How much do GPU respins actually cost?

[JDG1980]

As we all know, AMD is continuing to sell designs from 2012 (Cape Verde and Pitcairn) as part of their "new" R9 300 and R9 M300 series. Depending on who you ask, these incessant rebadges are either because AMD has massive back stock they need to liquidate (and apparently no one on staff who has heard of the sunk cost fallacy) or because it is too expensive to do respins to get these chips up to par with modern technology like GCN 1.2, FreeSync, etc.

So what I'd like to know is, just how expensive is this? AMD seemed to have no trouble affording a new lineup in 2012, but they've been trickling them out since then: Bonaire in March 2013, Hawaii in October 2013, Tonga in August 2014... and that's about it. The process hasn't changed since 2012, so what has? Why is it so insanely expensive to do respins, when that kind of thing used to happen all the time? How much money would AMD have to shell out if they wanted to create a GCN 1.2 successor to Pitcairn (on 28nm), and where would the money go? I know the exact figures aren't going to be available, but hopefully someone knows the order of magnitude. Are we talking about millions of dollars? Tens of millions? Hundreds of millions? Billions?

 

[RussianSensation]

Some possible explanations:

1. The products you speak of are OEM cards. We haven't seen the full retail line-up, including the products designated with X (370X, 380X, 390, 390X, etc.). Just because some parts were re-brands, doesn't tell us that all parts are re-brands for the rest of R9 300 retail cards. Thus far, we don't even have any info on retail R9 300 desktop cards.

2. It's possible AMD is doing strategic cash allocation towards Post-GCN products in 2016 and beyond. They may have realized that they missed the boat on Maxwell and at this point they could decide to conserve the cash for an all-new post GCN 14nm/16nm HBM GPU line-up starting in Q4 2016 and beyond. I personally do not agree with such a business strategy, but I have seen certain forum members state this as a viable possibility. It may be the case that AMD didn't think it would be worthwhile to redesign the entire GPU stack of GCN products < $349 and will just reuse Tonga and Hawaii for those products. However, until we see the entire R9 300 retail desktop and mobile line-ups officially announced by AMD, it hasn't been 100% confirmed that 95% re-branding strategy is a fact.

Based on how AMD emphasized 2x perf/watt with next gen architecture on its slides during their Financial Analyst Day but hardly spoke of 2015 as the ground-breaking GPU year for its graphics division, it may very well be true that the focus this generation was mostly on R9 390/390X/395X2 and just minor updates to Tonga/Hawaii (i.e., a more mature 28nm node). If true, looks like their mobile dGPU strategy until 14nm/16nm is also DOA, which is brutal because it means little pressure for NV to release much faster GPUs to supersede 960M/965M/970M and 980M.

In 1-2 months all shall be revealed.

 

[jpiniero]

Enough that it wouldn't make sense financially. Performance isn't the problem; it's the power consumption. Overhauling the lineup with Tonga wouldn't really change things. Expect rebrands except for the HBM models.

 

[JDG1980]

Some possible explanations:

1. The products you speak of are OEM cards. We haven't seen the full retail line-up, including the products designated with X (370X, 380X, 390, 390X, etc.). Just because some parts were re-brands, doesn't tell us that all parts are re-brands for the rest of R9 300 retail cards. Thus far, we don't even have any info on retail R9 300 desktop cards.

I hope you're right. But the fact that AMD trotted out Cape Verde and Oland rebrands for laptops - and that the "refined efficiency and power management" touted in the FAD slides turned out to consist of downgrading from GDDR5 to DDR3 - does not make me optimistic on this front.

2. It's possible AMD is doing strategic cash allocation towards Post-GCN products in 2016 and beyond. They may have realized that they missed the boat on Maxwell and at this point they could decide to conserve the cash for an all-new post GCN 14nm/16nm HBM GPU line-up starting in Q4 2016 and beyond. I personally do not agree with such a business strategy, but I have seen certain forum members state this as a viable possibility. It may be the case that AMD didn't think it would be worthwhile to redesign the entire GPU stack of GCN products < $349 and will just reuse Tonga and Hawaii for those products. However, until we see the entire R9 300 retail desktop and mobile line-ups officially announced by AMD, it hasn't been 100% confirmed that 95% re-branding strategy is a fact.

