A benchmarking-we-will go! K6-3 vs. WinChip 200.
Sandra benchmarks:
CPU: PR = CPU's PR rating as reported by Sandra
Dhry = Dhrystones
Whet = Whetstones
MMX: MMX = MMX iterations/s
FPU = FPU iterations/s
RAM: ALU = MB/s ALU
FPU = MB/s FPU
Real-world benchmarks:
MP3: % of CPU reported by M$'s W98 CPU load monitoring utility.
This utility frequently produced bogus results (e.g. 4%)
sometimes, and then sane results (e.g. 25%) later. Where
I got a believable number, I used it. YMMV wildly 
Sometimes I'll add an FPS number for Geiss.
JPG: Time to decode/view a 1280x960 .JPG as reported by a JPG viewer.
MAP: Seconds to render a 35000-node 3D image of a topographical
map of a location chosen at random. Both CPU dependant and
heavily RAM-constrained. The unit used for testing had only
32M of RAM. The goal was a test that didn't take undue
advantage of the K6's beefy 256K cache.
NSI: Just for yuks, the old Norton Utilities "SI.EXE" benchmark
from a pure DOS environment. The number is "number of times
faster than a 4.77 MHz IBM PC". Two numbers are given because
the benchmark divides by the number of 1/18-of-a-second clock
ticks, and it now runs so quickly that this is statistically
significant. But it's fun to see how far we've come from the
days when these numbers ("12 times faster than a PC!") were
actually taken seriously. You old-sk00lers will enjoy.
Thermal: CPU/Mobo "open" and mobo "closed".
All thermal numbers are based on:
1) "Opened: CPU/Mobo in degrees Celsius with all covers removed."
2) "Closed: CPU/Mobo in degrees Celsius with all covers removed."
"Closed" numbers only available for a couple of configurations.
If you run any tests with the covers removed, you'll have to guess
based on my numbers at how much hotter it'll run with the covers
in place. As with all benchmarks, YMMV. With my thermal numbers,
*expect* your mileage to vary. :)
3) I'm using the stock (passive) heatsink.
4) I have a very small fan (probably cools CPU temp by 5-6 degrees
Celsius based on my earlier WinChip-200 testing) blowing a very
light breeze across the passive heatsink. This probably gives
zero cooling effect on Q16.
5) No cooling measures taken for Q16. NOT recommended for actual use!
I ran my tests this way for benchmarking purposes only so I could
judge the success of my future Q16-cooling efforts!
6) Temperatures are recorded after at least an hour or two of operation
with a CPU-intensive task (MP3 playback with fullscreen visualization)
and no software-based "CPU cooling" solution in effect.
7) My Vcore for the K6-III was always around 1.9V. If you run
the chip at a higher voltage (2.0V to 2.2V), expect your heat
generation numbers to go up accordingly.
Summary:
A K6-III-200 is about twice as fast as a WinChip 200 and doesn't
add too much additional heat. Probably most bang for the buck
with minimal long-term risk to your hardware.
Taking a WinChip 200 as the baseline, we see a K6-III-200 runs
about twice as fast in real-world tasks (rendering a .JPG, or
drawing a big honkin' map) as the WinChip-200, even at the same
clock/multiplier settings. You also get MP3 visualizations that
you couldn't dream of on a WinChip.
If you're interested in an MP3 player, the CPU usage figures don't
tell all the story. A drop from ~50% of CPU to ~36% of CPU can be
obtained at 200 MHz, but the vastly superior (by a factor of 7!)
FPU performance gives you much better visualization performance.
Geiss should be OK at 166 or 200 MHz, and fine at 233 MHz and above.
Whitecap is fine even at 166 MHz. (I love Geiss, but if you're using
the IO's original keyboard, its lack of an ESC key is a real killer.
The only way out of Geiss without an ESC key is to pull the plug!)
WhiteCap is a little lighter on CPU load and almost as interesting
to look at, and you can quit out of it with a single mouse-click!
For the number-crunching crowd, you can take things up to about
"triple" the performance of the original a WinChip 200 if you've
got a good cooling solution for Q16 before you run up agains the
current limitation imposed by R302.
If you wanna go past 300 MHz, you'll have to work around R302, and
you'll definitely need a serious cooling solution, probably a combo
of of HLT instructions (software-based) and an active (heatsink/fan
combo) cooling of Q16 and L32. You're probably voiding your warranty
by even *thinking* about it
Again, these are benchmarks. Take 'em with the grain of salt they're
made of. YMMV, and it probably will. And given the heat concerns,
if it's fast enough for what *you* wanna do with the unit, consider
backing off on the CPU. Why burn your box down to get cycles you're
never gonna use?
OK, on to the numberz!
-------------------
WinChip 200 (3x66):
-------------------
Sandra 2000 benchmarks:
CPU: PR = PR201
Dhry = 253
Whet = 78
MMX: MMX = 169
FPU = 77
RAM: ALU = 53 MB/s
FPU = 56 MB/s
Real-world benchmarks:
MP3: ~50% (All visualizations were dog-slow, jerky, etc.)
