from l'inq
Nvidia GTX260 and 280 revealed
Missed targets and low yield
By Charlie Demerjian: Saturday, 24 May 2008, 5:57 PM
NVIDIA STAGED ITS regular editors' day, and once again, we were not invited. Luckily, that means we can tell you about it early.
With the usual flair of spin and statistics which almost no one questions for fear of being cut off, NV talked about two new GPUs, the GTX280 and GTX260. The 280 is the big one, the 260 is the mid range, what used to be the GTS.
The 280 has 240 stream processors and runs at a clock of 602MHz, a massive miss on what the firm intended. The processor clock runs at 1296MHz and the memory is at 1107MHz. The high-end part has 1G of GDDR3 at 512b width. This means that they are pretty much stuck offering 1G cards, not a great design choice here.
The 280 has 32ROPs and feeds them with a six and eight-pin PCIe connector. Remember NV mocking ATI over the eight-pin when the 2900 launched, and how they said they would never use it? The phrase 'hypocritical worms' come to mind, especially since it was on their roadmap at the time. This beast takes 236W max, so all those of you who bought mongo PSUs may have to reinvest if they ever get three or four-way SLI functional.
The cards are 10.5-inch parts, and each one will put out 933GFLOPS. Looks like they missed the magic teraflop number by a good margin. Remember we said they missed the clock frequencies by a lot? Here is where it must sting a bit more than usual, sorry NV, no cigar.
The smaller brother, aka low-yield, salvage part, the GTX260 is basically the same chips with 192 SPs and 896M GDDR3. If you are maths-impaired, let me point out that this equates to 24 ROPs.
The clocks are 576MHz GPU, 999MHz memory and 896MHz GDDR3 on a 448b memory interface. The power is fed by two six-pin connectors. Power consumption for this 10.5-inch board is 182W.
This may look good on paper, but the die is over 550mm, 576 according to Theo, on the usual TSMC 65nm process. If you recall, last quarter NV blamed its tanking margins on the G92 yields.
How do you fix a low yield problem? Well, in Nvidia-land, you simply add massive die area to a part so the yields go farther down. 576 / 325 = 1.77x. Hands up anyone who thinks this will help them meet the margin goals they promised? Remember, markets are closed Monday, so if you sleep in, no loss.
The 260 will be priced at $449 and go up against the ATI 770/4870 costing MUCH less. The 280 will be about 25 per cent faster and quite likely lose badly to the R700, very badly, but cost more, $600+.
Nvidia GTX260 and 280 revealed
Missed targets and low yield
By Charlie Demerjian: Saturday, 24 May 2008, 5:57 PM
NVIDIA STAGED ITS regular editors' day, and once again, we were not invited. Luckily, that means we can tell you about it early.
With the usual flair of spin and statistics which almost no one questions for fear of being cut off, NV talked about two new GPUs, the GTX280 and GTX260. The 280 is the big one, the 260 is the mid range, what used to be the GTS.
The 280 has 240 stream processors and runs at a clock of 602MHz, a massive miss on what the firm intended. The processor clock runs at 1296MHz and the memory is at 1107MHz. The high-end part has 1G of GDDR3 at 512b width. This means that they are pretty much stuck offering 1G cards, not a great design choice here.
The 280 has 32ROPs and feeds them with a six and eight-pin PCIe connector. Remember NV mocking ATI over the eight-pin when the 2900 launched, and how they said they would never use it? The phrase 'hypocritical worms' come to mind, especially since it was on their roadmap at the time. This beast takes 236W max, so all those of you who bought mongo PSUs may have to reinvest if they ever get three or four-way SLI functional.
The cards are 10.5-inch parts, and each one will put out 933GFLOPS. Looks like they missed the magic teraflop number by a good margin. Remember we said they missed the clock frequencies by a lot? Here is where it must sting a bit more than usual, sorry NV, no cigar.
The smaller brother, aka low-yield, salvage part, the GTX260 is basically the same chips with 192 SPs and 896M GDDR3. If you are maths-impaired, let me point out that this equates to 24 ROPs.
The clocks are 576MHz GPU, 999MHz memory and 896MHz GDDR3 on a 448b memory interface. The power is fed by two six-pin connectors. Power consumption for this 10.5-inch board is 182W.
This may look good on paper, but the die is over 550mm, 576 according to Theo, on the usual TSMC 65nm process. If you recall, last quarter NV blamed its tanking margins on the G92 yields.
How do you fix a low yield problem? Well, in Nvidia-land, you simply add massive die area to a part so the yields go farther down. 576 / 325 = 1.77x. Hands up anyone who thinks this will help them meet the margin goals they promised? Remember, markets are closed Monday, so if you sleep in, no loss.
The 260 will be priced at $449 and go up against the ATI 770/4870 costing MUCH less. The 280 will be about 25 per cent faster and quite likely lose badly to the R700, very badly, but cost more, $600+.
