Discussion Com'on AMD, bring us a Vega II Nano

[AtenRa]

Fury Nano was released in Sep 2015 , it used the AMD Fury chip produced by TSMC 28nm and the card was only 175W TDP.
It was the first high-end small form factor Graphics Card using the new HBM memory.
There were two things kept the card from being a huge success, it only had 4GB of HBM memory in a time 4k was starting to matter and its MSRP was extremely high at $649 when the competition was faster at the same price (GTX 980Ti)

https://www.anandtech.com/show/9621/the-amd-radeon-r9-nano-review

Since then AMD havent released a Vega 14nm Nano card leveraging the increased perf/w characteristics of the new 14nm GloFo process.
Now with the new Vega II produced using the TSMC 7nm process, I believe its the time to bring the Nano back where its belong.

In order for the new Vega II Nano to have a better success it should have at least 8GB of memory and priced at $499 to compete against the RTX 2070.

So, vote if you would like to see a Vega II Nano brought back to the market by AMD.

 

[whm1974]

I would like a cheaper Nano priced at $250 or less.

 

[Stuka87]

As an owner is a SFF PC, I would love a new Vega VII based Nano. A full vega won't fit, regardless of price. But a nano, that is smaller would work great. Declock it down to 1.4-1.5GHz, and set the voltage right, and the power usage should drop like a rock.

 

[AtenRa]

Every Radeon VII and hypothetical Vega II Nano is one less MI50 sold, period.

Only if you produce number of Vega II cards that fit exactly the MI50/60 sales.

RX Vega 64 went through the same issue irrespective of the mining craze. There is a reason why it took so long to bring a considerable supply of Vegas to the consumer market at reasonable prices. Sure, you can get one today for less than $400. It's because they're finally obsolete! If you want AMD to be able to comfortable sell Vega II Nanos you would have to wait until 2020/2021 or so.

Vega 10 die size is 495mm2
Vega 10 is produced at 14nm at GloFo , at the same time AMD also had
Ryzen 1 produced at 14nm at GloFo
Ryzen 1 APUs for both desktop and Mobile were produced at 14nm at GloFo
Polaris GPUs (both desktop and mobile) were produced at 14nm at GloFo

Vega was low volume at the start because AMD was volume constrained at 14nm GloFo not because of the server.

Vega 10 = 495mm2 = 112 dies per wafer , with 80% yields back in 2017 we get only 90 dies
Vega 20 = 331mm2 = 170 dies per wafer , with 80% yields we get 136 dies or 50% more

Today AMD only producing Vega 20 at 7nm TSMC and because currently they dont have any other 7nm product on the oven, they can produce more than double or even triple the Vega 20 dies compared to Vega 10 had at launch.

 

[JasonLD]

Unless AMD can't launch Navi until Q4 or later, I really don't see the point of Nano when Navi should be able to fill the performance range where potential Vega II nano will fit.

 

[DrMrLordX]

Only if you produce number of Vega II cards that fit exactly the MI50/60 sales.

Um, no? There is no "exact" level of sales. Whatever they produce, they'll sell, at least between now and maybe late 2019/early 2020. Then sales will taper off in anticipation of new products. AI/deep learning is that hot.

Vega 10 die size is 495mm2

You might think that mattered, but it really didn't, for the same reasons listed above. AMD was able to sell every Vega10 they made up through mid 2018 or so. Shortages were not related to die size or wafer availability.

 

[Mopetar]

Every Radeon VII and hypothetical Vega II Nano is one less MI50 sold, period.

I don't think that's quite the case. I believe that some of the early specs were misreported and Radeon VII doesn't have the same absurdly good FP64 rate as AMD's Radeon Instinct cards.

If that's because AMD has disabled working hardware, then it's a case of a card that could have been sold as an MI50 no longer being sold as one and you'd be correct.

If, on the other hand, it's the result of defects that would prevent the card from having the rated capability of the MI50 for FP64 capabilities, then no, that card could not have been sold as an MI50 and AMD would have not choice but to throw it away.

Usually the people paying for those expensive compute cards do so for the 64-bit floating point computation power. This consumer version has a tiny fraction of it, so you'd need to buy something like 5 or 10 times as many (I've heard a few different figures for the actual FP64 rate, so it varies) Radeon VII cards to replace an equivalent number of MI50 (or MI60) cards.

