AMD Radeon VII Announcement and Discussion

[Deleted member 13524]

That was GlobalFoundries' original plan.

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 all time?​

Navi first appeared publicly on roadmaps during the 2016 Capsaicin event, which was in March. There was no info about Navi's process then, nor Vega's for that matter.
Those news about GF skipping 10nm are from September 2016, so half a year later. During September 2016 AMD already had updated the roadmap with release schedules much closer to truth than what we saw during Capsaicin.

Regardless, I'm not saying Navi was definitely to be a 10nm chip back in March 2016. It could be 10nm, or maybe 7nm with a mountain of wishful thinking for GlobalFoundries to deliver 7nm volume production in Q2 2018.

I mean if we look at that roadmap from Capsaicin, we can tell Vega slipped by around 6 months, and that could have been due to HBM2 performance and/or volume, not hitting desired clocks, etc.
But Navi? Navi is on its way to be 18-21 months late. Now AFAIR that's a historical clusterfuck only surpassed by Canon Lake.

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.

What are you talking about? The RX480 released in June 2016, exactly when that roadmap said it would release. They wouldn't have a 6-month delay over the release presented 3 months prior.

 

[DmitryKo]

During September 2016 AMD already had updated the roadmap with release schedules much closer to truth than what we saw during Capsaicin.

According to this version, Navi 10/11 was to be a mid-range part similar to Polaris 10/11 but offering the performance level of Vega10/11.

Then again, the lower-end Vega 11 (Radeon Vega 32?) never materialized (AMD even went as far as trying to debunk the discrete Vega 11 codename) and Navi apparently slips by yet another year.

So AMD solved the problem by not disclosing GPU roadmaps anymore.


BTW, didn't even know Vega 10 x2 exists - quite obscure as it's only used by "Radeon Pro V340" MxGPU...

 

[yuri]

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.

This roadmap... Sure, the X-axis shows an unhealthy dose of optimistic planning but what about the Y-axis? Which Vega shows ~2x perf/watt of Polaris (RX 470)?

 

[Frenetic Pony]

This roadmap... Sure, the X-axis shows an unhealthy dose of optimistic planning but what about the Y-axis? Which Vega shows ~2x perf/watt of Polaris (RX 470)?

It did, for professional workloads when clocked right.

Been over how something in the Vega architecture vastly restricts it's performance on games compared to professional workloads. Also how something vastly restricted it's aimed for clockspeed, the Vega 64 was supposed to be clocked similarly to the Vega 7. If it had done what was hoped the Vega 64 would've been as fast as a 2080ti, instead, welp.

 

[Deleted member 13524]

Then again, the lower-end Vega 11 (Radeon Vega 32?) never materialized (AMD even went as far as trying to debunk the discrete Vega 11 codename) and Navi apparently slips by yet another year.

Vega 11 exists, in the form of Vega Mobile for the MacBook Pro, with a single HBM2 stack and 20 CUs.
The fact that AMD was counting on Vega 11 to replace Polaris 10 on compute platforms just goes to show how far off they ended up from their initial clock estimates.
Vega 11 would need to clock towards 1.7GHz in order to compensate for the lower amount of CUs compared to Polaris 10. The moment Vega hit an efficiency wall at 1.4GHz it just ceased to make sense, so they extended Polaris 10 as much as they could.

Another proof is the Vega 10 ROCm card they show in those slides, describing a Vega 10 doing 12 TFLOPs (~1450MHz) at 225W TDP.
That card also exists, it's the Radeon Pro
V320. It has a 230W TDP, but it's a cut-down 56CU chip clocked at.. 1GHz boost, with 7.6 TFLOPs.
The "pure ROCm" Vega MI25 ended up with a 300W TDP.

So AMD solved the problem by not disclosing GPU roadmaps anymore.

Perhaps it's for the best. I'm glad that AMD is on the path of announcing close to release. Raja's overhype was starting to become unbearable.

BTW, didn't even know Vega 10 x2 exists - quite obscure as it's only used by "Radeon Pro V340" MxGPU...

The slides only show ROCm GPUs. I don't think there ever was a consumer dual-Vega10 in the works.

 

[DmitryKo]

Vega 11 exists, in the form of Vega Mobile for the MacBook Pro, with a single HBM2 stack and 20 CUs.

I believe the mobile Radeon Pro Vega 20 is called Vega 12 in the latest LLVM drivers.

The fact that AMD was counting on Vega 11 to replace Polaris 10 on compute platforms just goes to show how far off they ended up from their initial clock estimates.