I wouldn't take that KitGuru article at face value. The FAD slides make it clear that the 2016 die-shrinks are going to be GCN derivatives, not a new clean-sheet architecture. I'm sure there will be refinements and optimizations, but it will still be GCN. Nothing inherently wrong with that; Intel's newest CPUs have a lineage going back to the Pentium Pro (P6).

Based on how AMD emphasized 2x perf/watt with next gen architecture on its slides during their Financial Analyst Day but hardly spoke of 2015 as the ground-breaking GPU year for its graphics division, it may very well be true that the focus this generation was mostly on R9 390/390X/395X2 and just minor updates to Tonga/Hawaii (i.e., a more mature 28nm node). If true, looks like their mobile dGPU strategy until 14nm/16nm is also DOA, which is brutal because it means little pressure for NV to release much faster GPUs to supersede 960M/965M/970M and 980M.

In 1-2 months all shall be revealed.

Tonga would be a reasonably decent competitor to GTX 960 if they started reserving decent silicon for the desktop; the one thing really missing is a HEVC decoder, which the 960 has. Hopefully it would be less trouble to add this (when porting over to GloFo?) than to do a complete redesign; AMD has one for Carrizo. What I'm trying to get a picture of is how much these specific moves cost. What does it cost AMD to do a minor respin of an existing chip?

But beyond Tonga, AMD doesn't have anything worthy of being part of the R9 300 series. The fundamental problem with Hawaii is that it is too big a chip, with too complicated a PCB and power delivery system, to sell profitably at the prices AMD and its AIBs can charge. The market has priced Hawaii at $249-$329, and I don't see how this is profitable. Cut out the Double Precision hardware (which is gimped on consumer Radeons anyway), increase the shader count to 3072, update the architecture to GCN 1.2, and reduce the bus width to 384-bit and I think they could make it work. But right now it looks like AMD is doing straight, lazy rebadges of as many products as they can.

 

[JDG1980]

Enough that it wouldn't make sense financially.

That's not really an answer, it's just a reiteration of AMD's (apparent) conclusion. What I want to know is the rough figures that went into determining that. How much would respins cost, to an order of magnitude?

Performance isn't the problem; it's the power consumption. Overhauling the lineup with Tonga wouldn't really change things.

You're underestimating Tonga because you're looking at the R9 285, which is trash silicon. The R9 M295X (full Tonga) in the Retina iMac pulls only 125W or so on full load, and clocks are only slightly lower than the desktop part. Getting a desktop full Tonga with 4GB to fit into 150W (one 6-pin connector) should be doable. That's competitive perf/watt with the GTX 960. And this is without considering the claims (I'm not sure if they are true) that GloFo's 28nm SHP process is more efficient for GPUs than TSMC's.

 

[ShintaiDK]

Price also depends on size. And its not a respin, but a new design. Assuming AMD can more or less automate the process, I would guess 50-75M$ for a GCN 1.2 Pitcairn. Design, mask, validation etc.

But you should ask IDontCare in the CPU section, he will be able to give a relatively accurate number.

 

[Qwertilot]

What has changed since 2012 is relatively easy to say - TSMC's 20nm process fell flat on its face for dGPU's

AMD's original plan - on which they'll have spent plenty of R&D money - will presumably have been a total refresh on 20nm sometime late last year/into this one.

A major set back for both companies of course, but its the sort of thing where having lots of R&D spending - some of it presumably going to making sure you've got back up plans! - really shows itself as useful. NV are doing well, AMD really not. Hence 28nm Maxwell and AMD's maybe slow reaction.

Also, if AMD can hit solid competition with the 960 with the 380/X by tweaking Tonga, and TitanX/980ti with the 390/X then are things that dire objectively?

Much more sensible line up than currently at least. Nothing much vs the 970/980 perhaps, but the time to fight them was 9 months ago, not now At this point you'd need to be actively better which would be very hard.

 

[ShintaiDK]

There is no hiderance as such in making 20nm dGPUs. Its just about cost.

