Playing the numbers game
By
Charlie Demerjian: Thursday, 29 May 2008, 4:54 PM
THE LAST TIME Nvidia's CPU mouthed off about Intel, the firm followed it up with the stunning NV5800 'Dustbuster'. This time, he mouthed off, and the successor to the GT200 had already taped out. NV is in deep trouble once again.
You heard that right, the successor for the GT200 chip has already taped out, and it too will be nothing special. Documents seen by the INQ indicate that this one is called, wait for it, the GT200b, it is nothing more than a 55nm shrink of the GT200. Don't expect miracles, but do expect the name to change.
There are several problems with the GT200, most of which are near fatal. The first is the
die size, 576mm^2, bigger than most Itanics. One might trust Intel to make a chip that big with decent yields, especially if it is basically an island of logic in the middle of a sea of cache. Nvidia using a foundry process doesn't have a chance of pulling this off.
Word has come out of Satan Clara that the yields are laughable. No, make that abysmal, they are 40 per cent. To add insult to injury, that 40 per cent includes both the 280 and the 260 yield salvage parts. With about 100 die candidates per wafer, that means 40 good dice per wafer. Doing the maths, a TSMC 300mm 65nm wafer runs about $5000, so that means each good die costs $125 before packaging, testing and the like. If they can get away with sub-$150 costs per GPU, we will be impressed.
So, these parts cost about $150, and the boards will sell for $449 and $649 for the 260 and 280 respectively, so there is enough play room there to make money, right? Actually, most likely yes. There are costs though, but not enough to kill profit for any one touching these.
The biggest cost is memory. The 512b memory width means that they will have to use at least 16 chips. This ends up making the board very complex when you have to route all those high speed signals, and that means more layers, more cost, and more defect fallout with the added steps. You also have to solder on eight more memory chips which costs yet more.
To add insult to injury, the TDPs of the 260 and 280 are 182W and 236W respectively. This means big copper heatsinks, possibly heatpipes, and high-end fans. Those parts cost a lot of money to buy, assemble and ship. Not fatal, but not a good situation either. It also precludes a dual GPU board without losing a lot of speed.
Basically, these boards are going to cost a lot of money to make, not just to buy. The $449 price is justified by the cost. The last round of GX2 boards
cost OEMs about $425, meaning that NV charges OEMs about 70 per cent of retail for high-end parts. After packaging, shipping and add-ins, there is almost nothing left for the OEMs, quite possible explaining why one of their biggest one is teetering on the verge of bankruptcy, kept alive because NV won't call their debt while still shiping to them. Watch for this to melt down once NV loses the high end.
So, you end up with an expensive chip on an expensive board that makes few if any people money. Fair enough, bleeding-edge parts mean bleeding-edge prices. The problem is that ATI is going to make a chip that competes with GT200, and lines up with it point for point. NV wins big Z Fill, ATI crushes them on Shader Flops. What this translates to in the real world is still up in the air, but it looks like the 770 and the 260 will be about equal for most things.
The GT200 is about six months late, blew out their die size estimates and missed clock targets by a lot. ATI didn't. This means that buying a GT260 board will cost about 50 per cent more than an R770 for equivalent performance. The GT280 will be about 25 per cent faster but cost more than twice as much. A month or so after the 770 comes the 700, basically two 770s on a slab. This will crush the GT280 in just about every conceivable benchmark and likely cost less. Why?
Because.
So, what is a company to do when it promised the financial world that ATI was lost, and GT200 would raise their margins by 100 basis points or so? Surely they knew what was coming a few weeks ago during their financial call, right? I mean, if word was leaking days later, the hierarchy surely was aware at the time, right?
The answer to that is to tape out the GT200b yesterday. It has taped out, and it is a little more than 400mm^2 on a TSMC 55nm process. Given that TSMC tends to price things so that on an equivalent area basis, the new process is marginally cheaper than the old, don't look for much cost saving there. Any decrease in defectivity due to smaller area is almost assuredly going to be balanced out by the learning curve on the new process. Being overly generous, it is still hard to see how the GT200b will cost less than $100 per chip. Don't look for much cost savings there.
The new shrink will be a much better chip though, mainly because they might fix the crippling clock rate problems of the older part. This is most likely not a speed path problem but a heat/power issue. If you get a better perf/watt number through better process tech, you can either keep performance the same and lower net power use, or keep power use the same and raise performance.
Given NV's woeful 933GFLOPS number, you can guess which way they are going to go. This means no saving on heatsinks, no savings on components, and a slightly cheaper die. For consumers, it will likely mean a $50 cheaper board, but no final prices have come my way yet. It will also mean a cheaper and faster board in a few months.
The GT200b will be out in late summer or early fall, instantly obsoleting the GT200. Anyone buying the 65nm version will end up with a lemon, a slow, hot and expensive lemon. Kind of like the 5800. It would suck for NV if word of this got out. Ooops, sorry.
What are they going to do? Emails seen by the INQ indicate they are going to play the usual PR games to take advantage of sites that don't bother checking up on the 'facts' fed to them. They plan to release the GT200 in absurdly limited quantities, and only four AIBs are going to initially get parts.
There is also serious talk of announcing a price drop to put them head to head with the R770 and giving that number to reviewers. When the boards come out, the reviews are already posted with the lower numbers, and no reviewer ever updates their pricing or points out that the price performance ratio was just blown out of the water. There is also internal debate about giving a few etailers a real price cut for a short time to 'back up' the 'MSRP'.
We would hope the reviewers are able to look at the numbers they were given on editors' day, $449 and $649, and take any $100+ last minute price drop with the appropriate chunk of NaCl. Just given the component cost, there is no way NV can hit the same price point as the 770 boards. "We lose money on each one, but we make it up in volume" is not a good long term strategy, nor is it a way to improve margins by 100 basis points.
In the end, NV is facing a
tough summer in the GPU business. They are effectively out of the Montevina market, and they are going to lose the high end in a spectacular way. Nvidia has no effective countermeasure to the R770, the GT200 was quite simply botched, and now they are going to pay for it. When all you have is a hammer,
everything looks like a 5800. µ
To be honest though, Nvidia have all the resources in the world available to them and they have still "possibly" fluffed it. However, as this was the Inq we should really wait and see.
