With the architectural previews hitting the wire this week, we took the time to actually look at potential cost of the GF100-based solutions and where the Fermi silicon stands right now. Our respected colleague Wolfgang Hegel provided us with the mockup of the TSMC’s 300mm wafer containing GF100 dies. As you can see in the image below, according to our estimates, nVidia can fit 94 GF100 chips on a single 300mm wafer as TSMC is able to make them.

Mockup shot of GF100 on 300mm wafer. Picutre Credit: Wolfgang Hegel
Mockup shot of GF100 on 300mm wafer. Picutre Credit: Wolfgang Hegel

Our original data cited that nVidia is able to put 130 dies on a single wafer but after learning the size of the chip [our sources estimate the die size at 24x23mm or 24x24mm], it was obvious that the chip was designed with the limitations of the original GT200 in mind, which was manufactured using TSMC’s 65nm node. If our sources are correct, GF100 comes with the die size almost identical to the GT200, at 570-571mm2. Long story short, one wafer should fit no more than 94 of these bad boys.

TSMC sells chips in two ways: you can purchase the whole wafer, in which the cost per die can get as low as possible – but you’re responsible for the yields in question. Second mode is purchasing dies alone, i.e. you pay much more per single die but ultimately – you don’t care what the yield is. Most of TSMC customers chose either one or another but with clients as big as AMD and nVidia are – those pricing models change on product range. However, in terms of die size and overall number wafers, nobody comes close to these two. Our sources claim that in case of GPU dies such as Cypress and GF100, both AMD and nVidia are ordering per wafer and they’re paying around $5000 per single wafer manufactured on 40nm Performance node.

This information gave us invaluable insight into the reasons why GF100 ended up delayed by a quarter, getting pushed from November 2009 to March 2010. First of all, TSMC screwed the pooch badly and the foundry wasn’t able to offer higher yields than 40% on their 40nm process node for better part of 2009. While Foundry execs claim that the yields of 40nm process reached the level of yields 65nm process had, the fact of the matter is that lies true on ASIC per ASIC basis. There are chips that yield excellently, and there are chips that yield horribly. Unfortunately, a complex design from AMD is currently having better yields than a less-complex part, thus we’re not surprised that AMD is considering 2010 as a transitional year before the beginning of the switch to GlobalFoundries.

In the case of nVidia GF100, the best they can expect at 40nm is 60% yield. Realistic goal is 40% and nVidia is going to end up with 24x24mm die costing $131 [40% yield] and in fairy tale land, nVidia is going to pay only $87 per chip. Now, if nVidia would have to pay $131 for a single GF100 – that would not be that bad, given the current cost of single AMD Cypress die is $96. For as long as you’re under $100 per die, you can count your blessings with TSMC’s 40nm process. Reality is that AMD can easily reach 60-70% yields with Cypress and they will drive the cost of a single Cypress ASIC to  anywhere between $54-64 [70 to 60% yield]. Thus, cost advantage AMD, no matter what happens. 

This is where the issue happened with this three billion transistor monster of a die. The yield of initial A1 silicon resulted in only nine working dies per wafer [which was the source of false rumor of 2% yield - in reality, 2% yield would be 1-2 dies per wafer], i.e. less than 10%. If we take a figure of $5000 per single wafer, the yield of around 10 percent put a price of single NV100 chip to around $550. With only Tesla businesses able to support that chip price [Quadro could sustain up to $350 per die and remain profitable], it was obvious that A2 revision was the way to go. Unfortunately, A2 revision didn’t come up with necessary yield increase. According to our sources, nVidia achieved a 25% yield with A2 silicon, coming up with around 24-26 functional dies per wafer. This puts the cost of a single chip into the $208 range, i.e. you can get 4.2 billion transistors [2x Cypress, i.e. HD 5970] for the price of a single three billion die and buy the cooling and the PCB. In case of nVidia, you still need to build the whole enchilada around the $208 piece of silicon.

Truth to be told, given that GF100 will probably beat both HD 5870 and HD 5970 in Tessellation and consumes less power than a HD 5970 – this isn’t as bad as we thought the first time we did the numbers game. Then again, AMD did ship over two million DX11 cards to date with nVidia shipping none.

What is important for the price of the die is the end price for the consumer. While nVidia will find tests in which it beats both the single and dual-GPU HD 5870 and the HD 5970, going with the $599 price tag would be suicide in this economy. We expect to see a price of $549 or $499 for the high end part, probably named "GeForce 380" and $349-399 for the "GeForce 360". The problem is that a $200+ die is killing any margin on the part and nVidia cannot price themselves out of the market. Thus, we expect to see Quadro and Tesla business significantly subsidizing for the price of desktop GeForce boards, at least until nVidia gets its yields into the 40% range. The good part about nVidia and 40nm is that again, according to the sources in the know – the company currently also has one of highest yielding 40nm parts on the market, the 2nd generation Tegra. Tegra uses 40nm low-power process though, so it is not directly comparable to Cypress or GF100.

nVidia Tegra 2nd Gen: Allegedly very good yields at 40nm process
nVidia Tegra 2nd Gen: Allegedly very good yields at 40nm process – then again, this is not a small die either.

AMD has a significant advantage because of almost a year of experience working with 40nm process and they know all the quirks, while nVidia’s decision to be really conservative with the manufacturing process is now coming to haunt them.

nVidia ramped up the GF100 manufacturing and is now getting ready for the launch. We know that mass production of GF100 boards will initialize in first week of February, with other AIB partners following in second, third and fourth week of February. However, we don’t expect that GeForce lineup will be nVidia’s bread and butter. Not this time around.

Update #1 January 21, 2010 23:10 GMT - In order to clarify the article, we’ve added information about Cypress die estimates and running costs. We also cleared the confusion in unit sizes, since we ran a wrong one [2.4cm equals 24, not 240mm]. We apologize for confusion that ensued.