[Rumor, Tweaktown] AMD to launch next-gen Navi graphics cards at E3

[swilli89]

I learned my lesson to not speculate about yields. And that is exactly what AdoredTV is doing. He is speculating about them.

I can't believe that guy is getting thrown around here still. He is still taken seriously? He had a massive amount of cake on his face from getting Ryzen 3000 so extremely wrong that he revealed he has absolutely no real sources, all just conjecture that isn't even that educated in my opinion.

Guys - you're arguing over semantics and definitions that can be easily re-made or perhaps have never existed in the first place. You can take a look at die sizes and maybe make an educated guess about wafer costs... but it doesn't matter. Price points are the only thing that obviously don't shift. Navi 10 COULD be sold by AMD has a midrange card at $300, but they chose to market it as a 2070 competitor and priced it such. And unsurprisingly they will sell much less... and AMD knows this of course!

They might be restricted on the amount of wafers they are allocated by TSMC - and if that is the case AMD will want to maximize the amount they get per chip. You guys act like the decisions these companies make are transparent and easily understandable, yet in the end all we can agree on is they are attempting to maximize their profits through their pricing and market segmentation strategies.

AMD felt that charging $449 for Navi 10 was their way to maximize revenue/profits and so they did. We can argue about what constitutes "mid-range" til we are blue in the face but it doesn't matter at all. What I CAN say for fact is that no, $449 is not a mid-range price point.

This should end the discussion - Navi 10 is a historically "mid-range" chip size (complete with traditionally mid-range 256-bit GDDR memory interface) that AMD is pushing as an enthusiast card. In that way everyone is right. From AMD's perspective its a mid-range chip in terms of cost to fabricate, and from consumer's perspective it is definitely not mid-range. Lastly, I'm not saying I agree that Navi 10 is priced the way it is, I like AMD as a company yet they don't put bread on my table, I still have to shell out for gaming related products at the end of the day so I would have LOVED to see 5700XT come in at $349. I saw this coming and when my 1060 couldn't perform in Apex Legends like I wanted, I went out and grabbed a used 1070 Ti for $200. Value has eroded like crazy for the end-user due to AMD and NV raising prices. Its the new normal! Definitions and expectations sadly have to be reset. Take it for what it is. At least AMD stopped Intel from charging a grand for 8-core processors and now we can get this performance level near $300 so its give and take.

 

[Ajay]

They might be restricted on the amount of wafers they are allocated by TSMC - and if that is the case AMD will want to maximize the amount they get per chip. You guys act like the decisions these companies make are transparent and easily understandable, yet in the end all we can agree on is they are attempting to maximize their profits through their pricing and market segmentation strategies.

I think it is very likely that AMD has a wafer allocation issue given that 7nm is still ramping and that there are alot of customers who want to fab 7nm designs at TSMC. AMD would likely want to shift the majority of their allocation to Zen2 chiplets. Now, TSMC seems to be a pay to play fab, that is, if you are willing to pay more, you get more wafers (as was the case with Apple) - so that's an option for AMD.

 

[Stuka87]

I think "midrange" is the wrong term to use.

RX 480/GTX 1060 were "Mainstream"
Vega/GTX 1080 were lower end "Enthusiast"

GTX 1080Ti fell under higher end enthusiast.

 

[swilli89]

I think it is very likely that AMD has a wafer allocation issue given that 7nm is still ramping and that there are alot of customers who want to fab 7nm designs at TSMC. AMD would likely want to shift the majority of their allocation to Zen2 chiplets. Now, TSMC seems to be a pay to play fab, that is, if you are willing to pay more, you get more wafers (as was the case with Apple) - so that's an option for AMD.

I agree its a wafer allocation issue. But we need to get away from talking about 7nm ramping.

TSMC was fabbing Vega 20 dies in October 2018. They were fabbing 7nm chips for Apple this time last year. 7nm may not be a completely mature node at TSMC, but its been at full volume ramp for many months. Right now yields aren't an issue, but wafer capacity is. And like you said, AMD is ramping Zen2 right now, so despite these being smallish dies, there is a cost of opportunity and AMD will want to maximize production of much higher margin products like Ryzen 3900X and EPYC parts.

