD's general execution given the ~$240M/quarter R&D expenditure was great. Otherwise their shares obviously wouldn't have risen over 1500% during the last 2 years.





As for Sony's influence on AMD's GPU architectures.. there's some evidence that Sony's ICE has been calling the shots since AMD won the original PS4 .
The Liverpool GPU preceded AMD's own dGPUs in adopting 8 ACEs for improved Async performance and the TrueAudio DSPs. Neo's GPU preceded AMD's dGPUs and SoC iGPUs in adopting Rapid Packed Math.
And now there are sources/rumors claiming Navi is being driven by Sony again.


Sony doesn't compete in the PC space and neither do they want to, but they have always had a strong grasp over the architecture that went into their consoles (unlike Microsoft). The only time they didn't was with the PS3 but reportedly that was dual-Cell (or was it Cell-based GPU?) not performing as intended and RSX being a late adoption that didn't go well.

So given AMD's lack of resources for GPU development, it just seems logical to me that Sony's relationship with AMD is one where they don't compete in the same space and contribute towards the AMD's general GPU arch. Which AMD can then use in their own products but can't use in semi-customs for Sony's competition within a certain time period - which would explain why Scorpio doesn't have RPM CUs, for example.
 
Seems awfully naive to me (or maybe convenient) to ignore AMD's anemic R&D expenditure in the ~3 years that preceded the release of Polaris + Zen + Vega, and the obvious impact it had on RTG's execution.

Most of us will agree that AMD's (specifically RTG's) marketing was terrible in the unnecessary hype they couldn't live up to (e.g. saying Polaris was coming "several months ahead" of nVidia's first FinFet GPUs, Vega being awfully late, saying Vega would clock above 1.7GHz, forcing Vega 56 down reviewers' throats at the last minute, and many others).
Regardless, AMD's general execution given the ~$240M/quarter R&D expenditure was great. Otherwise their shares obviously wouldn't have risen over 1500% during the last 2 years.





As for Sony's influence on AMD's GPU architectures.. there's some evidence that Sony's ICE has been calling the shots since AMD won the original PS4 .
The Liverpool GPU preceded AMD's own dGPUs in adopting 8 ACEs for improved Async performance and the TrueAudio DSPs. Neo's GPU preceded AMD's dGPUs and SoC iGPUs in adopting Rapid Packed Math.
And now there are sources/rumors claiming Navi is being driven by Sony again.


Sony doesn't compete in the PC space and neither do they want to, but they have always had a strong grasp over the architecture that went into their consoles (unlike Microsoft). The only time they didn't was with the PS3 but reportedly that was dual-Cell (or was it Cell-based GPU?) not performing as intended and RSX being a late adoption that didn't go well.

So given AMD's lack of resources for GPU development, it just seems logical to me that Sony's relationship with AMD is one where they don't compete in the same space and contribute towards the AMD's general GPU arch. Which AMD can then use in their own products but can't use in semi-customs for Sony's competition within a certain time period - which would explain why Scorpio doesn't have RPM CUs, for example.

https://forums.anandtech.com/thread...om-vega-forbes-article.2548845/#post-39463284
 
Ohh my. This is the worst rumor ever, and people still keep spinning it? Stuff like this is designed in stages, and never ever everyone is working on the same thing at the same time. One team starts, hands off the stuff to the next team, and then starts with the next product. In time everyone will probably have had their hands on a specific die, but never everyone at the same time.
 
It's been rumored before that a Vega refresh was coming. I still don't believe it but if it does and it's at 1080ti or better performance it must be at a decent price. I don't see why it wouldn't be something to hold the market over until Navi comes for gaming.

The crux of it, if this is all true, is a decent price. I've read a few posts in different forums of folk either keeping their 1080ti, getting a second one or upgrading to a 1080ti.
This is not a scientific barometer of how well Turing will sell. I'm just pointing it out to say that performance in that bracket can stll be attractive (at a good price).
 
