Does RAM timing really matter?

I ran a series of tests on my computer using various BIOS settings to see how much of a difference things make, with the following results:

Test 1: 1250 MHz(L1 on, L2 on), 100 FSB, 200.0 RAM, 2.5/3/3/6 timing
3468 Dhrystone 1740 Whetstone 1353/1253 MBps RAM bandwidth

Test 2: 1667 MHz(L1 on, L2 on), 133 FSB, 333.3 RAM, 2.5/3/3/6 timing
4585 Dhrystone 2301 Whetstone 1983/1864 MBps RAM bandwidth

Test 3: 1750 MHz(L1 on, L2 on), 140 FSB, 280.0 RAM, 2.5/3/3/6 timing
4817 Dhrystone 2418 Whetstone 2016/1872 MBps RAM bandwidth

Test 4: 1750 MHz(L1 on, L2 on), 140 FSB, 346.6 RAM, 2.5/3/3/6 timing
4818 Dhrystone 2418 Whetstone 2089/1966 MBps RAM bandwidth

Test 5: 1750 MHz(L1 on, L2 on), 140 FSB, 346.6 RAM, 2/3/3/6 timing
4818 Dhrystone 2418 Whetstone 2095/1981 MBps RAM bandwidth

Test 6: 1750 MHz(L1 on, L2 on), 140 FSB, 346.6 RAM, 2/2/2/5 timing
4832 Dhrystone 2418 Whetstone 2116/2009 MBps RAM bandwidth

Test 7: 1750 MHz(L1 off, L2 off), 140 FSB, 346.6 RAM, 2/3/3/6 timing
28 Dhrystone 29 Whetstone Memory bandwidth test cancelled after 20 minutes

Reference point: 486DX2 66MHz
85 Dhrystone 31 Whetstone

In particular, the difference between CAS 2 and CAS 2.5 is about a third of a percent, and the difference between the least aggressive timings and the most aggressive is 1.25%. For comparison, enabling the L1 and L2 caches increases performance by a whopping 17,107%!

Asus A7V333, Athlon XP2000+, 512MB Corsair PC2700 CL2, and assorted other parts.
Benchmark program: SiSoft Sandra

heheeh try that again with sandra... i'm sure you will see a huge difference at 140 fast then 140 slowest timings..

I wish I had read your post a few minutes earlier. I was asking a question from Crucial/Micron, and part of the response (on PC133 SDRAM) was the following:

Crucial Technology PC100 SDRAM DIMMs each have a letter near the end of the module's part number which designates the timings of the module. The following chart shows the different timings for each type of memory:

CAS Latency tRP tRCD
A 3 3 3
B 3 2 3
C 3 2 2
E 2 2 2

These different options are all equally compatible when used on a 100MHz bus, but if you use all of one of the above part type in your PC there is a small performance difference. The performance difference between an A and a B is 0.7% (the B being the faster of the 2). The difference between the B and the C is 0.1%, and the difference between the C and the E is 0.5% (the cumulative difference between the A and the E is 1.3%).

Click to expand...

What's remarkable is that they came up with a 1.3% figure, just like your 1.25% figure.

I guess won't be shopping for Mushkin Rev 3+ RAM after all ...

-- Paul

Well i will tell you this... It does make a huge difference with gaming!

I did some testing of my own and i seen an easy 20 frame differance in Quake 3 in demo 4 with 20+ giving to the higher timings...

Your findings may be different with each system though..

KnownKiller said:

Well i will tell you this... It does make a huge difference with gaming!

I did some testing of my own and i seen an easy 20 frame differance in Quake 3 in demo 4 with 20+ giving to the higher timings...

Your findings may be different with each system though..

Click to expand...

That's quite a difference! Nothing but the RAM timings were changed in those calculations? No FSB changes, etc.?

At the end of the day, I'm most interested in Memory-FSB-CPU interactions, as I do numerical analysis. If it's only a 1.3% improvement for those interactions (as opposed to adding in the effects of communication over the AGP with the vid memory) ....