And that this 32-40 CU Vega 11 part seems to have been spun and re-spun to death in the end, while Polaris 20/30 lives on. And that Kaby Lake G ended up using Polaris graphics engine as well.

I don't think there ever was a consumer dual-Vega10 in the works.

Sure, I'm just surprised they still use dual chip - even for quite strange stuff like Radeon Pro Duo Polaris.

 

[Deleted member 13524]

http://www.redgamingtech.com/amd-vega-based-radeon-vii-for-consumers-and-gamers-exclusive/

These guys seemingly leaked the Vega VII 2 weeks in advance, with pictures of the card.
They claim an availability of 20 000 cards for the first batch and 40 000 for the second batch.

Seems like a pretty wide release for a $700 card.

 

[Picao84]

http://www.redgamingtech.com/amd-vega-based-radeon-vii-for-consumers-and-gamers-exclusive/

These guys seemingly leaked the Vega VII 2 weeks in advance, with pictures of the card.
They claim an availability of 20 000 cards for the first batch and 40 000 for the second batch.

Seems like a pretty wide release for a $700 card.

Is that quantity world wide? If yes, I can't see how that is wide at all ..

 

[yuri]

Also how something vastly restricted it's aimed for clockspeed, the Vega 64 was supposed to be clocked similarly to the Vega 7. If it had done what was hoped the Vega 64 would've been as fast as a 2080ti, instead, welp.

This sounds a bit strange. Vega 64 top SKU have already been clocked 60% higher than Fiji (1050/1677MHz). It hits the powerwall hard past 1400-1500MHz. Vega VII got boosts to 1800MHz or 71% higher than Fiji.

This sounds like Pentium 4 era plans and hopes all over again - any NetBurst fireball CPU running at 10GHz would surely be capable. Yet, no such CPU has been produced.

 

[A1xLLcqAgt0qc2RyMz0y]

Limited supply: only 5000 available for entire production cycle

AMD Radeon Vega VII Rumored To Have Less Than 5000 Units Made
https://wccftech.com/amd-radeon-vega-vii-5000-units-64-rops-no-fp64-compute

AMD Radeon VII: less than 5000 available, no custom cards
https://www.tweaktown.com/news/64501/amd-radeon-vii-less-5000-available-custom-cards/index.html

Since the announcement I reached out to some industry contacts who said there will be "less than 5000" made.

...

There will also reportedly be NO custom Radeon VII graphics cards made

 

[Deleted member 13524]

Only tweaktown came up with that rumor and wccftech copy+pasted.

I'd say the other source I posted claiming 20k+40k is way more credible.
First because they actually leaked many Radeon VII details and pictures prior to announcement in the same newspiece, second because AMD isn't in the business of losing money on videocards. They've been pretty comfortable with not competing on the top end for over 30 months, so they wouldn't launch a card to lose money out of peer pressure now.

And also because...
Tweaktown is
Really
Terrible
With rumors

They've been in the business of making stuff up for a while.

 

[Ailuros]

Back in June 2012 the HD7970 1Ghz did 1 TFLOPs FP64 at $500.
AMD only started to artificially strip down FP64 throughput with Hawaii.

Which is almost 7 years ago, for which NV used to sell 1-2 years later for the Titan/Titan Black with 1.5/1.7 TFLOPs FP64 at a $999 MSRP, which they themselves noted afterwards that it was a mistake.

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.

FP64 is obviously castracted on R VII and I severely doubt that AMD reached that decision for no reason whatsoever.

 

[Deleted member 13524]

AMD responds to short supply rumors:

While we don't report on production numbers externally, we will have products available via AIB partners and AMD.com at launch of Feb. 7, and we expect Radeon VII supply to meet demand from gamers.

One thing that's changed and I don't think has been commented is the fact that AMD will be selling cards directly from their site. Maybe this is the way they found to avoid distributors and retailers to sell above MSRP to increase margins when faced with high demand, working as sort of a safeguard against mining booms.

Which is almost 7 years ago, for which NV used to sell 1-2 years later for the Titan/Titan Black with 1.5/1.7 TFLOPs FP64 at a $999 MSRP, which they themselves noted afterwards that it was a mistake.

Did they really make any statement about it being a mistake or are you assuming from the later Pascal and Turing Titan cards that use chips without functional FP64 throughput?
Because Titan V is selling right now with 1:2 FP64 throughput so it's certainly competing with the Quadro and Tesla V100.

FP64 is obviously castracted on R VII and I severely doubt that AMD reached that decision for no reason whatsoever.