As for Sony's influence on AMD's GPU architectures.. there's some evidence that Sony's ICE has been calling the shots since AMD won the original PS4 .
The Liverpool GPU preceded AMD's own dGPUs in adopting 8 ACEs for improved Async performance and the TrueAudio DSPs. Neo's GPU preceded AMD's dGPUs and SoC iGPUs in adopting Rapid Packed Math.
And now there are sources/rumors claiming Navi is being driven by Sony again.
What is the metric for deciding if the PS4's GPU preceded discrete? Bonaire was launched eariler and had the TrueAudio block on-die. Microsoft's audio block was similar in that it was a block of Tensilica DSPs with custom tweaks, so the concept of having a section of the chip composed of third-party DSP hardware with optional tweaks appears common. The less specialized and less-used DSP block on AMD's seems like it was consistent with being the basic offering AMD would make to clients before they made changes.

From Mark Cerny, the collaboration on the PS4 Pro made it sound like RPM was a roadmap item AMD offered to Sony, and Sony opted for it, rather than Sony introducing the concept into AMD's IP pool.
https://www.eurogamer.net/articles/...tation-4-pro-how-sony-made-a-4k-games-machine

The 8 ACEs in Hawaii may have been a Sony influence, although anything from Fiji onward becomes less certain. The 8 ACE blocks in AMD marketing resolved into 4 ACEs with dispatch units and 2 HWS blocks (which to my fragmented recollection may not be 2 separate HWS blocks silicon). Fiji apparently had initial silicon for HWS, and after Polaris AMD was more clearly diverged from the older compute front end.

From the Pro, the ID buffer and tweaks for checkerboard rendering were more clearly described by Sony as being customizations. That has not clearly shown in later discrete products. Perhaps the ID buffer has some analogous portions in the binning rasterizer, as it would need to track at least a subset of batched primitives by an ordered ID. The amount of work integrating DSBR into the pipeline and the age of some of the patents from AMD may put it before the ID buffer, where the primitive ID storage was exposed and automatically written out, while the rest of the functionality was not added to the console.

On the other console line, Microsoft's audio block was more extensively customized (not that AMD necessarily cared to follow what either vendor did for its discarded DSPs). The command processor tweaks disclosed for the Xbox are more in line with the various HSA and response queue/reservation retrofits to GPU firmware than what has been mentioned for the PS4. In the case of command processor "tweaks", it sounds likely that it's a case of leveraging an AMD capability that was freely available internally, given the flexibility of microcode engine updates--at least for GPUs with sufficient storage for the updates.

There were various other tweaks both console makers described as being selected from AMD's roadmap of later GPUs, such as DCC, primitive discard tweaks, and work distributor.

This seems more on equal footing so far. If there were some kind of disproportionate level of investment into Sony specifically, it may mean more questions as to why there isn't a similar level of demand from Microsoft.
 
That^ entire video is on point and covers my exact thoughts on AMD's direction.

I already predicted as much this past week. Chen already said a year ago, that VEGA's HBCC can already use any type of memory or storage type. (GDDR5/6/HBM/DDR4/SSD/etc..) And given AMD's 7nm leverage & all the plausible rumors... I can easily see a 7nm Radeon V96 & V128 @ $550 bucks, out for Holiday shopping season. With the average GPU performance 8~12% higher than the RTX2080 and using less power. (win/win)


AMD will have 7 months reign in sales, before Nvidia can respond with their 7nm chips in Aug next year...
 
That^ entire video is on point and covers my exact thoughts on AMD's direction.

I already predicted as much this past week. Chen already said a year ago, that VEGA's HBCC can already use any type of memory or storage type. (GDDR5/6/HBM/DDR4/SSD/etc..) And given AMD's 7nm leverage & all the plausible rumors... I can easily see a 7nm Radeon V96 & V128 @ $550 bucks, out for Holiday shopping season. With the average GPU performance 8~12% higher than the RTX2080 and using less power. (win/win)


AMD will have 7 months reign in sales, before Nvidia can respond with their 7nm chips in Aug next year...

