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

Didn't other leaks also claim 12TF SP performance? That would make for 6TF DP at half rate. That would more than double the W9100.

Consumer version, yes. I'm just assuming the professional version would come with somewhat lower clocks, as it happens with Hawaii solutions (you don't want your >$4000 card failing on your customers..).
But if consumer Vega 10 really does 1.5GHz, then professional Vega would probably do more than 1.2GHz.
 
Consumer version, yes. I'm just assuming the professional version would come with somewhat lower clocks, as it happens with Hawaii solutions (you don't want your >$4000 card failing on your customers..).
But if consumer Vega 10 really does 1.5GHz, then professional Vega would probably do more than 1.2GHz.

To my understanding this leak (or rumor, I dunno) was talking about the HPC parts with 24tflops of FP16, ergo 12 FP32. Vega10 wouldn't have 1:2 DP performance.
 
I doubt big Vega is server oriented, just that it's going there first. Cost isn't an issue there if constrained by component production. Not to mention the margins will be better.
The article uses "announce" as well as "launch" in different sections. Paper launches aside, the lead time between announcement and getting market can be longer for the non-consumer space.
Even if there's a volume constraint, there's sufficient precedent at this point for a high starting price for a new higher-performing product--unless the product cannot be described in that way.

Validation in theory should be what keeps the non-consumer products behind the consumer ones.
If there's product available, it's avoiding revenue and the benefits of exposing hardware to the real world to help software development and validation. How Vega could be validated for the more stringent requirements of the server space while not being ready for enthusiasts is unclear.
 
It wastes too much silicon and isn't that useful beyond a few specific markets. Which is probably why they mention Vega20 were an entire chip is dedicated towards that purpose. The 20 variant could compete against P100, but my guess is that it's a counterpart to a HPC Zen. That would be more practical for the extremely large dataset guys wanting X86.

Validation in theory should be what keeps the non-consumer products behind the consumer ones.
Normally yeah, but if the ultimate goal is pairing Zen and Vega in the server market the launch could be intended to allow partners to begin the validation process. There could actually be sufficient demand just for development before even reaching production environments. There would seem to be quite a few possibilities there given a shared memory pool that would take time to develop. The SSG might be a precursor for validation. The architecture of the chip is likely irrelevant to the tasks being adapted to it. Hardware costs, assuming they're even selling them initially, would be insignificant to those clients. Current constraints might also be pricing them out of the enthusiast market. It's possible they initially launch development kits as opposed to production hardware following months later.
 
It wastes too much silicon and isn't that useful beyond a few specific markets. Which is probably why they mention Vega20 were an entire chip is dedicated towards that purpose. The 20 variant could compete against P100, but my guess is that it's a counterpart to a HPC Zen. That would be more practical for the extremely large dataset guys wanting X86.

You're suggesting the FirePro W9100's successor will arrive in the form of Vega 20, a 7nm GPU that comes out almost 5 years after the former's release in the market.

I think I'll keep my opinion that Vega 10 will be a 2:1 DP GPU, if fully unlocked.
 
You're suggesting the FirePro W9100's successor will arrive in the form of Vega 20, a 7nm GPU that comes out almost 5 years after the former's release in the market.

I think I'll keep my opinion that Vega 10 will be a 2:1 DP GPU, if fully unlocked.

Theres an enormous void both in price and performance between rx480 and gtx 1080, how are they going to fill that gap if it's a DP part? They could disable 1:2 functionality on the consumer parts but they'd still be wasting a DP chip in a ~400 dollar card (1070 competitor)
 
Theres an enormous void both in price and performance between rx480 and gtx 1080, how are they going to fill that gap if it's a DP part? They could disable 1:2 functionality on the consumer parts but they'd still be wasting a DP chip in a ~400 dollar card (1070 competitor)

That's exactly what AMD did with Hawaii. In every R9 290, R9 290X, R9 290X2, R9 390 and R9 390X there's a 2:1 capable chip inside.

AMD can't afford to churn out 4 to 5 graphics chips a year. They're sticking to 2 graphics chips a year for the foreseeable future. This means they have to embrace a larger market with each chip.
 
