I have seen several Phenom 9500/9600 performance reports that show speed differences between the chip with the TLB workaround enabled and with the workaround disabled. They typically show a 5-15% performance drop with the workaround enabled (which is usually the default). My rendering tests showed 40% and 46% performance drop.
I have an AMD Phenom 9500, Gigabyte MA790FX-DS5 (shipped with BIOS revision level 3, with the TLB workaround toggle), and two Corsair XMS2 DDR2-800 2GB RAM sticks. I have a fresh Fedora 8 64-bit install, updated through pup using Livna and Fedora repositories.
The test uses the Radiance renderer that I use a lot for art purposes, and for which I have tested and tabulated nearly a hundred other machines. The single-thread test uses only 20MB of RAM, so I expect that the cache is especially active.
The single-thread test, with the TLB workaround enabled, clocked in at 7739s, or about as fast as an Athlon XP 2500+ or a 2.66GHz Xeon. Disabling the TLB workaround caused the rendering time to decrease to 4617s (40% decrease), or about as fast as an Opteron 250 (2.4 GHz).
The multithreaded test ran in 2142s with the workaround enabled, and 1154s with it disabled (46% decrease).
Does anybody else have real-world performance results for the step B2 Phenom TLB workaround?
Why would my tests indicate a much more significant performance slowdown than the other reported tests?
I have an AMD Phenom 9500, Gigabyte MA790FX-DS5 (shipped with BIOS revision level 3, with the TLB workaround toggle), and two Corsair XMS2 DDR2-800 2GB RAM sticks. I have a fresh Fedora 8 64-bit install, updated through pup using Livna and Fedora repositories.
The test uses the Radiance renderer that I use a lot for art purposes, and for which I have tested and tabulated nearly a hundred other machines. The single-thread test uses only 20MB of RAM, so I expect that the cache is especially active.
The single-thread test, with the TLB workaround enabled, clocked in at 7739s, or about as fast as an Athlon XP 2500+ or a 2.66GHz Xeon. Disabling the TLB workaround caused the rendering time to decrease to 4617s (40% decrease), or about as fast as an Opteron 250 (2.4 GHz).
The multithreaded test ran in 2142s with the workaround enabled, and 1154s with it disabled (46% decrease).
Does anybody else have real-world performance results for the step B2 Phenom TLB workaround?
Why would my tests indicate a much more significant performance slowdown than the other reported tests?
Comment