Thanks! -- Paul

Nope same everything.. Only thing i change was the timing of the ram from cas 2.5 4 way 3-6-3 2T to cas 2 4 way 2-6-2 1T... Thats it.

I did some testing of my own last night: I tried 2-2-2 and 3-3-3 at several bus speeds:

66 MHz: (I've never posted as a PIII 600 before! The thing idled at 6C above case temp.
Ints: 2.37% improvement
Floats: 2.77% improvement

133 MHz:
Ints: 3.58% improvement
Floats: 3.08% improvement

135 MHz:
Ints: 3.64% improvement
Floats: 3.36% improvement

The variation may well be due to rounding of bus speeds by Scisoft Sandra. (After all, bus speeds are more like 133.05, or 135.48, etc., but the MB/s are given in whole numbers, as are bus speeds in Sisoft Sandra.) So, really, we can't trust the decimal places for accuracy. This is probably more accurate:

66 MHz: (I've never posted as a PIII 600 before! The thing idled at 6C above case temp.
Ints: 2% improvement
Floats: 3% improvement

133 MHz:
Ints: 4% improvement
Floats: 3% improvement

135 MHz:
Ints: 4% improvement
Floats: 3% improvement

I can't run my memory at 2-2-2 at any FSB higher than 135 MHz. (After all, it is CAS3 RAM. , so I can't further validate, but I think it's safe to say that memory bandwidth generally isn't affected by more than 5% from the slowest timings to fastest timings at any FSB.

-- Paul

I ran similar test when putting my new system together over the past two weeks.

There was never more than about a 2% difference between CL2.5 and CL2.0 using the same timing settings.

There is a huge difference in running the memory at DDR400 at CL2.5 and DDR333 at CL2.0, however.

texasfit said:

I ran similar test when putting my new system together over the past two weeks.

There was never more than about a 2% difference between CL2.5 and CL2.0 using the same timing settings.

There is a huge difference in running the memory at DDR400 at CL2.5 and DDR333 at CL2.0, however.

Click to expand...

Cool.

Yeah, I think if someone wants to make real (and I mean substantial and worth the price of the investment) gains in memory performance, it's a matter of getting a better version of memory (DDR400 vs DDR333, etc.), rather than timings within the same version of RAM.

I was considering getting some expensive PC150 SDRAM so I could run 2-2-2 at my max overclock of 156 MHz FSB, but now that I see that the $180 investment would probably only boost my memory bandwidth by maybe 5% (substantial, true, but $36 per percentage point?), I see that a better investment would be a new motherboard with a memory capacity of more than 512 MB. (I'm running into memory barriers in my scientific computation.) And in turn, it would be an even better investment to save my money for a year and invest in a new motherboard, RAM, and CPU....

Well, I've gotta' run! Thanks for sharing your results, and have a good one! -- Paul

I think that Sandra is a poor benchmark for testing the differences between memory timings. AFAIK Sandra is basically raw thru put. Changing the timings change how quickly the cpu can access the ram. More appropriate benchamrks would be memory intesive apps like Quake 3 or even Seti - Sandra which simply streams. From experience, changing the timing on my mmorgan system from cas 2 to cas 3 increases a seti unit time by 30 minutes.

That's very interesting.

I'll try to do some computational benchmarking, say, I'll generate a few large random matrices, save 'em, and time the inversion times for various ram timings. -- Paul

"The Point" is in doing better with what you have. Not everyone can go out and splurge $300 on RAM that's 5% faster, but everyone can try to run the RAM they have at higher settings. Sure, if you're in the market for new mem, then you need to figure out what you want and can afford. Get what you feel is best for you and your wallet, and then push that, don't go trying to buy less than what you want and hope you can push it higher; that's not what OCing is all about, and it's a big gamble as to whether you'll get what you want from it.

Well put!!! -- Paul