Chris,
one question on the conclusion: was the Mandelbrot program set to use floating
point, or fixed-point arithmetic? I’m pretty sure the DSP version was
fixed-point (integer, scaled) arithmetic to make it run faster. The conclusion
might apply to the Pentium’s performance in integer tasks but not be relevant
to floating-point tasks.
- Mark
On Mar 3, 2023, at 4:21 PM, Chris Zach via cctalk
<cctalk@classiccmp.org<mailto:cctalk@classiccmp.org>> wrote:
[EXTERNAL EMAIL]
As part of fixing the Pro/380 I dug out and decided to get running my two Intel
systems. These are Compaq Deskpro/XE systems. One is a 4100 which has an Intel
486/100 (25mhz, quad clock), the other I upgraded with a Pentium P524T
overdrive chip at 83mhz (33mhz external clock).
The P524T was an interesting duck: It's a 5 volt pentium, 32 bit external bus
but they did double the amount of 64 bit on-chip cache so it can perk along
quicker than one might think. Not many were sold, but I have one and there you
go. It even has a little fan on the heat sink that is powered off the chip.
Cute.
The Deskpro/XE's were great systems, slimline, Compaq business audio, QVision
video interface with 2mb of RAM, IDE drive, and oddly enough a 3 slot ISA bus.
Most of the system ran at native 32 bit, so you just ran a slow network card in
the ISA. They also had up to 32mb memory, and an optional memory cache card to
speed things up.
The systems had issues, both on-board batteries were dead, resulting in me
having to find, download, run (not easy) and extract a setup floppy for this
model as you can't do the system settings without it. Not quite an EISA config,
but similar levels of stupidity in the ISA world. And one of them does not seem
to see the ISA bus, but not a big deal as it will just be a DOS floppy maker.
Anyway, finally got one of them running and decided to do some benchmarks.
Booting NextStep 4.2, and tried out a few basic tests.
Findings:
For general booting and such the Pentium does not offer that much of an
advantage. Time to go from login window to system quiet with 20mb memory (I
load several apps by default) is:
486/100-121 seconds
Pentium: 120 seconds
Installing and removing the 256k cache card (an option I have one of) doesn't
change the time much at all, maybe a second.
Boosting memory to 32mb brought that number down to 84 seconds. Moral: Memory
matters.
Then I figured I would try a CPU intensive app: Good old NeXT Mandelbrot. While
a true NeXT slab will kick the rear of any Intel chip (due to the on board
DSP56001) I figured I would put the Pentium up against the 486/100 and running
the 486 at 33mhz external bus (133mhz) in insane overclock mode.`So rendering
the "Valley of Fear" (a complex subset) resulted in:
Pentium, no external cache: 36 seconds.
Pentium, external cache: 34 seconds.
Not bad, cache really doesn't do a whole lot here.
486/100, no cache: 90 seconds. Wow, that is slow.
486/133, no cache: 65s. Faster, but very slow.
So the addition of the Pentium makes a huge difference on floating point CPU
intensive apps. I'm also guessing the extra large cache makes a difference as
well for highly iterative loads.
With this done I can continue looking for a 5.25 floppy to see about making
more PRO disks.
