massive FAH hardware research project

the_technocrat

Here's my guesses, let's fine-tune/add to these lists!

If are able to compare before and after logs when you switch out some different parts, let's see if we can find out wht benefits FAH the most.

A lot of people won't be able to test only one aspect of the hardware themselves, so let's work together and try to find out what changes do the most good.

For example, I might be able to overclock a 2.8Ghz processor to 3.0Ghz, but that also increases the fsb. If I compare my results (let's say a gromacs core) to someone with the same setup, but a stock 3Ghz processor (on a gromacs core), we'll be able to tell if increasing the fsb helps or not.

Sounds hard, and will probably take a while, but I think we can do it. And once we know, it could really guide us towards building the ideal folding rig.

Importance of CPU aspects
1. multiple cores (50% *virtual* decrease in time-to-completion)
2. Front-side bus speed (variable. 4x200mhz (800mhz fsb) to 4x266mhz (~1066mhz fsb) is 25% increase in fsb, and 25% decrease in ttc)
3. HyperThreading (~16-20% decrease in time-to-completion)
4. Speed (variable. 2.5Ghz to 3Ghz is ~16.7% decrease in ttc, for example)
5. On-die L2 cache (~10% decrease in ttc.)
6. Dual-channel memory vs. single channel (8-10% decrease in ttc)
7. Double-sided memory vs. single-sided (3-5% decrease in ttc)

Importance of GPU aspects
1. Speed?
2. Amount of memory? memory type? memory mhz?
3. pipelines?
4. bus architecture/speed? AGP vs. PCIe?

Thrax

1. FPU strength.
2. Clockspeed.
3. FSB.

Datsun-1600

p2124_Ribo_tryptiphan280

P4 840XE 2 x 1Mb cache @ 3.5Ghz = 32m 28s a frame or 54h 6m a Work Unit
P4 920 2 x 2Mb cache @ 3.5Ghz = 29m 34s a frame or 49h 16m a Work Unit

Both run Abit AW8-MAX mobos and Corsair 5400UL with 3.2.2.8 timings, so that is a pretty fair comparison of what the cache does.

To compare a Conroe with the same Ribo.

Conroe E6600 2 x 4Mb cache @ 3.0Ghz = 16m 29s a frame or 27h 28m a Work Unit
Allendale E6400 2 x 2Mb cache @ 2.7Ghz = 19m 55s a frame or 33h a Work Unit

Datsun 1600

Shalimar

the_technocrat wrote:

5. Dual-channel memory vs. single stick? (obviously, enough memory so we're not using the HD...)

On the same pc without Dual-channel memory, 1 gig in single channel mode equals a drop of approx 8 to 10%.

Those numbers are not set in steel but the performance hit is there.

Shal

mirage

Shalimar wrote:

On the same pc without Dual-channel memory, 1 gig in single channel mode equals a drop of approx 8 to 10%.

Those numbers are not set in steel but the performance hit is there.

Shal

Also, there is ~3-5% performance difference between one-sided and double-sided memory sticks. Double sided is faster.

Datsun-1600

From experience the later the motherboard Intel chipset the faster the frame time is, but only by a couple of percentage points, the non Intel chipsets are usually slower.

Datsun 1600

Leonardo

Intel overclocking on P4C 2.8, P4C 3.0, D820, D915, and D930 (multiple CPUs of each). As all multipliers on these CPUs were locked, all CPU overclocking was via FSB boost. DRAM overclocking was usually set at a conservative ratio to the FSB, therefor was not much of a factor in my Folding results. Although this data is not derived from strict empirical testing, I venture to say my assessment is very accurate.

With the processors I listed above, there is linear relationship between CPU overclocking and Folding production increases. The ratio of CPU frequency to production above the baseline rate (at stock) speed is very close to 1:1. In simple terms, a 25% overclock usually resulted in a 20-25% production increase.

Datsun-1600

From memory, a P4 3.46XE (1066fsb and 2Mb cache) @ 3.5Ghz did the same protein at a near identical time as a P4 570J (3.8Ghz, 800fsb and 1Mb cache) @ 4.0Ghz.

Datsun 1600

the_technocrat

Thrax wrote:

1. FPU strength.
2. Clockspeed.
3. FSB.

yes, but I'm trying to find a way to translate FPU strength into newegg search results...

the_technocrat

Datsun 1600 wrote:

From experience the later the motherboard Intel chipset the faster the frame time is, but only by a couple of percentage points, the non Intel chipsets are usually slower.

Datsun 1600

got an estimate of %? I wonder what the later northbridges have that benefits FAH? I'm sure it's some kind of memory management...?

the_technocrat

Leonardo wrote:

Intel overclocking on P4C 2.8, P4C 3.0, D820, D915, and D930 (multiple CPUs of each). As all multipliers on these CPUs were locked, all CPU overclocking was via FSB boost. DRAM overclocking was usually set at a conservative ratio to the FSB, therefor was not much of a factor in my Folding results. Although this data is not derived from strict empirical testing, I venture to say my assessment is very accurate.

With the processors I listed above, there is linear relationship between CPU overclocking and Folding production increases. The ratio of CPU frequency to production above the baseline rate (at stock) speed is very close to 1:1. In simple terms, a 25% overclock usually resulted in a 20-25% production increase.

great info, thanks.

the_technocrat

Datsun 1600 wrote:

From memory, a P4 3.46XE (1066fsb and 2Mb cache) @ 3.5Ghz did the same protein at a near identical time as a P4 570J (3.8Ghz, 800fsb and 1Mb cache) @ 4.0Ghz.

Datsun 1600

interesting. anyone know how much time it took?

the_technocrat

I've revised the list to make fsb more important, and added dual-sided RAM. Here's what we have so far, I'll get some more info on monday from the machines at my work:

still kinda hazy, but I think it's coming together...

Datsun-1600

the_technocrat wrote:

I've revised the list to make fsb more important, and added dual-sided RAM. Here's what we have so far, I'll get some more info on monday from the machines at my work:

still kinda hazy, but I think it's coming together...

Number 8 on your list should be a 3.46XE, the E6600 has a 334fsb and the E6400 has a 340fsb. Number 8 and 9 are equivalent to the 920 @ 3.5Ghz (14 x 250fsb), only with single core.

The main gist of what I am trying to say is that more cache is highly desirable when contemplating a CPU purchase for F@H, whether it be AMD or Intel.

Datsun 1600

the_technocrat

if anyone wants to fill in the gaps in that excel sheet and re-upload, feel free. I've made some of the changes D1600 mentioned. (I'm bound to be off somewhere... )

the_technocrat

Just got my brother's machine up and running. Unfortunately, both cores are working an amber units, so I can't compare to our previous gromacs numbers before I have to ship it out.

Dell Inspiron E1505
CPU: Intel C2D T2250 1.73Ghz, 2x1MB L2, 533mhz FSB
Memory: 2x512 dual-channel DDR2 SDRAM, 533mhz, 2 DIMM

Time for 1 core to complete 1 step: 26m 58s

muddocktor

For checking on the efficiency of your machine by platform, a guy on the Overclockers.com folding team has done quite a bit of research on it and has compiled a thread on this in their folding forum.

This is a link to the thread. If you have questions about how your machine stacks up to others of it's type or how 1 processor line compares to another with processing various wu's, this is a good reference.

lemonlime

Thought this might be useful.. I did a fairly large comparison between numerous K8 based processors for my AM2 platform review in May. Gives some contrast between Cache and clockspeed..

the_technocrat

NICE

I'll look up the stats on these a little later (busy here mon morning) and add them to the list. Unless someone knows off the top of their head...