Take a look at NVIDIA - they went Samsung 7nm EUV because TSMC is charging so much for 7nm wafers. EETimes cited multiple sources claiming Samsung simply undercut TSMC on bids for 7nm customers. TSMC knows they are the only game in town for 2019 volume orders for 7nm designs.. and they are charging accordingly. So its not so much 7nm ramping, but TSMC not having enough production lines switched to 7nm. Or.. it could just be entirely that TSMC is gouging everybody since there aren't other options until next year (Samsung 7nm EUV).

So we need to throw out way we know in terms of costs when it comes to talking about Navi's 251mm2 die size. For all we know they could cost the equivalent of a 400mm2 chip at 16nm.

 

[railven]

I think "midrange" is the wrong term to use.

RX 480/GTX 1060 were "Mainstream"
Vega/GTX 1080 were lower end "Enthusiast"

GTX 1080Ti fell under higher end enthusiast.

Pretty sure we had this rewording debate/spat of the terminology in 2012. Now the other shoe fell and everyone is being squeezed.

Plain and simple truth: we lost.

I was hoping that $1300 was the ceiling, but now - I'm not so sure. Definitely the ceiling for me (unless I get a surprise raise/promotion/XMas Bonus, then screw you all! I got mine! haha)

 

[Stuka87]

Take a look at NVIDIA - they went Samsung 7nm EUV because TSMC is charging so much for 7nm wafers. EETimes cited multiple sources claiming Samsung simply undercut TSMC on bids for 7nm customers. TSMC knows they are the only game in town for 2019 volume orders for 7nm designs.. and they are charging accordingly. So its not so much 7nm ramping, but TSMC not having enough production lines switched to 7nm. Or.. it could just be entirely that TSMC is gouging everybody since there aren't other options until next year (Samsung 7nm EUV).

Do we know nVidia went Samsung for cost? I have not seen this anyplace. For all we know TSMC doesn't have any capacity available and nVidia was then forced over to Samsung.

 

[linkgoron]

I think "midrange" is the wrong term to use.

RX 480/GTX 1060 were "Mainstream"
Vega/GTX 1080 were lower end "Enthusiast"

GTX 1080Ti fell under higher end enthusiast.

Clearly we should change the naming, the 2060 should be lower end enthusiast, the 2070 should be midrange enthusiast, the 2080 should be high end enthusiast and the 2080ti should be bleeding edge enthusiast.

14nm was also a "new" and expensive node when Polaris was released, and yet it was released @ $240. I'm sure we could say the same for every node shrink, but somehow a 251mm^2 die is now $400. Nvidia have shown in the last ~5 years that people will pay for more performance (people continuously claim that Nvidia just have to increase prices because of soaring R&D prices or wafer prices or huge dies or GDDR5X or GDDR6, and yet they're enjoying record profits), and AMD just want a piece for themselves. AMD could fight for market share, but they (sadly) prefer to go for margins. I do tend to agree that a lot of people just want AMD to compete in order to get cheaper Geforce cards, but that is besides the point.

 

[swilli89]

Clearly we should change the naming, the 2060 should be lower end enthusiast, the 2070 should be midrange enthusiast, the 2080 should be high end enthusiast and the 2080ti should be bleeding edge enthusiast.

14nm was also a "new" and expensive node when Polaris was released, and yet it was released @ $240. I'm sure we could say the same for every node shrink, but somehow a 251mm^2 die is now $400. Nvidia have shown in the last ~5 years that people will pay for more performance (people continuously claim that Nvidia just have to increase prices because of soaring R&D prices or wafer prices or huge dies or GDDR5X or GDDR6, and yet they're enjoying record profits), and AMD just want a piece for themselves. AMD could fight for market share, but they (sadly) prefer to go for margins. I do tend to agree that a lot of people just want AMD to compete in order to get cheaper Geforce cards, but that is besides the point.

Its widely understood that beginning at 28nm, cost benefit to switching to new nodes began to erode.

Actually, for someone on 1080p, they have been able to spend about $250 and the ability to play 60FPS or more at 1080P easily for the last 10 years. So yes consumers haven't got relative performance increases at a given price bracket that we got the decade prior, the ability to play modern games at a given price point has remained constant. Monitors creeping up to 4k and increasing frame rate limits have played a huge part in the demand for faster graphics cards.

