Daniel said:
Hi Stephan!
I see you follow the things as they are and not as they look.
Well, it's kinda hard to deny this issue if one gets benchmark results
like these:
http://stephan.win31.de/hdt722525dlat80-v44o.ps
http://stephan.win31.de/hdt722525dlat80-v44o.txt
Note the capped writes, while reads are fine. Since I had seen 55 megs a
second for graphics transfers in the past, I only had to put one and one
together.
BTW, I do like to do some more in-depth dabbling with old hardware,
partly because I'm too broke for the new interesting stuff
. (These
benchmark results for Intel's new "Sossaman" core CPU look nice, the
power consumption of the beast even more so. It would be great if these
came in socket 604, then one could use a nice würgstation, err,
workstation board with PCI-X and all the goodies without being killed by
(a) a look at the power bill or (b) $PARENTS after (a).) You may want to
read up on my current GA-686KDX adventures, including my first
(successful) attempt at BIOS modding. (Admittedly that BIOS mod wasn't
too hard to do, the right tools are easily to be found for a BIOS of
this vintage and I didn't do anything too outlandish. Sure was fun
though - and the result even works.)
Yeah, something about 50-60MB/sec you can have from the PCI Bus. ;-)
More.
I remember my Saturn II, SP3G with 512K second level kätsch.
^^^^^^
That's what I call an interesting interpretation of "cache"...
Sure
beats my "dadaschiet".
The PCI
Bus on this beast was unbeatable. It is a speed standard, till now
.
58MB/sec.
The most you can wring out of a 32/33 PCI bus is about 110...115 MB/s
for busmaster transfers (--> tecChannel). I have personally seen more
than 90 on a BX board (Adaptec SCSI Bench reading from an U2W and U66
attached drive at the same time) so I guess it wouldn't fall short here
(at least reading, i.e. PCI -> CPU). How much you can get on a 440
series chipset in CPU -> PCI direction will depend on the exact chipset
and on whether USWC (uncached, speculative write combining, or write
combining in short) is turned on for the graphics memory ranges. With
USWC, even a humble i440FX will manage to transfer more than 90 MB/s to
a Millennium II in LFB VESA modes, provided one has used FastVid (1.10
is the last version, it seems) or ctppro. The latter also is good for
SCSI host adapter tweaking, btw - set the address range with the SCSI
BIOS to WT instead of UC, and the SCSI stuff will operate faster in DOS,
as evidenced by the System Speed Test v4.78 disk benchmark (on the
GA-686KDX with a 53C895 based HA using the last 4.19.00 BIOS, maximum
STR grew from 39 to >41 MB/s - the maximum for the oldish Fujitsu MAJ I
used - , burst transfer from 33xxx to 51xxx MB/s). Without USWC, write
data rates seem to range from 30..35 MB/s (i440FX) to 55 MB/s (i440BX).
On a BX (or LX) board, using USWC allows quite impressive transfer rates
to AGP graphics cards, up to 252 "MB/s" (which probably is more like
MiB/s and would thus translate to ~264 MiB/s or almost 100% of the
bandwidth available in PCI mode), even more with the AGP overclocked
(BX@133). I have tried to enable AGP with 2X transfers by writing to the
chipset and graphics card registers but alas, no change whatsoever.
Maybe some more configuration of the graphics card prior to enabling AGP
mode would be necessary, and I haven't studied the AGP spec intensively
enough to know what this would have to look like.
One could try to circumvent the 55 MB/s write limit for disk controllers
by setting up their memory ranges for USWC, but I have no idea how one
would modify CPU MTRRs in WinNT/2k/XP/S2003 (nor what the cards may
think of this).
Stephan