WHEN WE SAID that Nvidia’s Fermi architecture was wrong, most people didn’t understand what we meant, they focused on the chips. With the release of the GTX460, one look at the die size puts the problem clearly into focus, and it is quite ugly for Nvidia.
If you recall, we were the first ones to say that the GF100 was huge, 529mm^2, months before anyone else had hard data. This was backed up 8 months later by some pictures on EXPreview. We can now tell you about GF104 die, it is 24.9mm x 14.7mm, and the die size is 367mm^2. (Note: The dimensions are rounded to one significant digit after the decimal point, they are a hair larger so multiplying those rounded numbers gets you 366mm^2. This is not an error, the 367 number is correct, some mad scientists directly measured the die.) For some odd reason, Nvidia didn’t give out any die size numbers when talking to the press, almost like they didn’t want people to know something before the reviews went up.
The closest site to publish an accurate number was the TechReport with 320mm^2. For that, Scott Wasson wins the mostly-not-crushed-much individually wrapped Kit-Kat stick that wound up in my travel bag after Computex. Congratulations Mr. Wasson, see me at IDF for your prize. No one said that the GF104 was larger than it’s closest rival, AMD’s Cypress/HD58xx, and that is what Nvidia didn’t want you to know.
The problem is simple for Nvidia, the economics of this part don’t work out, the underlying architecture is wrong, so the resultant parts start out with an uphill battle. This is a problem for Nvidia, not for the end user. If the GTX460 is priced at a loss, the consumer shouldn’t care, they get a deal, and that is the end of it. Retail buyers rarely care if the part is making a profit for the manufacturer.
Cypress/HD58xx is 334mm^2, which makes GF104 about 10% larger. If yields are the same for both parts, that would mean GF104 costs 10% or so more than Cypress to make. Sources in the fab community tell SemiAccurate that the GF104 is nowhere near the yield of Cypress.
As a humorous aside, both chips are made on the same process, TSMC’s 40nm, and literally at the same fab. AMD managed to cram 2.15 billion transistors into 334mm^2, about 6.44 million transistors per mm^2. GF104 has 1.95 billion transistors in 367mm^2, about 5.31 million transistors per mm^2. This means AMD’s Evergreen architecture is over 20% more space efficient than GF104 while delivering much more raw performance and vastly more performance per watt. When SemiAccurate teases Nvidia’s layout and physical design teams, it is for a reason.
Getting back to the story, to use ballpark figures, lets assume Cypress is now yielding at 75% for HD5870 and HD5850 combined, much higher if you add in HD5830. GF104 has one of it’s seven shader groups fused off leaving 336 functional shaders, and upping yields. Lets just ballpark that chip’s yield at 60% for reasons we can’t get into. If Nvidia decides to release more GF104 derivatives, it will up the effective yield, but anything lower on the performance scale would go up against Juniper, a 166mm^2 chip. That is economic suicide.
This means you can get about 150 GF104 die candidates per 300mm wafer, at 60% yields, that is about 90 good parts. Cypress has about 160 candidates, and about 75% yields, meaning 120 good parts per wafer. You can play with the number as much as you like, see what David Morgan did here for example, but the fact that Cypress is smaller and has notably higher yields is the take home message.
If a TSMC 40nm wafer costs $5000, that means each good 5850 or 5870 costs AMD about $42 to make. Each good GF104 costs Nvidia around $56. If you add in 5830s the number goes down for ATI, and the same holds true for any GF104 derivative that Nvidia releases. This is a problem for reasons we will get into later, but not the end of the world for Nvidia.
The Cypress based HD5870 and HD5850 cards seem to be selling for about $380 and $300 at etail respectively while the GF104 based GTX460 sells for $199 and $229 with 768MB and 1GB of memory. If you assume that the card models are split 50-50 for supply, that means AMD cards sell for an average of $340 while Nvidia cards sell for $215 on average. The two Nvidia cards straddle the lowest end Cypress variant, the HD5830, in the market, and are well below the price and performance of the HD5850.
Getting back to the GPU prices, you might recall that Nvidia’s silicon component costs about 25% more than AMD’s, and the resultant card sells for about 63% of AMD’s card sells for. Can you say squeezed margins, if they exist at all? ATI can lop $100 off their retail price and still make more money on Cypress than Nvidia currently makes on GF104.
Nvidia’s will make some money on GF104 only as long as AMD feels like letting them. Because TSMC can’t make enough 40nm wafers to go around, Cypress is currently supply constrained. Should GF104/GTX460 eat into AMD’s marketshare enough to effectively allow Cypress demand to be met by current wafer availability, AMD will drop prices and likely push Nvidia into the red on GF104. It isn’t a question of “Can AMD?”, it is a question of “Will they?”.
If you add in the low end HD5830 card, the numbers get better for AMD. ASPs go down, but so does cost because effective yield goes up. If Nvidia puts out a lower spec GF104 with 6 of the 8 clusters active, they will push yields up, but also push ASPs down a lot. That low end card would be competing with AMD’s Juniper, a 166mm^2 part that costs far less than half of what Cypress costs to make. Once again, Nvidia is bringing a knife to a gunfight because their architecture is fundamentally broken.
If Nvidia puts out a higher spec GF104 with all 8 clusters active, they have the ability to raise ASPs, but yields will take a huge pounding, as will supply. This may be good for polishing the halo for the gullible onlookers, but it is doubtful it would help their financial situation much.
In the end, the GF104/GTX460 is a good card for consumers, it is priced right, and slots into a gaping hole in the AMD lineup. That hole exists because it is better for AMD’s bottom line for it to be there. Nvidia poked a sleeping lion with the GTX460, and if it decides to roll over and swat the green team they have no defense.
To make matters worse, the Fermi architecture is not just wrong, it was also quarters late. About the time that Nvidia fleshes out the GF10x line with the upcoming GF106 and GF108 variants, AMD will have their next generation called Southern Islands out.
That will let AMD waterfall the prices on the current parts, and leave Nvidia without any place to turn, financially speaking. Best case, Nvidia has a window of a few months before AMD brings down the hammer. In the mean time, the boys in green have to be very careful, poking the sleeping lion might bring that hammer down faster. No matter which way they turn, Nvidia will get squished.S|A
Update: Clarified significant digits.