AMD Radeon VII Announcement and Discussion

[Ike Turner]

The downside to AMD keeping their cards so close to their chest is that too few people know what's going on, it would seem. It goes without saying that this is not the answer I got last week.

Who the hell was the clueless AMD rep who gave you all the bad info?

 

[Geeforcer]

It seems pretty clear that just like the roadmap indicated for years Vega 20 was never meant for consumer market; Radeon VII has “emergency stopgap” written all over it.

 

[Rootax]

It seems pretty clear that just like the roadmap indicated for years Vega 20 was never meant for consumer market; Radeon VII has “emergency stopgap” written all over it.

And a 4k / bandwidth limited stop gap. I'm pretty sure otherwise the perfs will be pretty close to Vega64 LC...

 

[DavidGraham]

Official word that there will be no FP64 support.

We reached out to AMD’s Director of Product Marketing Sasa Marinkovic to inquire whether or not the FP64 inclusion was real, and were told quite simply that “Radeon VII does not have double precision enabled.” That means instead of delivering 6.7 TFLOPS of FP64 like the MI50, Radeon VII will be closer to ~862 GFLOPS (it’s 1:32 with half-precision like RX Vega).

https://techgage.com/news/radeon-vii-caps-fp64-performance/

 

[BoMbY]

I just wish they would do a 64 CU LC version of Vega20.

 

[mrcorbo]

Man there is an awful lot of wasted silicon when using this chip in this implementation.

 

[Putas]

Where is this fetish for fully enabled chips coming from?

 

[Deleted member 13524]

A retail GPU yielding up to nearly 7 TFLOPs FP64 at a mere 700 bucks, is obviously way too good to be true.

Back in June 2012 the HD7970 1Ghz did 1 TFLOPs FP64 at $500.
AMD only started to artificially strip down FP64 throughput with Hawaii.
I doubt a 1:2 FP64 Radeon VII was ever going to cannibalize the MI50. The MI50 is apparently only selling inside specialized server stacks or even full server towers, so I wonder if this is really a valid concern.

Within the compute community, AMD cards have the considerable disadvantage of not running CUDA, so AMD couldn't ever ask the same money as a nvidia counterpart just for parity on floating point performance.
That's where a 1:2 FP64 "prosumer" Radeon VII made sense: gain market from nvidia for the cost/performance disparity and get those people to start using ROCm supported languages.

It seems pretty clear that just like the roadmap indicated for years Vega 20 was never meant for consumer market; Radeon VII has “emergency stopgap” written all over it.

Is Radeon VII really a consumer card? AMD sure seems to be pointing the card at the prosumer "niche" by constantly mentioning content creators.



I'd look at the VII as a consumer card had they cut down FP64 throughput, cut the memory stacks down to 3 for 12GB/768GB/s (or even 2 stacks of 2.4Gbps Aquabolt for 8GB / 614GB/s) and raised the core clocks by 5-15% compared to the Instinct equivalents.
The end result seems to be a product that doesn't really fit either a Pro card nor a prosumer nor a consumer one. Sure performance is on par with a RTX 2080, but that 4th HBM2 stack seems to be just taking space and increasing BoM IMO. It's not like the perception of value was ever going to significantly increase between 12 and 16GB VRAM, and I bet the bandwidth delta only really benefits heavy FP64 computations.. which the card apparently can't do either way.

 

[Rootax]

But they showed the card vs 2080 too, so, mixed message right there.

 

[Ike Turner]

Back in June 2012 the HD7970 1Ghz did 1 TFLOPs FP64 at $500.
AMD only started to artificially strip down FP64 throughput with Hawaii.
I doubt a 1:2 FP64 Radeon VII was ever going to cannibalize the MI50. The MI50 is apparently only selling inside specialized server stacks or even full server towers, so I wonder if this is really a valid concern.

Within the compute community, AMD cards have the considerable disadvantage of not running CUDA, so AMD couldn't ever ask the same money as a nvidia counterpart just for parity on floating point performance.
That's where a 1:2 FP64 "prosumer" Radeon VII made sense: gain market from nvidia for the cost/performance disparity and get those people to start using ROCm supported languages.


Is Radeon VII really a consumer card? AMD sure seems to be pointing the card at the prosumer "niche" by constantly mentioning content creators.



I'd look at the VII as a consumer card had they cut down FP64 throughput, cut the memory stacks down to 3 for 12GB/768GB/s (or even 2 stacks of 2.4Gbps Aquabolt for 8GB / 614GB/s) and raised the core clocks by 5-15% compared to the Instinct equivalents.
The end result seems to be a product that doesn't really fit either a Pro card nor a prosumer nor a consumer one. Sure performance is on par with a RTX 2080, but that 4th HBM2 stack seems to be just taking space and increasing BoM IMO. It's not like the perception of value was ever going to significantly increase between 12 and 16GB VRAM, and I bet the bandwidth delta only really benefits heavy FP64 computations.. which the card apparently can't do either way.

Full rate FP64 would have been of no help even for content creation because there isn't a single CC app that would utilize it. On the other hand 1Tb/s of BW & 16GB of VRam are additions that will improve performance in CC workflow in a tangible way.

 

[Geeforcer]

Is Radeon VII really a consumer card? AMD sure seems to be pointing the card at the prosumer "niche" by constantly mentioning content creators.



I'd look at the VII as a consumer card had they cut down FP64 throughput, cut the memory stacks down to 3 for 12GB/768GB/s (or even 2 stacks of 2.4Gbps Aquabolt for 8GB / 614GB/s) and raised the core clocks by 5-15% compared to the Instinct equivalents.

