> The System Bus speed is MHz as set by the CPU card. The RAM
> throughput numbers are the closest way I know to measure how the
> hardware and OS can take advantage of the RAM/Bus/CPU throughput.
The system bus *speed* is set by the CPU card, up to the limit of the
motherboard hardware. The *Bandwidth* in MB/s is (S * W) / 8.
S = Bus speed (MHz)
W = Bus width (bits)
This is reduced by bottlenecks in components (like the RAM, or the PCI card,
or the CPU, or the PCI bridge chipset). For example, for RAM it's really:
S W
----- * ---
1 + C 8
C = Wait states (clocks).
If the measured bandwidth is significantly lower than (S * W / 8) then
there's an additional bottleneck. For example, with the original 68000
the CPU could only read from the bus every other cycle, so the original
Macintosh's bandwidth and bus speed were the same: the 16-bit bus was
balanced by the 2-cycle access. This hasn't been true for any Mac with
a 68020 or better... the 68020 can saturate the bus.
If it's RAM, that would require amazingly slow memory.
In a Powermac 7500 you have 60ns FPM RAM, interleaved. That means for
block ops (multiple reads or write from the same page) you've got 66 MB/s
per SIMM, or 133 MB/s if you're doing interleaving.
With 70ns RAM, that goes down to 114 MB/s.
Xbench measured 105 MB/s for block writes, which is close enough to this
peak that I think I'm probably on target.
> RAM throughput in Quadras as tested is about 16 MB/sec max (see LEM),
> so the PMacs are faster than the Quadras. The bigger step forward in
> RAM/bus throughput was with the better memory controller introduced
> with the Sawtooth G4. Apple really crowed about it when they were
> introduced. De-interleaving the RAM in the 7300-9600 Macs reduced
> the tested speed by 10% or less
Depends on the operations. If you're not doing block ops interleave is
less useful. For random reads or writes interleave doesn't help at
all.
> All the tests
> do the same as far as the documentation states: move data from one
> place in RAM to another, which would consist of one read and one
> write step, thus giving you a number which approached half the real
> memory bandwidth.
So the real memory bandwidth you're talking about is 100 MB/s, not 50,
which is a few percent from what XBench measured and 30% below the 133
MB/s peak for the RAM.
> Even on the normally sprightly PCI
> Rage128 in the Beige G3/500, it's choppy compared to by AGP Rage128
> in my G4/450DP (nice and smooth).
Bus bandwidth. The PCI bus *theoretical* max is that same 133 MB/s as
uncached 60ns RAM in the PM7500 (33 MHz * 32 bits / 8 bits/byte). AGP
is 533 MB/s even for 1X.
Plus, with Quartz Extreme the GPU can do the translucent compositing and
it's probably got GB/s of bandwidth to its RAM.
> Again, I don't know how to test that aside from the full screen menu
> fading I mentioned above. What operations aside from menu drawing
> use this consistently? I'd like to test it in a number of machines I
> have access to.
Title bars, translucent windows (I use translucent Terminals all
the time so I can see what's going on in other apps without losing
any screen space), drop shadows on windows (that's why Shadowkiller
speeds things up) the Dock, ...
--
Unsupported OS X is sponsored by <http://lowendmac.com/>
Support Low End Mac <http://lowendmac.com/lists/support.html>
Unsupported OS X list info <http://lowendmac.com/lists/unsupported.html>
--> AOL users, remove "mailto:";
Send list messages to: <mailto:[EMAIL PROTECTED]>
To unsubscribe, email: <mailto:[EMAIL PROTECTED]>
For digest mode, email: <mailto:[EMAIL PROTECTED]>
Subscription questions: <mailto:[EMAIL PROTECTED]>
Archive <http://www.mail-archive.com/unsupportedosx%40mail.maclaunch.com/>
Using a Mac? Free email & more at Applelinks! http://www.applelinks.com
