You know what this is beginning to remind me of?

Back when my hair and beard were dark brown, I remember I'd take a Socket 7 motherboard with L2 cache on the board itself, and I'd add an AMD K6-III+ chip, and the L2 would "become" L3. Then the challenge was to run the speed of the FSB up as high as possible (usually determined by limited options on the motherboard) and select as high an even multiplier as you could get.

Your overclock of 215 on the FSB looks like it's really helping your times (or maybe that's the AVX and r557 build in combination).

I've just gotten a result from my 1045T that make me think the FSB is really helping. It was 60,284 seconds. That's at 2.7GHz. Without knowing all the particulars of the WU we really don't know anything.

I'm not thinking about trying to guess which part of a WU takes how long to do (at least not tonight) so I know I'm being incredibly sloppy with this "analysis from 30,000 feet."

BUT, it's amazingly "in the ballpark" of the numbers being generated by my other processors which are running 10-20% faster. Yet, the 1045T is in the general neighborhood of 20% faster than the *old* Phenom 9550 running at 80% of the speed.

(I'm saying that clock-for-clock, the 9550 and 1045 seem to share a "work done per clock" characteristic, but increases in the CPU clock multiplier beyond 2.7GHz don't seem to be "scaling." That leads me to believe that the FSB speed is the major bottleneck.)

BTW - I may be saying that entirely incorrectly under "modern lingo" with the RAM controller on the CPU, itself. Or not. I don't know.

I wonder what happens if you reduce the multiplier way down (run a 3.2GHz CPU at 2.2GHz) and jack the FSB up as fast as your RAM will allow while running a low, but even multiplier?

I'll have to try something like that when I have the time to goof-off that much.

I know, it's hard to believe I was ever considered "an expert" on Compuserve's PC Hardware Forum when it came to this stuff. "Back when I went to school we used to walk to school uphill both ways, eat rocks and dirt for dinner, and squeeze oil out of goldfish to read our school scrolls and do homework with lumps of coal on papyrus if we could see through our coke bottle glasses made with real coke bottles because the 6oz coke only cost a nickle. ...and we liked it!"

Yeah I'm rambling...

How much wingman hate do you think I'd generate by firing-up an old Socket 7 AMD chip (I don't know if under that postage-stamp sized heatsink and fan is a K6-2 or K6-III) installing Win2k (if that'll run), if I can find a HDD that doesn't exceed a BIOS limitation, and let it crunch whatever it can crunch provided the AT-style PSU doesn't smoke the minute it's turned-on?

I can see it now... 1,500,000 seconds on an AP WU.

Same concept. A good test along those lines could be to drop the memory speed a little & tighten the timings as far as they'll go while staying stable. The thing about cache, since we 'know' it's a big factor here & is designed to hide these latencies, is if cache misses are creating stalls, which are very expensive, then there is plenty of latency to be hidden & you're unlikely to be using anywhere near the rated bandwidth due to waiting cycles. IOW, if you are stalling due to waiting on memory (latency as opposed to bandwidth) then more CPU GHz won't help.

You could do some back of the envelope calculations: RAM latency(CL in cycles ) x 1/freq to get an idea of how many microseconds each setting might achieve/cost on cache misses, but probably easier to try each setting & closer monitor CPU core temperature at a fixed CPU fan speed. A CPu core averaging more instructions per cycle (not stalling so much on cache misses) should be marginally warmer.

Jason

---

"Living by the wisdom of computer science doesn't sound so bad after all. And unlike most advice, it's backed up by proofs." -- Algorithms to live by: The computer science of human decisions.

For Astropulse GPU is the way to go.
My 1100T will take 10 hours to do one while my 460 gtx will crank it out in two hours doing two at at time.

That's really the bottom line

You are right, of course.

I'm considering buying an ATI card (supposedly faster still) to crunch AstroPulse only. The idea is that with the money saved on electricity I can own the card.

I'm going to have to research that a little more carefully before I do anything.

Few more APs
FX 8120@ 3.9ghz 18multi 215fsb
Ram 2012 CL10

7.52-blanking 34509.92/sec 1singPulse 0repPulse
73.86-blanking 37737.71/sec 4singPulse 0repPulse

FX 8120@ 3.1ghz 14multi 215fsb
Ram 2000 CL10

0-blanking 42764.32/sec 4singPulse 3repPulse
4.84-blaning 42277.26/sec 3singPulse 0repPulse
9.53-blanking 42884.07/sec 3singPulse 0repPulse
16.66-blanking 43184.48/sec 1singPulse 1repPulse

temps back to "normal" about56c with 4 APs running,room temp was 22c.

Running Lunatics' optimized apps:

The Phenom 9550 @2.2GHz is completing WUs in about 90,000 seconds compared to the Athlon 64 x 2 @ 3GHz at about 120,000 seconds; and of course the Phenom is able to do 4 at a time instead of 2 at a time. (assuming you can get enough work to keep all the cores busy with AstroPulse)

Yeah, I know, either case is pretty sad by power-crunchers' standards.

But by the time you cut 25% off of the times of each unit and double the number of units, at least the Phenom is significantly less sad.

Non-AVX, FX-8120, stock clocks


http://setiathome.berkeley.edu/workunit.php?wuid=971631265
23,600.72 seconds
30 single pulses and 30 repeating pulses, exiting.
percent blanked: 85.10


http://setiathome.berkeley.edu/workunit.php?wuid=972110044
49,719.02 seconds
single pulses: 26
repetitive pulses: 30
percent blanked: 0.00


http://setiathome.berkeley.edu/result.php?resultid=2399604067
60,811.92 seconds
single pulses: 1
repetitive pulses: 0
percent blanked: 0.00


http://setiathome.berkeley.edu/workunit.php?wuid=966159816
63,535.79 seconds
single pulses: 30
repetitive pulses: 23
percent blanked: 0.00

With completion times all over the map like this, are we learning anything?

Why dont you use r555 or r557?
Its faster on your FX.

he is running non-AVX 548 on special request.

I'm happy to quit anytime now. I'm also happy to let it keep going.

I guess I'm just happy; but I want to know if you are still watching / interested?