Well, they are bundling it with 3 games...

The end result seems to be a product that doesn't really fit either a Pro card nor a prosumer nor a consumer one. Sure performance is on par with a RTX 2080, but that 4th HBM2 stack seems to be just taking space and increasing BoM IMO. It's not like the perception of value was ever going to significantly increase between 12 and 16GB VRAM, and I bet the bandwidth delta only really benefits heavy FP64 computations.. which the card apparently can't do either way.

Which all the more lends the credence to the "last-minute stopgap" theory. The rumor about Mike Rayfield departure and him being described as disengaged, disinterested and sighting they yet-announced Radeon VII as the evidence of this.

I think Navi (or at least its consumer version) going back to the drawing board rumor makes sense, which would obviously cause some short-term pain for potential long-term gain.

 

[Rootax]

Navi is going back to the drawing board ?

 

[Geeforcer]

Navi is going back to the drawing board ?

Just rumors obviously, but there have been mentions of "major re-spins" to "design changes"

Of course, Navi rumors have run the full gamut, from 2H18 (based on a past roadmap) to 2H2019 release date, various levels of interdependence with PS5 design, MCM - "Tensor-like cores" - High End - No, mainstream. I would not be at all surprised that when finally release it would be an vastly different design than initially conceived in 2016 and maybe even from what it was tracking to in early-mid 2018.

 

[Deleted member 13524]

Well, they are bundling it with 3 games...

And so have some Geforce Titan models, which are also "prosumer" cards.

Which all the more lends the credence to the "last-minute stopgap" theory.

I buy the "stopgap" part because Vega is clearly below Pascal and Volta in performance/power and performance/transistors. AMD needs to solve this, Vega was supposed to be mostly Polaris/GFX8 ISA with DX12.1 and high clocks and they failed with that last part. 7nm came and the clock bump was still pretty modest, so it's clear they need another architecture iteration to be able to better compete with nvidia.

I don't buy the "last-minute" part, though. How long is "last-minute"? AMD hasn't had a nvidia competitor on the top-end since Fiji from 2015, and they have known Navi's time to market for a while. They've had all the information they needed to know if they should or shouldn't launch a Vega 20 consumer card for a year, at least.

The only thing that changed in the meantime was the release of Turing cards.
Maybe they weren't counting on Turing getting rather lukewarm response, saw RTX 2080's price and figured out they could make money on a Vega 20 selling at a similar price.

Navi is going back to the drawing board ?

It's definitely not going back "now" or they wouldn't be announcing/releasing it in 2019, but Navi appeared first in AMD's roadmaps as projected for an early 2018 release.

Early 2018 wouldn't even be possible on 7nm, so I'm guessing Navi was a 10/12/14nm architecture at some point, which had to go back to the drawing board when GlobalFoundries decided they weren't going to compete on high-end processes anymore.

 

[DmitryKo]

It seems pretty clear that just like the roadmap indicated for years Vega 20 was never meant for consumer market; Radeon VII has “emergency stopgap” written all over it.

Which all the more lends the credence to the "last-minute stopgap" theory.

Indeed, it looks like they just needed something for CES publicity.

'VII' trademark was only applied for in November - typically that step would be taken much early on.

I think Navi (or at least its consumer version) going back to the drawing board rumor makes sense

Navi is going back to the drawing board ?

Major redesign is not likely at this point, but certainly it was pushed back even more, probably to end of 2019.

 

[DavidGraham]

I just wish they would do a 64 CU LC version of Vega20.

MI60 has a base clock of 1200MHz, compared to 1450MHz in VII. It seems 1800MHz is just a peak clock for all SKUs. VII was cut down to allow it to reach higher base/average clocks, and that also required it to have 2X 8pin power connectors. Funny how some here expected 2100MHz, despite MI60 1800MHz max clock. It seems 7nm didn't really fix most of Vega's power issues.

Of course, Navi rumors have run the full gamut, from

AMD officials are saying they are going to "divulge" more information on Navi in 2019, this doesn't sound like a release date to me. They already have a mid range refresh (RX 590) and a high end refresh (VII). They could do a low end Navi replacement before the end of the year.

 

[DmitryKo]

Early 2018 wouldn't even be possible on 7nm

That was GlobalFoundries' original plan at the time.

https://www.fudzilla.com/news/processors/41626-gloflo-confirms-skipping-over-10nm
https://www.tomshardware.com/news/amd-zen-10nm-7nm-intel,32619.html

so I'm guessing Navi was a 10/12/14nm architecture at some point, which had to go back to the drawing board when GlobalFoundries decided they weren't going to compete on high-end processes anymore

I can't recall AMD announcing any plans for the 10 nm node - didn't they intend to skip that node in sync with GF's much-publicized 7nm roadmap?

 

[Esrever]

That roadmap has been wrong since Polaris tbh. Polaris was half a year late, Vega was a year late, who knows how much Navi would be late by looking at this.

 

[Pressure]

That was GlobalFoundries' original plan at the time.

https://www.fudzilla.com/news/processors/41626-gloflo-confirms-skipping-over-10nm
https://www.tomshardware.com/news/amd-zen-10nm-7nm-intel,32619.html

I can't recall AMD announcing any plans for the 10 nm node - didn't they intend to skip that node in sync with GF's much-publicized 7nm reoadmap?​

Sorry to go off topic but your text is impossible to read with the dark theme.

 

[DmitryKo]

Sorry to go off topic but your text is impossible to read with the dark theme.

Apologies, stupid visual editor keeps inserting unneeded formatting tags.