Scaling of todays CUDA programs to GT300


can anyone informed put some light on the question whether a program I write today with CUDA 2.3 for a 1.3 device will more or less scale in speed according to the specs of the upcoming G300? There are two reasons I ask. First of all from the g80/g92 to g200 it is quite straight forward to expect relatively good scaling since those are build of (more or less) the same basic blocks. But now I read that for the G300 we will not have anymore a SIMD but a MIMD processor (though one could argue that todays GPUs are already MIMD chips build from SIMD blocks). I also read that double precision is organized different to todays architecture.

Now, can I expect that a program, which I write for todays SIMD architecture (especially double precision code), scales relatively good to the G300? Meaning: if I reach now like 80% peak performance will it be close to 80% peak performance on the G300 as well [assuming I got enough threads running]?. In that scenario I would expect that the MIMD architecture just let me things do I can’t (or better shouldn’t) do now (i.e. no problems with branching) while code written with SIMD in mind would still run well.

I ask because I don’t want to start a project now which will be obsolete as soon as I finished it.

Best regards

Well, you ask the question which can be safely answered as “do not worry it will scale perfectly well and exceed the most optimistic expectation”, since there are no specifics there is no risk for an explicit lie. I’m very interested as well to learn about MIMD architecture of G300 and to compare its MIMD potentials with Larrabee. NVIDIA does not go public about this topic so I suspect that g300 probably is not an impressive MIMD performer; I love to be wrong about it. Once I dramatically underestimated Intel core2/i7 MIMD capabilities and as result of such wrong judgment has wasted for 1 year for gpu so this time I will wait for more substantial technical information to make any investment in gpu/mimd development; hopefully NVIDIA guys read many such requests…