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

hmm, so the benchmarks leaked showing, in some cases, +80% compute performance compared to Vega64 are also likely fake
 
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).
No. TSMC has 2 variants of 7 nm, one meant for mobile (used by A12, Kirin, Snapdragon etc) and one meant for high performance (used by Zen 2, Vega 20, Navi and future NVIDIA chips etc). Mobile chips don't eat HPC allocations and vice versa
 
No. TSMC has 2 variants of 7 nm, one meant for mobile (used by A12, Kirin, Snapdragon etc) and one meant for high performance (used by Zen 2, Vega 20, Navi and future NVIDIA chips etc). Mobile chips don't eat HPC allocations and vice versa

Are you sure about that? I mean, they have two processes, but don't they share equipment or production lines?
 
Are you sure about that? I mean, they have two processes, but don't they share equipment or production lines?
Surely with processes this complex it's not just about flipping a switch on the product line, I'm quite confident they're completely separate to push out both
 
Are you sure about that? I mean, they have two processes, but don't they share equipment or production lines?

Already stated, no!

TSMC has expanded their 7nm process. And has invested heavily into their 7nm capacity. And understand, Dr Su inked their 7nm wafer deals some 17 months ago and knew about AMD's chiplet design and modular Vega/Navi/Zen architecture outlook. And went "all in" @ 7nm with TSMC with deals. Also knowing (way before we ever did) that Global was releasing them on their previous contract, due to changes in their own (Global's) companies direction of dropping 7nm from their roadmap.

Dr Su has mentioned many times and heralded how RTG can make use of TSMC's new process. And hinted as substantial investment into RTG over the past 2 years and that see sees graphics as AMD's future. AMD is one of TSMC's biggest 7nm contracts. With other big players trying to squirm their way into TSMC's 7nm goodness.
 
  • Like
Reactions: ECH
Vega 64 with 1:2 double precision, additional deep learning instructions and interconnect.
Nothing of this is particularly gamey.
And significantly higher clocks on TSMC 7nm compared to GF 14LPP.
A Vega 10 clocked 30% higher than 1.5GHz (~2.0GHz) would perform at the same level as a 1080Ti or 2080.

You also could build one being 35% faster and you could build one, that's twice as efficient.
Where do those 35% come from?

I see 30% claimed by TSMC compared to 16FF+, but Vega 10 isn't made on that process. It's made on GF's 14LPP which according to various statements ( from @Nebuchadnezzar et al) is some 10-20% worse than TSMC's 16FF+.
So we're looking at ~45% to 55% higher clocks on Vega 20, at ISO power.


Does 32GB VRAM, half-rate FP64 or external IO links sound anything like "gamer stuff"?
I doubt they'd make a gaming Vega 20 with 4 stacks.
But 3 HBM2 stacks of 4GB could provide 768 to 920 GB/s and 12GB total. so 58-85% higher bandwidth than a Vega 10.
If the Vega 20 clocks at around 2GHz, a 3-stack implementation would provide a similar bandwidth/GFLOP to Vega 10.
 
And significantly higher clocks on TSMC 7nm compared to GF 14LPP.
A Vega 10 clocked 30% higher than 1.5GHz (~2.0GHz) would perform at the same level as a 1080Ti or 2080.

I think you're right wrt the process offering much higher clocking headroom.

But while it's likely that Vega missed its frequency / performance targets, pretty sure consistently hitting > 25 % higher clocks also involve some redesign to the basic shader units. I'm willing to bet AMD didn't commit resources for this, given their statements so far and the target market

We should further expect from professionally oriented GPUs to be clocked very conservatively. Pure performance is just one variable in the many determining the products' success (viability rather, it's 2018 RTG we're talking bout). They've increased bandwidth, hugely beefed up the FP64 part, increased VRAM significantly. And they cannot change the software ecosystem overnight.
If they'd be binning chips to squeeze an extra 10-20 % shader performance, they'd be throwing alot of otherwise good chips away. All that to change just one variable
 
Where do those 35% come from?