Not while increasing margins....... And that has been the goal from last year..... IF they do it this way, margins will not increase unless they can price higher than nV counterparts, that means, beat them in performance and perf/watt. I find that highly unlikely.
 
Not while increasing margins....... And that has been the goal from last year..... IF they do it this way, margins will not increase unless they can price higher than nV counterparts, that means, beat them in performance and perf/watt. I find that highly unlikely.

Colour me shocked, horrored and Surprised.......
We already know you think that Vega is nothing more then Polaris so i guess we will just wait and see.....

I still think im right in that thax to the 20nm issues and the choices AMD made they will only catchup in terms of technology life cycles with Vega. Fiji was a hack job, Polaris was the Maxwell competitor(delayed and moved from 20nm to 14nm) and Vega is the pascal competitor ( on time).

I guess we will see who's right, not that you would ever concede anything.........
 
Colour me shocked, horrored and Surprised.......
We already know you think that Vega is nothing more then Polaris so i guess we will just wait and see.....

I still think im right in that thax to the 20nm issues and the choices AMD made they will only catchup in terms of technology life cycles with Vega. Fiji was a hack job, Polaris was the Maxwell competitor(delayed and moved from 20nm to 14nm) and Vega is the pascal competitor ( on time).

I guess we will see who's right, not that you would ever concede anything.........


There is no such thing as a hack job man not with projects that take longer than 3 years to make. Fiji was miscalculation. Unknown variable was Maxwell. They had no control over that. Were they expecting Pascal, with its crazy clocks? I think not either. If they were, then Vega should be well above Pascal, not close to it. 225 watts sounds crazy from what we have seen so far.

Vega had just about as much time as Polaris, and that is the problem. Look, that 225 watts rumor, What I think its from is just the chip itself. Yeah after seeing how AMD spun its marketing wheels with Polaris, color me confused. Cause I don't care how they spun their words, it came out as if it was a last years product in the metrics I mentioned. If Vega is that much better in 6 months, then Polaris should never have been there, because I don't see them gaining marketshare from Polaris......

PS I don't need to concede to anything, your cloud server crap can go to hell, it makes no sense, and that is it. Please come back in 20 years and make your point, you don't understand how corporations work, and what infrastructure needs are, and what is needed to sustain current business, and that is all that matters. $ in the now.
 
Last edited:
Consumer version, yes. I'm just assuming the professional version would come with somewhat lower clocks, as it happens with Hawaii solutions (you don't want your >$4000 card failing on your customers..).
But if consumer Vega 10 really does 1.5GHz, then professional Vega would probably do more than 1.2GHz.

The leak was for server roadmap and professional cards so it's the other way round.
 
That's exactly what AMD did with Hawaii. In every R9 290, R9 290X, R9 290X2, R9 390 and R9 390X there's a 2:1 capable chip inside.

AMD can't afford to churn out 4 to 5 graphics chips a year. They're sticking to 2 graphics chips a year for the foreseeable future. This means they have to embrace a larger market with each chip.

I don't think AMD wanted to do that with Hawaii. A lot didn't go according to plan with Hawaii.

Hawaii debuted as a $550-400 part (halo-level at that time), but quickly dropped it to $300-400 levels amid stiff competition from GM204.

Shoot, AMD infamously low-balled the performance requirements of the 290's market segment and had to crank up its fan speed to painful levels at the last minute to eek out additional performance.

Also remember how Hawaii was supposedly designed to operate at a laughable 95C in order to exploit increased thermal transfer rates?

Ultimately, you simply can't sustainably sell a ~275W 512-bit 1:2 DP-capable chip for ~$300. You don't plan for lean margins like that.

Hawaii was a hell of a workhorse and absolutely invaluable to AMD in 2014 & 2015, but I can't fathom that Hawaii went according to plan. Therefore, I don't see AMD intentionally trying to repeat their experience with Hawaii.
 
Ultimately, you simply can't sustainably sell a ~275W 512-bit 1:2 DP-capable chip for ~$300. You don't plan for lean margins like that.

Hawaii was a hell of a workhorse and absolutely invaluable to AMD in 2014 & 2015, but I can't fathom that Hawaii went according to plan. Therefore, I don't see AMD intentionally trying to repeat their experience with Hawaii.

Then what chip will replace Hawaii in the high-DP compute market?