Do we know nVidia went Samsung for cost? I have not seen this anyplace. For all we know TSMC doesn't have any capacity available and nVidia was then forced over to Samsung.

Toms EETimes
From the reports linked:
"Separately, one source said that Samsung is aggressively undercutting prices for its 7-nm node with EUV, offering some startups a full mask set for less than a multi-layer mask (MLM) set at its rival. TSMC introduced the MLM mask sets in 2007 to lower costs for small-volume runs. They are said to be about 60% of the cost of a full mask set. "

TSMC has the capacity, but at what cost that comes at is unknown. We know AMD is using up capacity for both graphics and CPU chips, and we know of several other customers of TSMC's 7nm wafers. If the above report is to believed, Samsung has been offering up to a 40% discount versus TSMC when it comes to mask sets. This, coupled with cheaper wafers, might have been too attractive for NVIDIA to pass up. And they would be in a position to risk switching to Samsung since AMD will, at best, equal their Turing with their eventual full stack of Navi 7nm products using TSMC.

As the chart above illustrates.. 5700 series are a direct market replacement for Vega series. They are several brackets up from where Polaris occupied. Feels similar to the 7970 launch that saw AMD increase its flagship price by 40%.

 

[ozzy702]

That chart is wrong for at least the 590... for "GCN" and "14nm".

 

[DrMrLordX]

I learned my lesson to not speculate about yields. And that is exactly what AdoredTV is doing. He is speculating about them.

Then don't speculate.

The slide posted by Crisium provided some good data to infer further.

https://forums.anandtech.com/thread...vi-graphics-cards-at-e3.2564009/post-39850637

If yielded costs are 2X from14/16nm to 7nm and yield should (must ?) be lower for 7nm, then costs of wafer is less than 2X increase, maybe a lot less (33% less?).

RTX2060 445mm^2 ~= RX5700 251mm^2 in cost to fab?
Nvidia uses custom TSMC 12nm process. (safe to assume is not cheaper than generic 16nm previously used)
RTX2060 = 6GB vs RX5700 = 8GB.

This seems to be the main difference in cost between the two cards. This could indicate that AMD is getting as high a margin as Nvidia on those two competing products, or at least very close.

Thanks. Food for thought . . .

I seem to remember a time when companies did talk about this, and I'm sure that TSMC and Samsung would not keep such information from their shareholders - are there not TSMC/Samsung Financial Analyst / investor Days that deal with these questions of yield and such?

There might be, but I haven't seen coverage of those in the forums to the extent that we have/had coverage of Intel investor days.

I agree its a wafer allocation issue.

Why? Wafer demand is down from the cell phone/tablet sector. AMD should have all the allocation they need, possibly at a discount.

For all we know they could cost the equivalent of a 400mm2 chip at 16nm.

If what @maddie posted is true, then see above.

 

[swilli89]

Why? Wafer demand is down from the cell phone/tablet sector. AMD should have all the allocation they need, possibly at a discount.



If what @maddie posted is true, then see above.

If you were TSMC, and you knew AMD depended solely on you, would you offer them discounts? AMD will order as much as they think they can sell, no more.

according to chipwiki "The 7nm node will come in two variants, one optimized for mobile applications and a second one optimized for High performance applications. "

So I'm not sure if production lines can produce either varient of the 7nm process, but TSMC uses a different version of their 7nm process for mobile as compared to GPU/CPU for PCs.

 

[DrMrLordX]

If you were TSMC, and you knew AMD depended solely on you, would you offer them discounts? AMD will order as much as they think they can sell, no more.

Only if TSMC had a lot of spare capacity at the time of contract negotiation. As for AMD "ordering only what they can sell" . . . I guess if they had this price in mind from the start, I wouldn't order very many either! But that is not wafer constraint. That is the marketing department insufficiently aggressive in their pricing.

 

[swilli89]

Only if TSMC had a lot of spare capacity at the time of contract negotiation. As for AMD "ordering only what they can sell" . . . I guess if they had this price in mind from the start, I wouldn't order very many either! But that is not wafer constraint. That is the marketing department insufficiently aggressive in their pricing.