I see 30% claimed by TSMC compared to 16FF+, but Vega 10 isn't made on that process. It's made on GF's 14LPP which according to various statements ( from @Nebuchadnezzar et al) is some 10-20% worse than TSMC's 16FF+.
So we're looking at ~45% to 55% higher clocks on Vega 20, at ISO power.
I saw it in an AMDs slide labelled „Leading Edge 7nm Process Technology“ here: https://www.tomshardware.com/news/amd-7nm-gpu-vega-gaming,37228.html
 
But while it's likely that Vega missed its frequency / performance targets, pretty sure consistently hitting > 25 % higher clocks also involve some redesign to the basic shader units. I'm willing to bet AMD didn't commit resources for this, given their statements so far and the target market

We should further expect from professionally oriented GPUs to be clocked very conservatively. Pure performance is just one variable in the many determining the products' success (viability rather, it's 2018 RTG we're talking bout). They've increased bandwidth, hugely beefed up the FP64 part, increased VRAM significantly. And they cannot change the software ecosystem overnight.
If they'd be binning chips to squeeze an extra 10-20 % shader performance, they'd be throwing alot of otherwise good chips away. All that to change just one variable

They don't have to use the same clocks for every SKU. The workstation Vega 10 cards (FE, WX9100, SSG) have different clock settings compared to the RX Vega 64.



I saw it in an AMDs slide labelled „Leading Edge 7nm Process Technology“ here: https://www.tomshardware.com/news/amd-7nm-gpu-vega-gaming,37228.html
That presentation is previous to GF dropping 7nm, so the 35% could be an average of the gains between the GF and TSMC's 7nm nodes over the current 14LPP.
They also don't mention if it's for turbo or base clocks.
Still, assuming those slides refer to a Pro version then we should compare them to the WX9100 with a 1500 MHz clock. 1.35 * 1.5 = 2025 MHz.
 
A Vega 10 clocked 30% higher than 1.5GHz (~2.0GHz) would perform at the same level as a 1080Ti or 2080.
Still need to figure out the primitive shader front end which could give a substantial boost in addition to clocks. If that allowed scaling to multiple chips, a viable part could be made with Turing's MSRP. Boosting throughput in addition to scaling wider with higher clocks would be inline with the Epyc strategy and be competitive if they got it working.
 
They don't have to use the same clocks for every SKU. The workstation Vega 10 cards (FE, WX9100, SSG) have different clock settings compared to the RX Vega 64.

Of course. But do pro skus vary their clocks wildly between them? Also here I expect they would try to limit the number of skus, as it is easier to market less products especially with customers not so price sensitive.

So yeah, maybe they could provide, after a few months of Vega 20, a limited edition Titan-inspired consumer card, with top binned cards 0higher clocks. That's as far as I'd expect them to go but I still think it's a stretch.

That presentation is previous to GF dropping 7nm, so the 35% could be an average of the gains between the GF and TSMC's 7nm nodes over the current 14LPP.
They also don't mention if it's for turbo or base clocks.
Still, assuming those slides refer to a Pro version then we should compare them to the WX9100 with a 1500 MHz clock. 1.35 * 1.5 = 2025 MHz.

That specific 35% slide mentions comparison between processes. Process performance could (and posibly rather would) refer to process parameters other than density and power. So it could reference current increases and other things relevant to electrical engineers but which only indirectly impact performance of chips designed on said process
Ofcourse, without more context it's impossible to tell.
I'd be shoked to see a 2ghz Vega 20 card, sold at a counter in any significant volume.
 
That presentation is previous to GF dropping 7nm, so the 35% could be an average of the gains between the GF and TSMC's 7nm nodes over the current 14LPP.
They also don't mention if it's for turbo or base clocks.
Still, assuming those slides refer to a Pro version then we should compare them to the WX9100 with a 1500 MHz clock. 1.35 * 1.5 = 2025 MHz.
This press release from AMD after GloFo announced their exit out of 7nm for now seems to indicate that AMD was either well aware of GloFos plans in advance or that they had planned initial 7nm products for TSMC fabrication anyway. So it would make little sense to include a purely hypothetical figure into their forecasts, lowering expectations or sandbagging.