- Hawaii is 3 years old.
- It seems to me AMD is sticking with 2 graphics chips/year.
- 2016 has had its fill.
- 2017 is Vega 10 and Vega 11
- By late 2017 Hawaii will be 4 years old

If Vega 10 can't replace Hawaii in all its markets, then you're suggesting AMD will just stop competing in that market for several years in a row?
People are shocked enough that AMD isn't competing in the high-end consumer market for 3 or 4 quarters..
 
Personally, I think only the bigger chip will have 1:2 DP and it will target both the professional market and the high end market (with the DP disabled). The high end market is still a high margin market. They can price it between Titan and 1080.

Why some people are assuming that Vega 10 with the 1:2 DP removed will be sold at $300-400? What kind of logic is that? This will be at least a 12TF part if the rumor is true. This will be AMD halo product.
They don't need to beat Nvidia in terms of perf/watt. They just need to beat it in terms of perf/$ while still having a reasonable power consumption.
 
For the question of Vega 10 DP rate ... Well the probem i see is more a question of timeframe:

We know that Vega is developped since a good time, and it is logical ( and what we was thinking ) that it was a big architecture aimed at both professional and gaming market: So in this aspect we can suppose that it was logically 1:2 DP enabled for replace Hawaii old W9100, S11000 etc, we see today it have to double FP16 ( and surely Int 8 ) performance. So something in the line of Hawaii for DP and Pascal for the rest.

Vega have been developped since 3 years maybe, and was surely expected to be launched in 2016 initially. But as we have seen with GP100 and HBM2, it seems HBM2 was or too costly, or not fully ready for a 2016 launch. ( GP100 exist but real availability is for Q1 2017 ).

So if Vega 20 is the "big enabled DP " chips, this mean DP have been disabled on Vega10, i really doubt they had release 2 version of it. Then the timeframe of Vega20 could be a logical move if it come quickly after Vega10,
but honestly, with a release on the end of 2017 ( meaning surely Q1 2018 ) for Vega20; One year after Pascal P100 availability, that will not be a so good move from AMD.

2016>end of 2017- Q1 2018, that make a long delay for release a Vega professional sku.. and as with 1080-1070, its like let the market of high end compute gpus to Nvidia P100 for 1 more year.

Maybe AMD have change is plan in between, have redesign Vega10, removed the DP logic etc, but just for release a Vega20 with it 6months, 1year later ?
 
Last edited:
Why some people are assuming that Vega 10 with the 1:2 DP removed will be sold at $300-400?

I don't know either.
@ImSpartacus pointed out that's what happened to Hawaii when GM204 came out, so it could happen again to Vega 10 when Volta is released a few quarters later.
I think if AMD was predicting to be undercut on every corner at all times, then they could just as well give up on discrete graphics cards altogether.
 
I don't know either.
@ImSpartacus pointed out that's what happened to Hawaii when GM204 came out, so it could happen again to Vega 10 when Volta is released a few quarters later.
I think if AMD was predicting to be undercut on every corner at all times, then they could just as well give up on discrete graphics cards altogether.
Is there any actual reason to believe Volta has suddenly jumped 1 year early from the roadmap?
Yes, there was one rumor saying that, but that's just one rumor - there's also a rumor it would be 7nm which rules out 2017 release, and then the actual fact that it has always been 2018 on the roadmap (since it returned to it)
 
Is there any actual reason to believe Volta has suddenly jumped 1 year early from the roadmap?
Yes, there was one rumor saying that, but that's just one rumor - there's also a rumor it would be 7nm which rules out 2017 release, and then the actual fact that it has always been 2018 on the roadmap (since it returned to it)
I think some of the discussion has been around the Summit supercomputer which will use Volta and is planned for 2017.

Summit Fact Sheet said:
The machine, named Summit, will arrive in 2017 and be ready for users in 2018.
I'm not assuming that any Volta GPU will show up anywhere other than in Summit before 2018 though. Even Summit won't actually be available until then.

Regarding 7 nm, this rumor says that
WCCFTech said:
NVIDIA is expected to drop 10nm process and go straight for 7nm with their post-Volta GPUs,
which would imply that Volta is 14/16 nm.
 
Back
Top