That's a fair point. AMD would be able to sell more if they reduced their prices, which they could do with a price cut from TSMC.

However, given what I've seen reported, TSMC isn't giving out any 7nm discounts any time soon.

 

[crisium]

As the chart above illustrates.. 5700 series are a direct market replacement for Vega series. They are several brackets up from where Polaris occupied. Feels similar to the 7970 launch that saw AMD increase its flagship price by 40%.

Then what's the market replacement for Polaris?

3 full calendar years, and if you bought a RX 480 for $240 then your upgrade path for the same price is the... ~15% faster 590? Really? No one would ever ugprade from 480 to 590, and rightfully so.

And 5700XT will probably only be ~20% faster than the Vega 64. That's no upgrade either, even if you pretend it's a replacement towards the much larger Vega.

Years have passed, and AMD has no upgrade path for the same price.

Nvidia did the same with the 2000 series, as they moved up price brackets. I guess that's the good thing we can say about the RX 5700 series - they are same as Nvidia in not offering an upgrade path to users who have bought their cards 2-3 years ago. Though even the 1660 Ti is at least 35% faster than the 1060 for a similar price. Not a worthy upgrade, imo, but comapred to what AMD is offering to their 2-3 year old cards? Yeah, Nvidia is actually doing a little better.

 

[DisEnchantment]

I am not sure we can rely on all these HBM, GDDR5/6 cost being thrown around by these youtube experts.
From personal experience (working in a big corporation manufacturing high volume non consumer systems ) big ODMs don't buy RAM, NAND, NOR, Display, DSPs, regulators etc... from the market.
The silicon supplier comes to us, brings their catalog, give a presentation and sometimes a demo if it is a major component like a new SoC ..... loooong before the parts appear on their website.
Negotiations continues for months and if we decide to buy (after reviewing a bunch of bids, company policy forbids buying without a proper bid process) we sign a contract to supply. The components gets shipped from their assembly (some cases via a third party ) to our logistics depot from where it goes to our assembly line.
The cost are nowhere near market prices that is all I can say. And sometimes even when the supplier wants to end production due to newer part coming out they cannot because we have long term contracts.
You can even get customized packaging for the chips, different operating temperature ranges (which is important to us).
Sometimes we even buy before the chips are produced which in some cases the supplier does not even know the price it will be on the market when it launches. In such cases we sometimes pay ridiculously low price per piece and sometimes higher but rarely crossing market price. (If someone is paying market for these parts then our auditing and purchasing will likely have a word with the guys who authorize such a thing)

 

[Kenmitch]

That's a fair point. AMD would be able to sell more if they reduced their prices, which they could do with a price cut from TSMC.

A CPU/GPU bundle deal would be better than going into a price war scenario. AMD would get more sales while eliminating a potential sale to both Intel and Nvidia at the same time. Price war strategy only lowers pricing overall and doesn't guarantee a sale at all.

 

[tviceman]

Then what's the market replacement for Polaris?

3 full calendar years, and if you bought a RX 480 for $240 then your upgrade path for the same price is the... ~15% faster 590? Really? No one would ever ugprade from 480 to 590, and rightfully so.

And 5700XT will probably only be ~20% faster than the Vega 64. That's no upgrade either, even if you pretend it's a replacement towards the much larger Vega.

Years have passed, and AMD has no upgrade path for the same price.

AMD's own slides says the Navi 10 card shown on stage is 14% faster than Vega64.

 

[mohit9206]

A CPU/GPU bundle deal would be better than going into a price war scenario. AMD would get more sales while eliminating a potential sale to both Intel and Nvidia at the same time. Price war strategy only lowers pricing overall and doesn't guarantee a sale at all.

Any reason why AMD never does CPU+GPU bundles?

AMD's own slides says the Navi 10 card shown on stage is 14% faster than Vega64.

Isn't that pathetic? You would expect a 40% jump if its supposed to replace Vega 64. But we all know its a Polaris 590 replacement despite what some others here say.

 

[Ajay]

That's a fair point. AMD would be able to sell more if they reduced their prices, which they could do with a price cut from TSMC.

However, given what I've seen reported, TSMC isn't giving out any 7nm discounts any time soon.

