Is there a performance database?

[Joe O]

Pentium
Pentium MMX
Celeron
Pentium PRO
Pentium III
Pentium IV
AMD K6-2
AMD K6-III
Athlon
Duron
Tbird
Xp/MP
Tbred
How do these stack up? What work are they best suited for? Minimum speeds? Recommended memory?

[Lumly]

You mean like this...

http://www.mersenne.org/bench.htm

8)

[Prime95]

Originally Posted by Joe O

How do these stack up? What work are they best suited for? Minimum speeds? Recommended memory?

My recommendations (from worst CPU to best) - you can do whatever you like though:

AMD K6 - to be blunt they are better off on a different distributed project.
They have lousy FPUs.

Pentium, Pentium Pro - a little underpowered by todays standards. Good for factoring, or if very patient a double-check.

P-II, slow celerons - double-checking, but factoring or first time tests are OK

P-III, fast celerons - first time checking or double-checking

Duron, all Athlons - first time checks, maybe 10,000,000 digit tests

P4 - first time checks or 10,000,000 digit tests

As to memory - most of the time the client uses very little (say 1MB factoring and 10MB running LL tests)

[Stormblade]

Well, that list gives you an idea of how fast they are in respect to each other, but not really what they are suited for..

Currently, the client itself is pretty well set up for deciding what work to do on what boxen.

less than a Pentium 233 = Factoring
Less than a P2 550 = Double checks
of course, you can do either of these with faster machines if you want short "workunits" tho it is currently better to do factoring on a non P4 system (the SSE2 code in the beta doesnt kick in till the 64 bit or higher passes)

As for the amount of memory used, of course the more you use, the better chance you have of finding a factor during P1 stage 2 testing and saving yourself a lot of time.. I think current doublechecks (8M range) will use 2-300 MB for optimum checking. Dunno about first time LL or 33M exponents.. Anyone care to help out on these?

Did this answer your question better?

[edit]
Bah, George beat me to it by a minute.. :mrgreen:

[willmore]

One other note, for an SMP system--pre-opteron--you will have main memory bus bandwidth contention issues if more than one processor is doing an LL test (or an FFT based calculation) of any kind.

For a 2x SMP (like my little 2x333PII) I have one CPU doing DCs and the other doing factoring as factoring stays in L2 cache and doesn't clog the bus.

For higher then 2-way SMP, it may start to depend on the exact topology of your computer. Most Intel systems use 'clusters' of 4-way or 8-way 'nodes'. There is too much variation to give a hard and fast rule.

With Opteron systems, (based on what I've read, I'm not lucky enough to *have* one) if each processor keeps to local memory, each processor may calculate at full speed even when doing FFTs (used in LLs, DCs, P-1s, and ECM)--sinceeach processor has a local chunk of memory directly attached. Given some of the BW figures I've seen thrown around, one of these processors might be able to run at full speed from a relatively close neighbors memory.

[Tasuke]

On the opteron, a processor working out of annother's local memory will react as though it was using a memory controller/chipset topology. Running out of its own bank, the Opteron will have a 64bitX2channel system, and the latency will be nearly halved due to the on-chip, Full core speed memory controller. This may not be as noticable on the Opteron, as it will be using the SSE2 instruction set, which seems to do a good job of hiding the bottleneck, thus the reason why SDR on a p4 performs so well in comparison to DDR, where on an other platform, DDR can be better than a processor upgrade(like a fast AXP).

[Lord_Humungous]

Is there a list that expresses performance in P90 CPU years?

I saw this on Primenet: (*Measured in calibrated P5 90Mhz, 32.98 MFLOP units: 25658999 FPO / 0.778s using 256k FFT.), but I'm not sure what to do with it.

Since I horribly lazy, I was hoping someone had come up with a list of different uP/systems.

[Prime95]

Originally Posted by Lord_Humungous

Is there a list that expresses performance in P90 CPU years?

Buried in the footnotes of http://www.mersenne.org/status.htm page is the factoid: A PII-400 is equal to 5.5 P-90s. I suppose you could compare your machine timings to the PII-400 timings on that same
web page to determine your P-90 equivalents.

[Lord_Humungous]

Close enough. Thanks George!

[Complex33]

If you’re interested I have thrown together an Excel Spreadsheet that relates three FFT lengths to a number of processors and their equivalent P90 years. The information was gleaned from the benchmark page. Quite interesting to see which processor is best for which type of exponent (as shown in green).

Hope this is of some use.

http://www.teamprimerib.com/xls/Prim...0Relations.xlsoddslot


Complex33

[Lord_Humungous]

I'm hoping you didn't do that just for me...right? ops: