Until now the GT440 was an OEM only card. It had 144 cuda cores (shaders) and was clocked at 594MHz. As it's CC2.1, only 2/3rds of the shaders are presently utilizable at GPUGrid. Basically this made the card operate as if it was a 96shader card, which makes it no better than a GT240.

Now NVidia have bestowed upon us a [96 Shader] retail version of the GT440.
Although it comes with a substantially higher clock; 810MHz, which in theory means the card should perform to within 10% of the 144shader OEM version, it is basically a GeForce GT 430! While I cannot confirm the GT440 is CC2.1, due to the many changes, going by the GF108 spec, I suspect it is, leaving you with 64 useful shaders.

- Don’t rush out to buy one!

Did you forget to say that the retail GT440 has 96 CC2.1 shaders?

MrS
____________
Scanning for our furry friends since Jan 2002

Yes, totally - and that was the main point!
[Edited], Thanks

At least 6months away, if it ever happens.

The retail GT440 is basically a GT430 with higher reference clock speeds and the option of DDR5. It will only perform as if it has 64 shaders, and would therefore not match up to a GT240. Although it has a higher clock than a GT240 my guess is that the 6.13 app would also slightly underperform compared to the GT240 with the 6.12 app, so it might only be about ½ as fast as a GT240; it would definitely not make it onto my recommended GPU list.
At 9Watts less (56W TDP) and approximately 10% faster (my guesstimate) the OEM version might be a better bet, but it would still not make it onto my recommended list.
At 106W TDP perhaps a GTS 450 (NON-OEM version) is an option (109W TDP)? It should at least slightly outperform a GT240.

More of a side note: I'm not convinced this "2/3 of shaders useable" is something a new GPU-Grid app could fix. It's actually up to nVidias hardware to detect independent instructions in the instruction flow for each set of "pixels". And up to nVidias driver to arrange these instructions in a way where it becomes easier for the hardware to detect said independent instructions. If it's detected, the additional shaders will be used, independent of the application.

The problem could be any of these two, or "just" a lack of independnet instructions in the algorithm, i.e. many atoms / pixels may be in flight at once, but for each of them any instruction uses the output of the previous one. Which would be normal for a time iteration, depending on how complex each iteration step is (don't know).

MrS
____________
Scanning for our furry friends since Jan 2002