JPG: 3.1 seconds
MAP: 75 seconds
NSI: 331/350
Thermal: ~45-50C on both mobo and CPU in closed configuration.
--------------------
K6-III-166 (2.5x66):
--------------------
Sandra 2000 benchmarks:
CPU: PR = PR201
Dhry = 398 (Which just goes to show you that PR ratings
Whet = 205 are just "PR"! Beats hell outa WinChip!)
MMX: MMX = 404
FPU = 595 (Beats SERIOUS hell outa WinChip by 7:1 margin!)
RAM: ALU = 67 MB/s
FPU = 76 MB/s (Even at same bus speed. Cool!)
Real-world benchmarks:
MP3: 36% (5.7 FPS in Geiss looks marginally-OK, Whitecap looks fine.)
JPG: 1.9 seconds
MAP: 45 seconds
NSI: 530/550 (The entire .EXE fits in the 256K of cache)
Thermal: 45C CPU / 61C Mobo, closed, ~1 hour of MP3 with fullscreen vis.
Heatsink was hot to touch, comparable to WinChip 200 "feel".
--------------------
K6-III-200 (3.0x66):
--------------------
Sandra 2000 benchmarks:
CPU: PR = PR241
Dhry = 490
Whet = 247
MMX: MMX = 486
FPU = 716
RAM: ALU = 70 MB/s
FPU = 79 MB/s
Real-world benchmarks:
MP3: 30% (6.9 FPS in Geiss looks OK due to slow refresh of LCD screen)
JPG: 1.5 seconds
MAP: 40 seconds
NSI: 609/643
Thermal: 48C CPU / 64C Mobo, closed, ~1 hour of MP3 with fullscreen vis.
Heatsink was hot to touch, comparable to WinChip 200 "feel".
---------------------
K6-III-250 (2.5x100):
---------------------
Sandra 2000 benchmarks:
CPU: PR = PR301
Dhry = 614
Whet = 309
MMX: MMX = 610
FPU = 898
RAM: ALU = 95 MB/s
FPU = 107 MB/s (Bus speed to 100 MHz shows up)
Real-world benchmarks:
MP3: 26%
JPG: 1.2 seconds
MAP: 38 seconds
NSI: 758
Thermal: 36C CPU / 56C mobo "open".
Not run in closed config. See "-300" numbers for guess.
--------------------
K6-III-266 (4x66):
--------------------
Sandra 2000 benchmarks:
CPU: PR = PR320
Dhry = 653
Whet = 324
MMX: MMX = 648
FPU = 955
RAM: ALU = 73 MB/s
FPU = 80 MB/s
Real-world benchmarks:
MP3: 31%
JPG: 1.2 seconds
MAP: 34 seconds
NSI: 788
Thermal: 36C CPU / 56C mobo "open".
Not run in closed config. See "-300" numbers for guess.
------------------
K6-III-285 (3x95):
------------------
Sandra 2000 benchmarks:
CPU: PR = PR344
Dhry = 702
Whet = 344
MMX: MMX = 697
FPU = 1026
RAM: ALU = 93 MB/s
FPU = 103 MB/s
Real-world benchmarks:
MP3: 26%
JPG: 1.2 seconds
MAP: 30 seconds
NSI: 900/954
Thermal: 38C CPU / 62C mobo "open".
Not run in closed config. See "-300" numbers for guess.
This system was unstable with the stock R302 even with voltage
tweaked from 1.872 to 1.92V. Oddly enough 4.5x66=300 *was*
stable. High bus speeds evidently draw more current than high
multipliers.
I also tried 3x100=300, but was unable to keep it running for long
enough to do meaningful tests, even with R302 jumpered over with a
piece of bare wire. Unsure whether it was the under-voltage (1.9V
vs. 2.2V spec for K6-III, far too hard on Q16) confusing the CPU,
or a real voltage *drop* induced by high current draws of the chip,
I aborted any testing on configurations unstable with the stock R302.
There are reports of varying success at high bus speeds with the
K6-III. If you can run stably at 3x95 or 3x100 or 3.5x95, more
power to ya. Just be sure you can adequately cool Q16.
--------------------
K6-III-300 (4.5x66):
--------------------
Sandra 2000 benchmarks:
CPU: PR = PR361
Dhry = 737
Whet = 371
MMX: MMX = 731
FPU = 1078
RAM: ALU = 74 MB/s
FPU = 81 MB/s
Real-world benchmarks:
MP3: 30% (9.5 FPS in Geiss looks very nice on the IO's screen.)
JPG: 1.1 seconds
MAP: 30 seconds
NSI: 893/943
Thermal: 38C CPU / 61C mobo "open"
55C CPU / 80C mobo "closed".
Heatsink was *TOO* hot to touch, very slight discoloration
around Q16 after 1-2h of torture. Not recommended without
serious cooling solutions for Q16. L32 was also hot enough
that I didn't want to hold my finger against it.
This system was stable with the stock R302 resistor and Vcore tweaked
down to 1.95V, unlike the competing 3x95 = 285 MHz config.