Sorry, not happening. There will not be a gaming version of Vega 20.
 
Sorry, not happening. There will not be a gaming version of Vega 20.

Well that begs the question, what exactly is "Vega20" ..?
When you know for certain and can tell me that, then I'll start to believe your claimless predictions. I said 7nm Vega for gaming...


And I bet it happens before 2018 ends.
 
It's really a stretch to believe that AMD would already sell 7nm gaming Vega in 2H 2018 for 550 Dollars with more than 64 CUs and in the video there is even speculation that AMD would change the memory interface to GDDR6 because HBM2 is so expensive.
That's all a bit too dreamy isn't it?
 
Well that begs the question, what exactly is "Vega20" ..?
When you know for certain and can tell me that, then I'll start to believe your claimless predictions. I said 7nm Vega for gaming...


And I bet it happens before 2018 ends.

and there won't be 7nm Vega for gaming this year either. TSMC 7nm production will be occupied by A12, Kirin 980, and Snapdragon 855, then later Zen 2 silicons. Reason why Vega 20 is going to be released on limited quantity later this year (which may or may not happen).
 
That^ entire video is on point and covers my exact thoughts on AMD's direction.

I already predicted as much this past week. Chen already said a year ago, that VEGA's HBCC can already use any type of memory or storage type. (GDDR5/6/HBM/DDR4/SSD/etc..) And given AMD's 7nm leverage & all the plausible rumors... I can easily see a 7nm Radeon V96 & V128 @ $550 bucks, out for Holiday shopping season. With the average GPU performance 8~12% higher than the RTX2080 and using less power. (win/win)


AMD will have 7 months reign in sales, before Nvidia can respond with their 7nm chips in Aug next year...
You better be right, because if not, and AMD does not have a clear winner at hand with both the advantages of advanced process technology and High Bandwidth memory against Turing, then things will look to start really bleak once Nvidia also makes this shift.

That said, some predictions in videos and websites seem a bit optimistic. AMD has stated that it would require major rework for the command processor to scale out beyond 4 shader engines. We have not seen shader engines with more than 16 CUs. In addition to that, the parts with higher CU numbers per SE tended to scale less than optimal for adding more units.

Sure, the HBCC won't care how it manages it's internal tags and where it maps its data chunks to, but I guess the memory controllers will and need to be redesigned or rather switched. A certain amount of transistors will have to go into doubling the number of memory channels and probably into increasing the L2 cache as well. Then there's half-rate DP (not thinking they'd go full rate), which is going to cost as well both in terms of transistor count as well as in terms of energy because of the increased size of the adders and multipliers.

As for manufacturing tech, in my book, you will have the density advantage going to 7nm, but i am not sure about it being coupled with either the quoted perf increase or the quoted efficiency. You sure could build a chip that's twice as many transistors, and you could build one, that's half as big. You also could build one being 35% faster and you could build one, that's twice as efficient. But when you design your chip, you try to strike a balance out of this triangle that gives you the most benefits for your projected market. Remember, in Vega, a lot of the extra transistors compared to Fiji went into making the chip reach higher clocks (for example shortening data paths by reducing the number of consumers connected to certain caches).
 
Well that begs the question, what exactly is "Vega20" ..?
When you know for certain and can tell me that, then I'll start to believe your claimless predictions. I said 7nm Vega for gaming...
Vega 20
  • 7nm
  • the same core config as Vega 10 = 64CU, 4 SEs, etc.
  • additionally: half-rate FP64, some new instructions for INT8 and FP8 (I don't remember this, so feel free to add the correct instructions) => GFX9 in a higher revision than Raven
  • new IO = PCIe4, fast external GMI links - XGMI
  • new memory = two additional HBM2 links above Vega 10; allows 32GB VRAM
  • TDP 150-300W = similar to Vega 10, so probably 1.6+GHz
Does 32GB VRAM, half-rate FP64 or external IO links sound anything like "gamer stuff"?
 
Back
Top