IIRC, TSMC just qualified 7N for HVM late Q1 this year (Apple ran with early risk production - ie really low yields). TSMC will have ~100 tape outs this year - pretty confident that there are no discounts. D0 is turning out to be a problem, especially with > 300 mm^2. They were having issues with D0 defect rates on < 100 mm^2 early in the year.

If AMD is sales limited, then they, obviously, won’t have an allocation problem - that’s a bit of bad news.

 

[Ajay]

Any reason why AMD never does CPU+GPU bundles?

No idea, not a marketing guy, seems like it would be a great Black Friday deal.

 

[tajoh111]

That's a fair point. AMD would be able to sell more if they reduced their prices, which they could do with a price cut from TSMC.

However, given what I've seen reported, TSMC isn't giving out any 7nm discounts any time soon.

Rule 1 of any foundry is you never run it with empty capacity. The highest cost of running and building a fab, is the depreciation expense from being obsoleted in a few years. Running them empty does nothing but increase their own cost per wafer since the vast majority of cost is a fixed cost, not variable cost.

Ru

Then what's the market replacement for Polaris?

3 full calendar years, and if you bought a RX 480 for $240 then your upgrade path for the same price is the... ~15% faster 590? Really? No one would ever ugprade from 480 to 590, and rightfully so.

And 5700XT will probably only be ~20% faster than the Vega 64. That's no upgrade either, even if you pretend it's a replacement towards the much larger Vega.

Years have passed, and AMD has no upgrade path for the same price.

Nvidia did the same with the 2000 series, as they moved up price brackets. I guess that's the good thing we can say about the RX 5700 series - they are same as Nvidia in not offering an upgrade path to users who have bought their cards 2-3 years ago. Though even the 1660 Ti is at least 35% faster than the 1060 for a similar price. Not a worthy upgrade, imo, but comapred to what AMD is offering to their 2-3 year old cards? Yeah, Nvidia is actually doing a little better.

One more thing to mention is when the rx480 launched at $239, it launched at near 14nm peak wafer pricing. Wafer costs for TSMC was $4700(the cost for TSMC not companies using the wafers), with a 45% margin, they were charging $8400 which is much closer to the pricing 12500 figure being thrown around for 7nm. Prices have fallen now for wafers 14/16nm wafers so the difference is larger, but if AMD/Nvidia could make profits at 240/250 with wafers at $8400, you know AMD is making bank at $450 with wafers pricing at $12500.

 

[Guru]

It's barley better than NV in value and hence not good enough because of brand value or lack thereof. no buyer preferring NV due to brand will switch to AMD with this pricing. The card will only sell to AMD fans. low volume. I don't get it. With low volume it will be hard recuperating R&D costs regardless of margin.

It's even worse if NV super rumors are true. that will require AMD to lower price shortly after release. Better to undercut them right now and make the super-line DOA.

I really don't get it. The pricing only makes sense if:

1. 7nm yields are that bad (but then the Radeon VII not really be possible)
2. Cartel...
3. Supply limit due to high Zen 2 chiplet demand

They will recuperate their R&D money from the PS5 and Xbox 4, most of their 7nm is going into zen 2 and so they have no need for big market share in the desktop gpu space, they would rather have higher margins. The volume will come from ps5 and xbox4, those consoles together will sell about 200 million units judging from the previous generation sales.

Again their new design is very small and lean, so they have a huge and wide margin to lower prices. They can probably lower their prices by 40% on these chips and still make decent profits on them. RX5700XT chip is barely larger than Polaris RX 480/580 chip.

 

[jpiniero]

you know AMD is making bank at $450 with wafers pricing at $12500.

I dunno about the actual wafer cost, kinda thinking it's more than that; but you also have to take into consideration that defect rates/yields at 7 nm are worse.

 

[DrMrLordX]

I dunno about the actual wafer cost, kinda thinking it's more than that; but you also have to take into consideration that defect rates/yields at 7 nm are worse.

Didn't we hear earlier this year that defect rates on Zen2 chiplets were around 70%?

 

[Glo.]

Didn't we hear earlier this year that defect rates on Zen2 chiplets were around 70%?

With very small die, like Zen 2 is, 70% yield is... terrible news for large die like Navi 10.