It wasn't obviously castrated when @Ryan Smith asked them a week ago:

https://twitter.com/x/status/1083959608371175424

For now, I'd say there are diverging reports.
Maybe it's capped, maybe it's uncapped, maybe it's uncapped but AMD doesn't want to make it public too much. Maybe AMD is testing the waters by sending different info to different people to gauge the interest on FP64 to see if they'll enable FP64 on the firmware/drivers or not.
Who knows...

 

[entity279]

AMD responds to short supply rumors:
we expect <insert random enthusiast-gamer oriented cpu/gpu here> supply to meet demand from gamers

famous last words

 

[Ailuros]

Did they really make any statement about it being a mistake or are you assuming from the later Pascal and Turing Titan cards that use chips without functional FP64 throughput?
Because Titan V is selling right now with 1:2 FP64 throughput so it's certainly competing with the Quadro and Tesla V100.

For which the Titan V is selling at 3x times the MSRP ($2999) of those early 2012/3 Titan SKUs.

For now, I'd say there are diverging reports.
Maybe it's capped, maybe it's uncapped, maybe it's uncapped but AMD doesn't want to make it public too much. Maybe AMD is testing the waters by sending different info to different people to gauge the interest on FP64 to see if they'll enable FP64 on the firmware/drivers or not.
Who knows...

If AMD should consider a professional solution in the future with nearly 7 TFLOPs FP64 amongst others (presupposition the chip is actually capable of an as high rate which I don't know), it would be mighty naive of them to sell it at just 700 bucks. Make it twice as much and they're still better positioned against whichever NV solution, but not at a consumer performance GPU MSRP.

 

[Deleted member 13524]

For which the Titan V is selling at 3x times the MSRP ($2999) of those early 2012/3 Titan SKUs.

The first Titan was an experimental product for a niche, AFAIK it's the first "prosumer" graphics card. From there on, nvidia had been raising the price for the Titan line. The Titan RTX costs almost as much as the Titan V and it has 1/12th of the latter's FP64 performance.
I doubt any of the Titan cards ever cost more than $500 to make (except maybe for the dual-GPU Titan Z), so upwards of $800 it's probably just a matter of how much more money they can make on top-end enthusiasts.

If AMD should consider a professional solution in the future with nearly 7 TFLOPs FP64 amongst others (presupposition the chip is actually capable of an as high rate which I don't know), it would be mighty naive of them to sell it at just 700 bucks. Make it twice as much and they're still better positioned against whichever NV solution, but not at a consumer performance GPU MSRP.

AMD can't ask the same as nVidia for similar FP64 performance, especially on prosumer cards, because they don't support CUDA which is the de facto standard for GPGPU compute and they lack tensor cores.
They'd always have to compete on price.

 

[Ailuros]

The first Titan was an experimental product for a niche, AFAIK it's the first "prosumer" graphics card. From there on, nvidia had been raising the price for the Titan line. The Titan RTX costs almost as much as the Titan V and it has 1/12th of the latter's FP64 performance.
I doubt any of the Titan cards ever cost more than $500 to make (except maybe for the dual-GPU Titan Z), so upwards of $800 it's probably just a matter of how much more money they can make on top-end enthusiasts.

That still doesn't change the fact that you can find a relatively cheap NV SKU today with a high DP rate.

AMD can't ask the same as nVidia for similar FP64 performance, especially on prosumer cards, because they don't support CUDA which is the de facto standard for GPGPU compute and they lack tensor cores.
They'd always have to compete on price.

Where did I propose that AMD should ask the same MSRPs as NV?

 

[Ryan Smith]

Alright, I finally have the Radeon VII FP64 performance matter sorted out with AMD.

Contrary to earlier statements, it is being throttled. Radeon VII's rate will be 1:8, versus Vega 20's native 1:2 rate. So if you want the max rate, you're still going to have to buy MI50/MI60 or whatever Radeon Pro card they eventually come out with.

It is notable though that 1:8 is still twice the native FP64 rate of all other Vegas. So it'll still be a lot faster at FP64 operations than Vega 56/Vega 64.

 

[Ike Turner]

Alright, I finally have the Radeon VII FP64 performance matter sorted out with AMD.

Contrary to earlier statements, it is being throttled. Radeon VII's rate will be 1:8, versus Vega 20's native 1:2 rate. So if you want the max rate, you're still going to have to buy MI50/MI60 or whatever Radeon Pro card they eventually come out with.

It is notable though that 1:8 is still twice the native FP64 rate of all other Vegas. So it'll still be a lot faster at FP64 operations than Vega 56/Vega 64.

Thanks for the update Ryan. Still boggles the mind that RTG's communication is so lackluster.

 

[vipa899]

I actually like reference more then 3rd party designs, hard to find reference though.