Slashdot Mirror
Comparative CPU Benchmarks From 1995 to 2004
Lux writes "The guys over at Tom's Hardware Guide have been busy recently! They've compared over a hundred different architectures dating all the way back to the Pentium 1 in one huge benchmarking effort. Looking to upgrade an older system? Unlike most benchmarks, which compare modern systems to other modern systems, these charts can help you figure out if the cost of upgrading is worth the speedup or if you should hold off for a bit longer."
You can get this kind of info elsewhere
They haven't been busy recently. They just updated the guide they did quite some time ago. Not very much new to see here...
Because megahertz is not totally where it's at, contrary to Intel's marketing hype. I'm currently running an Athlon XP-M 2500+ overclocked to 2.4ghz, and it performs faster than a 3.2ghz Pentium 4. An Athlon 64 3200+ runs at 2.2ghz iirc, and outperforms a 3.2ghz p4 by an even larger margin.
With multi-core chips and on-die memory controllers, the benefits of performance will be felt, even if the clock speed is constrained to 4ghz for now.
Part 1: http://www.tomshardware.com.nyud.net:8090/cpu/2004 1220/index.html4 1221/index.html
Part 2: http://www.tomshardware.com.nyud.net:8090/cpu/200
http//injoke.org -- Culling The Interesting
I can't get to the article, so I have to wonder what kind of benchmarks they used that were consistant on all those platforms. Old benchmarks tend to freak out on newer hardware, and I can't imagine newer benchmarks running properly on older hardware.
http://crummysocks.com
The article is a bit slashdotted but it looks like it doesn't go back all that far.
Just a teaser, I have been running a collection of benchmarks since the Pentium 90.
At the time, I was involved in a huge UNIX engineering workstation benchmark. I felt we needed something more constant than the applications to compare performance (the engineering apps constantly change). So I quickly assembled everything I could find that could be easily run. These are mostly 'toy' benchmarks, but the results are still interesting.
For these int benchmarks, higher is better:
c4.s c4.64 dhry21 hanoi heapsort nsieve nsieve TOTAL
Kpos/sec Kpos/sec MIPS mvs/sec high High Low
MIPS MIPS MIPS
P 90 92.7 94.2 68.6 51.2 43.55 111.0 33.3 494.6
md64b 4050.1 4167.8 4914.3 2708.8 3333.7 3333.7 610.4 21782
Float: Higher is better, except for the fft's.
flops20 fft tfftdp
MFLOPS MFLOPS MFLOPS MFLOPS TOTAL time time
(1) (2) (3) (4)
P 90 13.3 12.8 18.1 23.8 68.0 3.07 16.81
amd64 1120.9 1004.3 1480.9 1834.7 5440.8 0.04 0.42
The P90 was running RedHat. The AMD64 is my new desktop, a 90nm 3000 OC'd to 2430 Mhz. My data also includes systems from DEC, HP, IBM, Sun and SGI. I also ran 10 matrix multiply benchmarks as part of the effort.
I have never gotten around to publishing the results or the collection of benchmarks.. Maybe it is time.
Actually I shouldn't give Tom's Hardware a hard time (like everyone else seems to). As articles go, the reviews of high-end ink-jets, the 8-channel RAID6 card and the Viewsonic media center were quite interesting (and a lot more recent than the CPU round-up too).
These days though, my favourite reviewer is Dan (who posts here now and then). Dan seems to understand that a million graphs showing you the statistically insignificant difference between the latest mobos / graphic cards / processors / ram sinks don't really make a great site.
The 486 has had a built-in math coprocessor ever since it debuted. After the 486 was around for a while, they made a stripped down version without a coprocessor called the 486SX. The plain 486's were called 486DX.
You could get a coprocessor for the 486SX, but not the DX. From what I've heard, the original 486SX's were actually re-badged 486DX's whose math coprocessor unit was either not functional or just disabled. When you bought the 487SX "co-processor" you were actually buying a fully functional 486DX that disabled the other CPU on the board.
However, beware of the 486DX50 vs the 486DX250. The 486DX 50 was a true 50Mhz part whereas the DX2 were only 50Mhz internal to the chip with the bus running at 25Mhz. Same thing for 486DX2 66's. Most programs ran slower on them than a trus 486DX50 due to the slower (33Mhz) bus speed.
Yup. Then Intel had to confuse the issue by releasing the 486DX4. Just as the DX2-50 had a 2x multiplier with a 25 mhz bus and a 50 mhz core speed, you'd think the DX4-100 would have a 4x multiplier with a 25 mhz bus and 100 mhz core speed. But it was actually a 3x multiplier, with a 33 mhz bus speed. They should have caused it a DX3
It would make about as little sense. On many CPUs (particularly the CISC CPUs), instructions take wildly different amounts of time to complete. A NOP might complete in one clock cycle, while an obscure legacy instruction might take twenty. Running only NOPs, the CPU would be 1 BIPS if driven at 1 GHz. Running the other instruction, it'd be 50 MIPS. Somewhere in the middle would be the truth.
The fact is, it is next to impossible to accurately digest CPU performance down to a single number that can be comparable across architectures and across the variety of actual usage. Even something as obvious as instruction cache size has a different real world effect depending on the sizes of your instructions!
AMD was a second-source for Intel CPUs up through the 286 era. I believe this arose out of IBM's requirement to have a second source for whatever CPU it picked for its PC. It appears Wikipedia corroborates my story.
--JoeProgram Intellivision!
Never mind that the multipliers would be through the roof, on a box from the days where multipliers weren't even needed at all sometimes.
'twas a i486 with a fully functional FPU and a Voodoo video card, so running half-life wouldn't be so dificult if it was really at 247 mhz.
the project, of course, was a joke. it was called E.U.N.U.C.H. for "The Extreme Use of Nearly Universal Cooling Hardware", the page is here. enjoy.
What ? Me, worry ?