I suppose if you don't need that or your workload uses other types of calculations that aren't hobbled compared to the Radeon Instinct cards, then Radeon VII is obviously a much better buy. I think AMD knows what people are in the market for though.

 

[DrMrLordX]

If that's because AMD has disabled working hardware, then it's a case of a card that could have been sold as an MI50 no longer being sold as one and you'd be correct.

When have we ever seen a GPU die sold that had FP64 disabled due to defects?

 

[Mopetar]

331mm2 is not that big of a die and yes I believe yields should be around 70% or better today at those sizes.

That seems optimistic unless you mean that the yields are 70% if you factor in dies that can be salvaged.

Otherwise, this implies that the yields for Zen 2 would be almost 93%. AMD would be getting around 700 functional 8-core chiplets per wafer. Unless they were still trying to fix a few minor issues with a new stepping, at those yields they should be shipping much sooner.

It's possible to think that maybe they want to finish some tweaks that that they think are necessary, which is what is holding back production, but I'm more inclined to think that the yields aren't quite that good yet. There's probably a good reason that Nvidia went with 12nm for their latest GPUs even though 7nm would have let them shrink them a lot smaller. My guess is that NVidia ran the numbers and even with the smaller dies, the yield wasn't good enough for the economics to work out as favorably.

 

[Stuka87]

When have we ever seen a GPU die sold that had FP64 disabled due to defects?

I don't think he meant disabling FP64 entirely, but disabling some of the FP64 related hardware so that the performance is lower than that of the Mi50. Although I am unsure as to if this is even an option.

 

[DrMrLordX]

I don't think he meant disabling FP64 entirely, but disabling some of the FP64 related hardware so that the performance is lower than that of the Mi50. Although I am unsure as to if this is even an option.

He actually did.

The issue at hand is this: AMD can sell ridiculous amounts of Vega20 dice as pro/AI cards for over $1k. So they have no financial incentive to sell any of them in consumer cards, except to keep the iron hot for their next gen consumer product (Navi). @Mopetar raised the possibility that AMD had some Vega20s with defects that made full FP64 performance unworkable (for whatever reason). Personally I've never heard of that happening before.

 

[Mopetar]

He actually did.

The issue at hand is this: AMD can sell ridiculous amounts of Vega20 dice as pro/AI cards for over $1k. So they have no financial incentive to sell any of them in consumer cards, except to keep the iron hot for their next gen consumer product (Navi). @Mopetar raised the possibility that AMD had some Vega20s with defects that made full FP64 performance unworkable (for whatever reason). Personally I've never heard of that happening before.

I don't know if it's even that likely to occur. There's certainly some part of the silicon that's necessary for the chip to process FP64 instructions, but unless it's a large enough part of the circuitry, the odds of any defect occurring in those locations becomes incredibly small. With the rumors of 20,000 Radeon VII cards being available, they would have had to produce hundreds of thousands or even millions of Vega 20 dies just to get 20,000 that happen to fail as MI50/MI60 products, but could still make a useful consumer GPU.

There are probably better explanations. I just find it useful when examining something to break down all of the different possibilities, even if some of them are really unlikely. It really doesn't make sense for AMD to take cards that could be sold for several thousand dollars and intentionally disable them to be sold as $700 consumer cards.

One other theory (based on a lot of rumor and speculation) was that prior to getting initial Navi silicon back, AMD had started ramping up Vega 20 to compete with Turing so that they'd have something. Navi was reported to be better performing than AMD had expected (and later Turing being released with more focus on ray tracing than really driving traditional performance) so they decided to shift away from using an Vega 20 dies in consumer products, but still had some extra dies they had in anticipation of needing Vega for a consumer card.

I'm not inclined to believe that what I speculated is the case. It's certainly possible, but I don't believe that AMD would naturally have enough dies with that type of problem. Even then it seems as though they could salvage those to sell another model (say MI40) that just has even fewer cores.

 

[DrMrLordX]

I just find it useful when examining something to break down all of the different possibilities, even if some of them are really unlikely.

That's fine, and it can be instructive to follow such trains of thought. I agree that it is unlikely that the tiny silicon defects that would make FP64 unworkable at 1/2 rate while leaving the rest of the card functional would be probable or even possible.

One other theory (based on a lot of rumor and speculation) was that prior to getting initial Navi silicon back, AMD had started ramping up Vega 20 to compete with Turing so that they'd have something. Navi was reported to be better performing than AMD had expected (and later Turing being released with more focus on ray tracing than really driving traditional performance) so they decided to shift away from using an Vega 20 dies in consumer products, but still had some extra dies they had in anticipation of needing Vega for a consumer card.

