I'm working for a mechanical engineering firm that specializes in vibration modeling (mainly for NASA). We're trying to see if GPGPU computing is a good fit for our applications (models).
We're considering using either a Radeon 3870 or a FireStream 9170 (possibly a Radeon 4850, if it's practical in the near term?). For the purpose of proving the applicability of GPGPU computing for our model codes, would a dedicated Radeon 3870 suffice or is a FireStream 9170 required? With the FireStream 9250 on the cusp, I'd prefer to start with the Radeon 3870 (or possibly the 4850) and move over to the FireStream in September.
What's the scoop?
---jski
I just noticed the Radeon 4870 announcement today - are there Radeon Linux drivers for it that could be used for GPGPU purposes? Are am I pushing te raggedy edge?
--jski
Hi all,
Radeon 4850 seems to be not supported yet but it works for me - on 64-bit Ubuntu 8.04 system, latest fglrx, FireStream SDK Beta. As I've got it yesterday, I had no time to do anything useful yet and just played with some demos.
Sample results (from optimized_matmult_d):
Width Height Iterations Time Gflops
1024 1024 100 1.235000 161.943320
1024 1024 1000 9.271000 215.726459
2048 2048 1000 72.445000 220.857202
4096 4096 100 59.657000 214.559901
Note that as 48xx cards are still not (officially) supported, I'd not recommend this configuration to any production purposes but I'd certainly wait for it instead of settling on previous generation chips. Also note that this is a single-precision calculation demo.
My (early) opinion is that RV770 is a really good chip for number crunching (at least for single precision floating point numbers), the only thing I object about the whole card is its heat sink being permanently hot (it's a 1-slot design, so I'm not too surprised about it). But it should change as soon as vendors release non-reference design cards with bigger heatsinks.
Regards,
rle