On Thu, Jan 23, 2003 at 04:16:24PM +0100, Gabriel Paubert wrote: > On Thu, Jan 23, 2003 at 01:48:29PM +0100, Geert Uytterhoeven wrote: > > On Thu, 23 Jan 2003, Gabriel Paubert wrote: > > > On Thu, Jan 23, 2003 at 12:12:56PM +0100, Geert Uytterhoeven wrote: > > > > > BTW, I get 45 MB/s for the buffer-cache on my 66 MHz SDRAM. > > > > > > Looks still very low, very conservative bridge timings? This means a > > > cache line read (32 bytes) every 700ns or so. Even if you dirty every > > > cache line, doubling the traffic, this means 350ns per burst transfer. > > > > Note these are the results of `hdparm -T', not from a real memory benchmark. > > Real memory read performance is ca. 100 MB/s, which is reasonable for > > machines > > of that age, cfr. http://home.tvd.be/cr26864/Linux/PPC/MemSpeed.html > > Ok, this means 90 or 135 MB/s on the bus, depending on whether > copy_to_user() uses dcbz or not. It should AFAIR, but I've not > looked at the code in a very long time. Depending on how the chips > is configured, dcbz may take quite some time too. In any case a far > cry from the theoretical half GB/s. > > Waiting for the PPC970 to see how much of the theoretical 6.4 GB/s > you can actually get in RL...
Any idea if the PPC970 bus is really hypertransport like so many people sy, or something else ? Friendly, Sven Luther
