[TPU] NVIDIA GeForce GTX 650 Pictured - Page 10

I really don't care enough to damn the line up based on the name.

I hope to have 576 core 128bit Kepler GDDR5 6GHz @ selling $120-130. Something that actually run as fast as a factory OCed 7770.

It don't need, but it should have symmetric design, at least on the flag chip. 5 SMX/3 clusters is strange.

So I did some calculation, and found a possible way how can they do that. Basically they would do something like it has been done with PS3 Cell, it was printed with 8 CPUS, but released with 7. (6 usable for programmers)

They will print 6 SMX, and then test them. When all 6 would work, they would disable one, because 660 has 960 cores. (They can also keep it, and later decide to release it as GTX 660-1152.) When one of these 6 SMX will not work, they would erase it by LASER, and have correct specification to release it as GK106-400.
When manufacturing defect would affect two clusters each on a different memory controller, they would erase them and release it as GK106-250.
Now when manufacturing defect would affect memory controller/ROP cluster, or both affected clusters would be on the same controller, they would erase memory controller and both SMX connected to it, and they would release it as GK106-200.

But it seem strange: There are lot of GK106-400, which is correct, they did it to massively improve production rate of GK106-400. There are lot of GK106-200, majority of which can go to OEM. However there are just few of GK106-250. (Of course memory controller defects would be rare, thus the above table isn't exactly symmetric.)

The die size, 221 mm^2, actually supports the 6 SMX printed, 5 healthy SMX kept idea. They will pay for it by using about 1/8 - 1/6 more on wafer, but it would considerably improve yields.

If they went this way, then GK106-200 would have 768 CUDA cores, and it would basically behave as a GT 640, just it would be twice as fast... (GT 640 had great scores in Diablo 3, thus GK106-200 would be great Diablo 3 card. Of course it would be bandwidth starved as hell.) The numbering GK106-200 and GK106-250 also suggest they only changed number of ROP and memory controllers, the rest of the chip remained the same.
Edited by Raghar - 9/6/12 at 9:22am

I was joking originally. You taking everything so seriously creates a great troll opportunity I have chosen not to exploit.

You can't troll someone that doesn't really care

EVGA makes that many of one card because their motherboard division is dying they were the last to deliver a p67 board and its been on the decline since then.

kudos for 13 of the same card?