460-470-480 market share, which architecture to target?

Version 2.1 is a bit different than 2.0, is some cases it maybe worse. Will all new gpu be based on this arhitecture? Will 470 and 480 be in production or it iwll be moved to 460? I am writing consumer cuda app so need a gpu for tweaking. I assume 460 should be much more popular based on its price and market position. However, a lot of 470-480 were sold already.

Well, 460 comes with sm_21 rather than sm_20 for 470/480 so I would go with the highest version number available as I can’t really see nvidia dropping support for features that were in sm_21 in sm2_X…

Well, 460 comes with sm_21 rather than sm_20 for 470/480 so I would go with the highest version number available as I can’t really see nvidia dropping support for features that were in sm_21 in sm2_X…

The best place to get market share stats is from the Steam Hardware stats: http://store.steampowered.com/hwsurvey/videocard/

This is updated monthly. Right now, there are more GTX 470, then GTX 480, then GTX 460, and finally GTX 465 in that order (I’m only looking at the 4xx series). But as you can see at that site, the momentum of the GTX 460 is much higher than the 470 or 480, so I would suspect that it would certainly surpass the 480 in the near future, and probably surpass the gtx 470 in the not-to-distant future. I’d probably target the 460. But remember, whatever code you write for any of these cards should work fine for all the others in the 4xx series. Unless you’re using heavy double floating point, in which case, you’d need to target the Tesla series instead of the Geforce series.

The best place to get market share stats is from the Steam Hardware stats: http://store.steampowered.com/hwsurvey/videocard/

This is updated monthly. Right now, there are more GTX 470, then GTX 480, then GTX 460, and finally GTX 465 in that order (I’m only looking at the 4xx series). But as you can see at that site, the momentum of the GTX 460 is much higher than the 470 or 480, so I would suspect that it would certainly surpass the 480 in the near future, and probably surpass the gtx 470 in the not-to-distant future. I’d probably target the 460. But remember, whatever code you write for any of these cards should work fine for all the others in the 4xx series. Unless you’re using heavy double floating point, in which case, you’d need to target the Tesla series instead of the Geforce series.

Thanks.
I wonder if NV will make high end 2.1 cards, or is it just branch for volume niche? Interesting what will be plans before totally new architecture.

Thanks.
I wonder if NV will make high end 2.1 cards, or is it just branch for volume niche? Interesting what will be plans before totally new architecture.

The speculation is that there will be a dual GPU version, perhaps called GTX495. This isn’t any announcement or clue from NV, it’s just logical since two GF104 chips on one PCB would draw just under 300W, so it’s manufacturable. (2 GF100 chips won’t fit that wattage even if they’re clocked slower than the GTX465 is).

Performance in games would be about 20% better than the GTX480, and likely similar for CUDA.

NV costs would likely be cheaper than the GTX480, just looking at the fact that two GTX460 boards are cheaper than one GTX480.

So technically and financially it makes sense. And it uses only parts which NV has already been shipping, so it’s not waiting for anything.

I was hoping to see some announcement at GTC last week, but I guess we’re all still just waiting. Jen-Hsun did say last week that there would be a “mid life kicker” to Fermi, but that’s the extent of our knowledge.

The speculation is that there will be a dual GPU version, perhaps called GTX495. This isn’t any announcement or clue from NV, it’s just logical since two GF104 chips on one PCB would draw just under 300W, so it’s manufacturable. (2 GF100 chips won’t fit that wattage even if they’re clocked slower than the GTX465 is).

Performance in games would be about 20% better than the GTX480, and likely similar for CUDA.

NV costs would likely be cheaper than the GTX480, just looking at the fact that two GTX460 boards are cheaper than one GTX480.

So technically and financially it makes sense. And it uses only parts which NV has already been shipping, so it’s not waiting for anything.

I was hoping to see some announcement at GTC last week, but I guess we’re all still just waiting. Jen-Hsun did say last week that there would be a “mid life kicker” to Fermi, but that’s the extent of our knowledge.

… and will future Tesla products incorporate compute 2.1 features?

… and will future Tesla products incorporate compute 2.1 features?

No. Compute 2.0 is superior to 2.1 for Tesla… much higher DP throughput and ECC memory, mostly.

No. Compute 2.0 is superior to 2.1 for Tesla… much higher DP throughput and ECC memory, mostly.

I’m talking about the redesign of the SM’s.

I assume it was also important in increasing the performance per watt otherwise why did they do it.

Such a change is equally applicable to the design of the next Tesla chip given these must meet a certain heat/power limit.

I’m talking about the redesign of the SM’s.

I assume it was also important in increasing the performance per watt otherwise why did they do it.

Such a change is equally applicable to the design of the next Tesla chip given these must meet a certain heat/power limit.

Do you guys have any news on how hard it is to fully utilize the GTX460? I heard talk of increased register bandwidth pressure that made it more difficult to reach peak flops on this arch than < 2.1 archs. Apparently one needed to do some arch specific optimizations to get more ILP ? Will there be new drivers for this?

This makes me skeptical about buying 460s for CUDA. Does anyone with hands on experience have some advice?

Do you guys have any news on how hard it is to fully utilize the GTX460? I heard talk of increased register bandwidth pressure that made it more difficult to reach peak flops on this arch than < 2.1 archs. Apparently one needed to do some arch specific optimizations to get more ILP ? Will there be new drivers for this?

This makes me skeptical about buying 460s for CUDA. Does anyone with hands on experience have some advice?