ACML gpu benchmark numbers

[codedivine]

I hope people with different types of cards will post info here.
Running a 4870 512mb at 780mhz core ,1000mhz GDDR5 memory.
Anyone running 4850? 4870x2? 4850x2?
Anyone running a dual/triple 4870 or 4850?
Or 4670 (which cannot do double precision but you can try sgemm)?

(The above numbers on ubuntu 8.10 amd64)

edit : Major update : the CORRECT numbers on my GPU are:

170 gflops for DP and 380 gflops for SP.
Remember to run your CPU at full speed (i.e. disable powersave)!!

[rpg.314]

What size matrix? The bus transfer time is included I presume. What is the size needed to max out the speed. Is there an option to time the "compute on gpu" alone?

[codedivine]

The data transfer time is probably included since its timing the call to DGEMM and the use of GPU is transparent.

The timing applications are included as examples in ACML download. Heres the output. The first is matrix size, the second runtime in seconds and the third is mflops.

./time_dgemm.exe:

400 2.715632 47
600 0.029877 14459
800 0.037001 27674
1000 0.042463 47099
1200 0.056387 61291
1400 0.071061 77229
1600 0.089047 91996
1800 0.106845 109167
2000 0.134802 118692
2200 0.165079 129004
2400 0.207295 133375
2600 0.243919 144113
2800 0.289596 151604
3000 0.350922 153880
3200 0.416189 157466
3400 0.488798 160819
3600 0.574795 162339
3800 0.663308 165449
4000 0.760757 168253
4200 0.868719 170568
4400 0.984924 172975
4600 1.135350 171464
4800 1.269783 174190
5000 1.431244 174673
5200 1.865354 150757
5400 2.045647 153950
5600 2.262121 155266


./time_sgemm.exe

400 1.441326 88
600 0.024520 17618
800 0.027590 37114
1000 0.034138 58585
1200 0.039476 87547
1400 0.049521 110821
1600 0.057001 143717
1800 0.069265 168396
2000 0.092858 172305
2200 0.101135 210569
2400 0.117272 235758
2600 0.143149 245562
2800 0.160566 273432
3000 0.190883 282895
3200 0.217412 301436
3400 0.256443 306531
3600 0.314809 296408
3800 0.329136 333430
4000 0.384163 333191
4200 0.439259 337331
4400 0.498265 341922
4600 0.569931 341571
4800 0.616383 358841
5000 0.698239 358043
5200 0.801469 350875
5400 0.861945 365369

[rpg.314]

would ahve been nicer if you had plotted it. But no matter, for dgemm, the max occurs at about 5000x5000 size, which makes sense..