GTX295 shortage, where can I find them (except overpriced on ebay) Need to build CUDA rig, but can&#

So, I need some cards, even dual pcb would do, help, I can’t wait until Fermi comes…

I’m surprised there hasn’t been more discussion about this but the availability of the entire GTX 200 series is becoming a serious problem considering that Fermi isn’t due out for another month (and I’m hearing rumours that it might not be available in volume even then). I know some people here are using Tesla and others are not buying large quantities of cards but there must be people affected by this, surely?

And if the pricing I’m seeing is accurate ($480 for a GTX 470 1GB or $680 for a GTX 480 2GB) then that is also a serious problem since you can buy two GTX 260 1.8GB for less than the cost of a single GTX 470 1GB.

Those rumored prices seemed shockingly high to me, but then I went back and looked at the previous major releases. The 8800 GTX debuted at $600, and the GTX 280 started at $650 (!!?).

But yeah, they definitely need to keep the retail channel stocked with GT200 cards until the Fermi cards come down 30% in price.

Fermi is late. It is understandable.

If you can find a GF200 card it is selling at a 10-30% premium. The last GF200 chips were probably manufactured in November. Knowing that Fermi was going to be late, they should have ordered more. You can buy 5000, 6000, 7000, 8000, and 9000 series cards. Will GF200 cards ever be available again?

$50 - $100 cards ( GT210, 220 and 240 ), older cards up to $200 and then a large gap to $600 GF400 cards does not seem like a winning stategy.

The next scandal waiting to happen is card manufactures so desperate for GF200 chips they start repairing broken cards and try to sell them as new.

The ebay cards aren’t always overpriced. You just have to keep watching.
Here’s one from today, brand new, for $370. That’s a bargain! (Admittedly the best price I’ve seen in weeks… this is one to buy now if you’re in the market.)

As far as I’m concerned its a disaster. It makes even the GTX 470 too expensive for us. If we can’t get the GTX 260 anymore then we’re probably going to have to go back and see what we can do with the GTS 250. I actually seem to remember some OEM launching a dual GPU card (similar to the 9800 GX2) and thinking they were mad but it actually almost makes sense now. I think if I were nVidia I’d be rushing like hell to make a 40nm GTX 260.

The 2GB 260 and 285 cards are nowhere to be found (ok, found one 2GB 285 card priced at 700 bucks!).

In terms of Gflops the 260 216 givse most bang for the buck in terms of Gflops/dollar, but including the external hardware costs (components to build the system), the GTX275 is better value. The GTX295 is the clear winner. So for scaling things up it doesn’t make sense to go for anything else than the 295, anyone know how many Gflops the GTX480 can do?

OK, I’m seeing plenty of GTX 260s and 285s on Newegg still. (Only one GTX 275 model in stock, though.) The only models to have sold out are the high memory 1792 MB and 2 GB versions.

The GTX 295s are out of stock, suggesting that those chips are being diverted to the better selling cheaper models. I agree, though, if the GTX 260s start to sell out before June, then you know there is a serious problem.

It’s going to be something like 512 multiply-adds per clock cycle, so if the rumors of 1.25 GHz are true (down from a projected 1.5 GHz), that comes out to be a peak of 1280 GFLOPS.

Which makes it horrible value when it comes to CUDA calculations, right?

I don’t think it is horrible value. In terms of raw floating point and memory bandwidth it is going to be close to the two GPUs in a GTX 295 combined. (Depending on whether you count that extra dual issue MUL in the GT200 as a significant boost or not. It’s always been hard to tell how much that really helps you.) The GTX 480 certainly will not have as much FLOPS-for-the-buck as a GTX 295 (assuming they were still in stock for purchase at $550 each), although the value of a GTX 295 depends on your problem being partitioned efficiently over multiple GPUs.

Basically, it sounds like Fermi will not be a huge win for people who have compute-bound single precision floating point applications which can be trivially split over multiple GPUs. That is not terribly surprising, as those kinds of problems almost always benefit by using a lot of not-quite bleeding edge hardware. (Hence the effectiveness of clusters of PCs.)

For comparison, the 9800 GX2 was comparable to the GTX 280 in raw single precision floating point (again, depending on whether you count the dual-issue MUL) when it came out. The real win in the GTX 280 was the performance of two previous generation GPUs in one, plus all of the increased flexibility of new features, like double-precision, better memory coalescing, and more registers. If your code already ran great on 9800 GX2s when the GTX 280 came out, then it would have been silly to upgrade. The new GPU would offer no improvement.

However, if your code was limited by memory coalescing or double precision, or not enough registers to get good occupancy, then GTX 280 was a win. Similarly, with Fermi, if your code is limited by a lack of caching, or slow double precision, or not enough shared memory, or poor C++ support or slow atomics, then the GTX 480 will be a win and have great value.

So yeah, depending on your problem, the GTX 480 will range from “Who cares?” to “Freaking Awesome!”. There is no one-size fits all answer here.