 

[JDG1980]

Price also depends on size. And its not a respin, but a new design. Assuming AMD can more or less automate the process, I would guess 50-75M$ for a GCN 1.2 Pitcairn. Design, mask, validation etc.

But you should ask IDontCare in the CPU section, he will be able to give a relatively accurate number.

Just curious, where exactly is the line drawn between respins and new designs? For example, if Tonga's UVD block was updated to include HEVC but the rest of the chip is left unchanged, is that considered a respin or a new design?

 

[JDG1980]

Also, if AMD can hit solid competition with the 960 with the 380/X by tweaking Tonga, and TitanX/980ti with the 390/X then are things that dire objectively?

Much more sensible line up than currently at least. Nothing much vs the 970/980 perhaps, but the time to fight them was 9 months ago, not now At this point you'd need to be actively better which would be very hard.

Nvidia's current Maxwell lineup consists of six cards (750, 750Ti, 960, 970, 980, Titan X). We've already penciled in a seventh (cut-down GM200), and it's possible there may also be a 960Ti at some point in the future. For AMD to compete with only three out of seven spots is stretching things pretty thin.

In my opinion, AMD needs two additional chips, on top of Fiji and a slightly modified Tonga (with updated UVD for HEVC, and HDMI 2.0). One should be an updated replacement for Hawaii, with GCN 1.2+, 3072 shaders, 384-bit bus, and 6GB GDDR5. This would be put up against the 970 (and 980?); even if it wouldn't beat them decisively, it would at least be cheaper to make due to the reduced power and board routing requirements, and could win on perf/$ without Hawaii's embarrassingly high power consumption levels. The other new chip should be designed to fit into a 75W power envelope (no PCIe power connector), in order to compete with the 750 and 750Ti in the OEM upgrade and HTPC markets. It would also be GCN 1.2 and should cram in as many shaders as possible within that limit - maybe they could stretch to 1024? They should at least be able to match Bonaire's 896. The bus width should be 128-bit if that provides acceptable performance, otherwise 192-bit. The 750 and 750Ti sell quite well, and AMD has had no response to them for a very long time. AMD's best card without an external power connector is the old Cape Verde rebrands, and these just can't even come close to stacking up to Maxwell.

Yes, this would mean some R&D spending, but AMD can't abdicate so many spots in the GPU market for 18 months. I don't see how doing so is even survivable. They lose way too much mindshare and marketshare.

 

[ShintaiDK]

Just curious, where exactly is the line drawn between respins and new designs? For example, if Tonga's UVD block was updated to include HEVC but the rest of the chip is left unchanged, is that considered a respin or a new design?

A respin is usually fixing bugs or slightly improve electricals.

 

[Qwertilot]

Yes, this would mean some R&D spending, but AMD can't abdicate so many spots in the GPU market for 18 months..

I'm not saying this is what they want to do it might be what they've found themselves having to do. They've already basically gone missing from mobile GPU's for quite a long time now - which isn't something anyone sane would choose to do either

I don't see how doing so is even survivable. They lose way too much mindshare and marketshare.

The market share is, I fear, mostly gone already for many of those segments. The 750ti has had ~15(!) months, and the 970/80 8+, Xmas thrown in All basically to themselves.

No doubt some continuing sales but a majority of the people who wanted a card with those sorts of characteristics will have already got one. Even if they do manage to put your (very sensible!) 'ideal' chips it'd serve more to stem the bleeding than anything else.

Might have been too much money for them to try, or maybe they have done some of it. We'll see

Happily the next gen looks to be set to be enough of a jump to motivate a lot of upgrades and so a lot of market share back in play so long as their mind share is broadly intact/they show up this time! With a full line up.

390/X crucial for the mind share of course. I'm sure they'd prefer it to be more like 12 than 18 months.....

 

[tviceman]

I think we're going to get 3 Fiji desktop SKU's, an additional Tonga sku (for a total of 2), and I don't know about the rest.

 

[Silverforce11]

I think we're going to get 3 Fiji desktop SKU's, an additional Tonga sku (for a total of 2), and I don't know about the rest.

I think 3 Fiji SKU is a solid prediction. Such a big die it should be able to harvest twice for two extra lower variants. That would fill the performance & $ gap from top to mid-range.