ISTR reading that all else being equal, NV pays less per die/wafer than AMD. Can anyone corroborate this, or am I mistaken?
TSMC Wafer Pricing - NV vs. ATi
- Thread starter ShaidarHaran
- Start date
Wafer pricing for very large customers is always an interesting question. Obviously it always depends on order priority when in a supply constrained environment, as per the basic laws of supply and demand, but that much is obvious so let's focus on other aspects. Look at the wireless business: some companies like Atheros and CSR have been very conservative on process nodes, keeping some key products on 130nm for a very long time and really optimising every bit out of that process. Others, such as Broadcom, switched to 65nm much more aggressively.
From TSMC's point of view, they need both: some early adopters to push the new process along as well as justify the investment, and some laggards to maximise utilisation of the old nodes. I don't know anything for sure here, but I'd certainly expect CSR to have been proposed better 130nm/90nm pricing than Broadcom at one point, and Broadcom getting preferential 65nm rates in the early days of the node. Different deals for different customer requirements, with complex feedback effects on strategic decisions at every level of the ecosystem. You need to take such things into account when you consider what NVIDIA gained from being very aggressive process-wise in its early days, and what they gained from being much more conservative from 90nm to 55nm.
Then we get into complex political considerations, such as potential exclusivity deal, both formal and informal... What would it be worth to TSMC for Jen-Hsun to promise Morris Chang, who's also a personal friend, that they'll stick to 100% TSMC on both 40nm and 28nm? Quite a bit, I suspect. Similarly, what might TSMC put on the table for AMD to off-source the Ontario 'Fusion APU' to their 40nm process? It's not possible to simply abstract away such things, and you'd be foolish to think you could put a number on it without very deep insider knowledge.
It is not at all plausible that NVIDIA has systematically paid less than AMD on wafer costs to TSMC. But it's certainly very likely that they have at some points in time, and the same is true the other way around. What about 40nm? I don't know, and neither does anyone else. Jen-Hsun Huang said in a 2009 CC (Q3?) that they were paying lower 40nm wafer prices than just about everyone else (I don't remember the exact phrasing), and that would probably be based on a direct statement to him by Morris Chang. But that was for the very early ramp, where NVIDIA was moving more wafers than AMD (that says more about RV740 than NV really) and presumably helping kick-start the process. There's little reason to believe that must necessarily still be true today, and whether it is would probably depend on entirely different reasons.
Do remember to have some perspective, though. It's very unlikely to ever be a massive difference.
From TSMC's point of view, they need both: some early adopters to push the new process along as well as justify the investment, and some laggards to maximise utilisation of the old nodes. I don't know anything for sure here, but I'd certainly expect CSR to have been proposed better 130nm/90nm pricing than Broadcom at one point, and Broadcom getting preferential 65nm rates in the early days of the node. Different deals for different customer requirements, with complex feedback effects on strategic decisions at every level of the ecosystem. You need to take such things into account when you consider what NVIDIA gained from being very aggressive process-wise in its early days, and what they gained from being much more conservative from 90nm to 55nm.
Then we get into complex political considerations, such as potential exclusivity deal, both formal and informal... What would it be worth to TSMC for Jen-Hsun to promise Morris Chang, who's also a personal friend, that they'll stick to 100% TSMC on both 40nm and 28nm? Quite a bit, I suspect. Similarly, what might TSMC put on the table for AMD to off-source the Ontario 'Fusion APU' to their 40nm process? It's not possible to simply abstract away such things, and you'd be foolish to think you could put a number on it without very deep insider knowledge.
It is not at all plausible that NVIDIA has systematically paid less than AMD on wafer costs to TSMC. But it's certainly very likely that they have at some points in time, and the same is true the other way around. What about 40nm? I don't know, and neither does anyone else. Jen-Hsun Huang said in a 2009 CC (Q3?) that they were paying lower 40nm wafer prices than just about everyone else (I don't remember the exact phrasing), and that would probably be based on a direct statement to him by Morris Chang. But that was for the very early ramp, where NVIDIA was moving more wafers than AMD (that says more about RV740 than NV really) and presumably helping kick-start the process. There's little reason to believe that must necessarily still be true today, and whether it is would probably depend on entirely different reasons.
Do remember to have some perspective, though. It's very unlikely to ever be a massive difference.
So the answer is no. Thank you very much.
Silent_Buddha
Legend
Yes but once you get over a certain number of units AND you are supply constrained as TSMC is, then prices will level out. After all, if one vendor doesn't order as many due to price, the demand is high enough than any excess wafers due to that will sell to someone else.