For AMD's sake, I hope you are right. There is also the possibility that Navi will be slower than RX Vega 64 at launch, making it a $100-$200 midrange card as an upgrade for Polaris. I would rather not see a Navi product that is that slow. There is still the "tile" rumor that AMD will be able to launch faster Navi products by just adding tiles, meaning that eventually Navi would catch up to and surpass Vega20 in gaming performance. But we know so little about Navi that it is impossible to speculate about that now.[/quote][/QUOTE]

 

[Mopetar]

That's fine, and it can be instructive to follow such trains of thought. I agree that it is unlikely that the tiny silicon defects that would make FP64 unworkable at 1/2 rate while leaving the rest of the card functional would be probable or even possible.

It should be possible, or rather is almost certainly has to be possible. If it weren't AMD wouldn't be able to disable performance for Radeon VII except through software, and there would be a massive incentive to patch the functionality back in since it's worth several thousands of dollars per card if you can do that.

However, I don't think there are a huge number of transistors devoted to that functionality. That makes the likelihood of them being defective low in itself. If it's only a few hundred transistors (maybe even on a per CU basis) that control this behavior, the odds become astronomically low.

To illustrate, imagine taking the entire state of Connecticut and sectioning it off into little one meter by one meter squares. If you do this, you'll end up with slightly more of those little squares than Vega 20 has transistors. Now drop a rock down from orbit and try to get it to land in one of something like 25,000 of those squares that are scattered all over the state. Carnival games would seem less rigged by comparison.

Even if I'm off by an order of magnitude (or two) you're still looking at an event with around one one-hundredth of a percent chance of occurring. Sure it can happen, but you'd need to make about 25,000,000 Vega 20 dies before you would have a 50% chance of seeing 5,000 dies with a defect that unlikely, even after assuming that I'm off by up to two orders of magnitude.

It was one of those things that seemed more plausible than it really was when I first thought about it, but after bothering to explore the idea a little more and scrutinize it a little, I'm possibly more likely to be killed by ball lightning inside of my house while thinking about whether or not it could happen than for it to actually happen.

 

[GodisanAtheist]

I think the RVII is really more of a tactical release by AMD, likely using the worst binned MI50 dies, to capitalize just a little bit on NV's RTX stumble.

AMD knows there is a short term opportunity cost in turning AI GPUs into Gaming GPUs, but the long term gains in keeping their brand fresh outweighs the cost.

To the topic at hand: NV has never released an official Nano version of their GPUs, but their AIB partners always find a way to fit even their 80ti tier of chips into remarkably small sized boards (even if it's not quite Nano sized).

Don't know why Nano needs to be an official AMD product and not something that board partners just relase on their own volition (like one of them did with Vega).

 

[Dribble]

I would very much like you to analyse the BOM cost of the Radeon 7 card that drove you to the conclusion they already sell them at a loss.

[Fudzilla] Radeon VII production cost $650+

I'm guessing it's a $5000 card in MI50 Instinct form so not surprising it's pretty expensive to make.

 

[AtenRa]

[Fudzilla] Radeon VII production cost $650+

I'm guessing it's a $5000 card in MI50 Instinct form so not surprising it's pretty expensive to make.

This article doesnt provide any source of its claim that 16GB of HBM 2 cost $320 to AMD.

 

[Stuka87]

This article doesnt provide any source of its claim that 16GB of HBM 2 cost $320 to AMD.

Pretty sure they took the story that said Vega56/64 8GB of memory cost $160, so obviously double that would cost $320, right?

Vega64 cost: https://www.fudzilla.com/news/graphics/43731-vega-hbm-2-8gb-memory-stack-cost-160

Vega VII cost: https://www.fudzilla.com/news/graphics/48019-radeon-vii-16gb-hbm-2-memory-cost-around-320

 

[AtenRa]

Pretty sure they took the story that said Vega56/64 8GB of memory cost $160, so obviously double that would cost $320, right?

Vega64 cost: https://www.fudzilla.com/news/graphics/43731-vega-hbm-2-8gb-memory-stack-cost-160

Vega VII cost: https://www.fudzilla.com/news/graphics/48019-radeon-vii-16gb-hbm-2-memory-cost-around-320

Heh, next year if AMD will release a 32GB HBM 2 card they will release a new article claiming the price of the memory will be $640