I don't care either way anyway, since I have no stock in either AMD or their competition.
 
This press release from AMD after GloFo announced their exit out of 7nm for now seems to indicate that AMD was either well aware of GloFos plans in advance or that they had planned initial 7nm products for TSMC fabrication anyway. So it would make little sense to include a purely hypothetical figure into their forecasts, lowering expectations or sandbagging.

I don't care either way anyway, since I have no stock in either AMD or their competition.

Could be the other way around, AMD was the biggest client for Glofo 7nm, AMD have annonce to Glofo they was going TSMC whatever is the reason, and Glofo have cancel their 7nm. ( Better to focus on their other clients area )
 
Unlikely, due to the WSA agreement. Or rather the WSA bound both parties so they had to negociate their way out of it.

Also GF initiating the move made economical sense; their other processes were sucessful and needed less investement than bleeding edge
 
GF was going to get money from AMD regardless of whether AMD manufactured chips at TSMC, Samsung, or GF. So that wouldn't be a reason for them to ditch 7 nm.

GF having AMD as the main and potentially sole user of 7 nm, however, is a potential problem for them. Especially if their existing manufacturing lines are seeing increased use from other consumers, thus generating a profit. This is made more stark when you see how much difficulty other foundries are having moving down to smaller and smaller nodes and how incredibly expensive it is to do so. You really need manufacturing partners (TSMC with companies like Apple and NVidia and to a lesser extent AMD and others) to be able to afford the cost of moving to smaller and smaller nodes or be in a position where you yourself can benefit immediately and gain a competitive advantage in the marketplace (Samsung and Intel). GF has neither of those, so it's extremely difficult and costly for them to stay on the bleeding edge.

In that sense it may make sense for them to ditch 7 nm. In which case they'd have to go to AMD and ask AMD to redo the WSA. Something I'm sure AMD was more than happy to do.

Regards,
SB
 
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...


  • Radeon Vega with 96 and 128 CUs
  • $550 price point
  • Available between Thanksgiving and Christmas 2018
  • 8-12% higher performance than 2080
  • Lower power consumption than 2080
What would you like to bet, since all this can “easily” happen?
 
  • Radeon Vega with 96 and 128 CUs
  • $550 price point
  • Available between Thanksgiving and Christmas 2018
  • 8-12% higher performance than 2080
  • Lower power consumption than 2080
What would you like to bet, since all this can “easily” happen?

It would cost too much. The Vega 20 is targeted squarely at the high end AI market because they're the ones that can afford to pony up for a giant chip built on a just released node. The reason the 20 will cost so much is the same reason Nvidia didn't delay a few months to put Turing on 7nm. When new nodes come out error in building the chips is high.

If the error crops up in a non critical area you can "bin" chips, disable that part (a shader core, etc.) and sell it at a lower price. But if an error crops up in a critical area it could kill the whole chip. If your building relatively tiny chips, such as Apple's A12 at 83mm (squared), and a critical error shows up on average of once every 400mm, then a little over 20% of A12 chips would be defective. Bad but not the end of the world for trillion dollar company Apple. Now a Vega 64 is 486mm (squared) in die size. Even squeezing that to 7nm, under half size, that's still every other chip being defective, that's a lot of wasted money.

And these are just example numbers, which don't take into account things like transistor cost not going down for nodes anymore. So even on 7nm, if Vega 20 is just double a Vega 64, then it costs double to make it. Meaning there's no way AMD is hitting a reasonable price point for gamers with it, not including lost chips. As a time goes on in a node each silicon foundry refines its process and bring errors down, so it's conceivable we'll see Navi as a much smaller, mainstream chip ($200-$300) by the middle of next year, but Vega 20 isn't being released at any reasonable price point this year.
 
I see it the same Pony, but this should be added: unlike Apple products, some faulty graphics chips can be salvaged in cut-down products.
 
Back
Top