Nouveau Fermi, GeForce GTX 580 Linux Benchmarks

[Veerappan]

Indeed, if you look at this comparison

http://openbenchmarking.org/result/1102278-CALI-OANVIDI95,1102272-CALI-OANVC0164,1101019-IV-RADEON22575&compare

It seems opensource fermi can already reach of up to over 70 % of the blob performance, while opensource radeon 5770 is stuck below 20% of blob performance in this particular benchmark. So this makes my curious (and a bit jealous) myself.

Isn't the r600g driver still using an unoptimized shader compiler? If so, it's possible that the r600g driver is using only the fat vector processing unit in each processing unit (out of 4-limited and 1 complete units which are available). That could explain the 20% upper limit issue.

Note that I haven't read the compiler code myself, and I'm going only on what I seem to remember from Phoronix reports from last year.