Regards,
SB
Regards,
SB
Wow - just plain wow! How you just compressed a very elaborate and multi-faceted answer into a single word is just astonishing.
Wow - just plain wow! How you just compressed a very elaborate and multi-faceted answer into a single word is just astonishing.
I realize Arun's answer was rather in-depth, but I gathered that in the end the answer to the question is that NV likely does not pay less per wafer/die than ATi. If I'm mis-interpreting by way of over-simplification please feel free to correct me.aaronspink
Veteran
Well lets through some dirt into the conversation, there is less and then there is LESS. $1 on a $3-5K item isn't really a big deal. .5-1.5K is. I think any advantage is closer to the former than the later.
That makes sense. I suspect NV does pay less per wafer, but the amount is likely negligible, or darn close to.
Exactly.
Naively yes. But when hundreds of millions of dollars are changing hands do you really think the potential for "flexibility" is in the $1 on $3-5K range? Dell sure paid the same as everyone else for those Intel CPUs didn't they?
Silent_Buddha
Legend
Dell sure paid the same as everyone else for those Intel CPUs didn't they?
Yes but at the time Intel not only weren't supply constrained but they were faced with a competitor with a superior product (Athlon 64/Pentium 4 and Opteron/Xeon) giving Dell a whole lot of negotiating leverage as they were the number 1 OEM at the time.
The situation here is completely different. Not only isn't there realistic bulk 40 nm competition but TSMC also remain supply constrained. There could be some value in threatening to move at some nebulous date in the future, but most don't seem to think GloFo is going have much smoother node transitions than TSMC, and I believe TSMC still has more capacity at competing nodes.
In the Dell case, if Intel had been unable to prevent signficant update of AMD chips at Dell, it could have set them back quite a bit leading to them selling even less chips. So it was a case of sell a LOT more chips at a lower price. Or sit on a LOT more inventory that is continuously depreciating in value.
In this case, if Nvidia decides to lower their 40 nm wafer purchases (which they cannot) then TSMC would have no problems finding someone who would be more than happy to buy more wafers.
Regards,
SB
Yes, the situation is different. TSMC isn't giving Nvidia kickbacks to prevent them from going to another fab. That wasn't the point - the point is that the notion that outsiders "know" specifics of large contracts between firms is silly. This isn't some regulated utility we're talking about. In terms of leverage you seem to be looking at short-term considerations. The longer term implications of TSMC's relationships with Nvidia and AMD are a lot more interesting.
Bottom line is that for the last few years the common perception seems to be that AMD's smaller dies earn them higher margins yet this is never borne out in the financials. Therefore simplistic die-size analysis is irrelevant and out of touch with reality.
Um.. not really? People that have more to do (than us) with the semicon industry said that the deals between TSMC and IHVs area variable but mostly dependent on the number of dies ordered given the same process node which is quite reasonable.
I don't know what exactly do you mean by simplistic analysis but no, die-size analysis is far from being irrelevant. Of course, it may be less accurate than one may thing but that's it.
I don't know what exactly do you mean by simplistic analysis but no, die-size analysis is far from being irrelevant. Of course, it may be less accurate than one may thing but that's it.
Smaller dies cost less. That is not in doubt. So far, AMD has traded higher margins for higher unit share. Now that they own a majority market share, I expect real margins to rise going forward. Cypress prices are a good indication of that. IMO, this is following the natural path of market evolution.
No it's not in doubt at all and I don't have a problem with that assumption all other things equal. I do have a problem with blanket assumptions of similar net wafer costs to both firms though.
Which is supportive of the argument that their margins did not benefit from lower die sizes as ASP's were also lower. There is a constant mantra that AMD could slash prices at will as if they're raking in the cash, which they aren't. In terms of what happens next, well that's anyone's guess and has just as much to do with how well (or poorly) Nvidia executes as it does with the die sizes of AMD's next chip family.
Well..., I don't.
Silent_Buddha
Legend
This has also been the 3rd straight Quarter, I believe, that margins have risen for AMD overall as a company, despite their CPU division "probably" droping in margins. I just can't believe margins are all that great for CPU's like the Quad core Phenoms and salvage tri-core Phenoms selling for ~100 USD and less.
It also coincides quite well with, not being in a ruinous price war with Nvidia. Unlike Nvidia during the Rv770 vs GT200, they couldn't fall back on high margin professional sales (Quadro/Tesla) to help offsent falling margins in the consumer space. Their FireGL sales still lag quite far behind Quadro with no Tesla equivalent (I believe).
Despite that, they still managed competitive and sometimes higher margins than Nvidia during that same period.
Regards,
SB
They've got Cypress HPC/GPGPU cards coming in Q3.
Similar threads
