Sempron 2800+ 64-bit E6 benchmarks

[hurtstotalktoyou]

Hey, all. I just got my new Sempron 64 2800+ setup, and I thought I'd share some benchmarks

SETUP:
Sempron 2800+ (64-bit, E6 revision)
Epox EP-8KDA3J (nForce3 250GB)
512 MB PC3200 3-4-4-8 DDR SDRAM
5400 RPM ATA/100 hard disk

AUDIO TEST:
Convert a 38:27.51 44.1KHz waveform to 48 KHz in Adobe Audition 1.0 using 300 quality setting

no overclocking: 1.6 GHz CPU, 800 MHz HT, 200 MHz RAM
458 sec. (base)

overclock #1: 1.92 GHz CPU, 720 MHz HT, 200 MHz RAM
432 sec. (6.0%)

overclock #2: 2.13 GHz CPU, 798 MHz HT, 177 MHz RAM
414 sec. (10.6%)

As you can see, the performance benefit from a 33% CPU overclock was less than 11% in the real world. This might be due to the 11% underclocked RAM, or perhaps the hard disk bottleneck, but the end result remains rather disappointing. Hopefully a video benchmark might provide a wider gap.

I'll run the video tests, later.

 

[Hacp]

Thats the highest you'll go? . Have you tried tightening your timings with the 177 MHZ ram?

 

[Concillian]

Originally posted by: hurtstotalktoyou

As you can see, the performance benefit from a 33% CPU overclock was less than 10% in the real world. This might be due to the 11% underclocked RAM, or perhaps the hard disk bottleneck, but the end result remains rather disappointing. Hopefully a video benchmark might provide a wider gap.

If you plot the three out with CPU speed vs. time to encode is a straight linear relationship (Rsquare = 0.9998). Because CPU speed accounts for 99.98% of the time difference between the tests, it's unlikely that the RAM speed has much of an impact at all here.

You have to be careful with statistics and time measurements.. It's not just 10% in the real world, as it takes a CPU twice as fast to do something in half the time. That is, if you were to process something twice as fast (100% improvement in speed) your time would be HALF (229 seconds). The actual improvement is ~19% in the test you performed.

19% system performance for your particular test out of a 33% increase in a single component isn't too bad.

 

[Duvie]

I have done some testing here....

http://forums.anandtech.com/messageview...hreadid=1531414&enterthread=y&arctab=y

Granted ram was higher speed but lower timings...generally the besweet I have tested from years is mainly a raw cpu speed...therefore oy should have scaled linear...

 

[hurtstotalktoyou]

The video benchmarks are in.

4000 frames of DivX > MPEG2

1.60/800/200 -- 109 sec. (base)
1.92/720/200 -- 95 sec. (14.7%)
2.13/798/177 -- 89 sec. (22.5%)
2.28/855/190 -- 83 sec. (31.3%)
2.40/900/200 -- 79 sec. (38.0%)

Much better than the audio benchmarks, but 22.5% performance increase is still less than the 33% clock increase, or a 38% performance increase for a 50% clock boost.

 

[hurtstotalktoyou]

Originally posted by: Concillian

If you plot the three out with CPU speed vs. time to encode is a straight linear relationship (Rsquare = 0.9998). Because CPU speed accounts for 99.98% of the time difference between the tests, it's unlikely that the RAM speed has much of an impact at all here.

You have to be careful with statistics and time measurements.. It's not just 10% in the real world, as it takes a CPU twice as fast to do something in half the time. That is, if you were to process something twice as fast (100% improvement in speed) your time would be HALF (229 seconds). The actual improvement is ~19% in the test you performed.

19% system performance for your particular test out of a 33% increase in a single component isn't too bad.

I'm not sure where you're getting the 19% figure. It took 458 sec. to process x min. of audio at stock speed. That means it can process x/458 min. per sec. At 2.13 GHz, it can process x/414 min. per sec.

[(x/414) - (x/458)]/(x/458) = (458/414)-1 = ~.106 = 10.6%

 

[Concillian]

Again you aren't understanding that the scaling with time cannot be used as a linear function. Plotting a logarithmic function based on this data predicts you will be at half the time (100% faster) at a CPU speed of 3.37 GHz, which is only 5% off from a perfect 1:1 relationship. You are VERY close to perfect scaling with that Divx --> MPEG2 conversion, you just don't realize it because you aren't interpreting the data correctly.

 

[Concillian]

Originally posted by: hurtstotalktoyou

I'm not sure where you're getting the 19% figure. It took 458 sec. to process x min. of audio at stock speed. That means it can process x/458 min. per sec. At 2.13 GHz, it can process x/414 min. per sec.

[(x/414) - (x/458)]/(x/458) = 458/414 = ~.106 = 10.6%

So you're saying that if you double the processor's speed you expect it to take 0 seconds? Because that's the only way to get a 100% increase in performance the way you're showing it.

Time functions are not linear, they are logarithmic. 2x processor speed = 1/2 time. 100% improvement = 50% time.

458 / 2 = 229
458 - 414 / 229 ~= percentage benefit (with some error because you are making a linear approximation of a logarithmic function)

Another way to gauge how well it scales with CPU power with a logarithmic function is to plot the data in excel and fit a log function to it. Then you can extrapolate to half time and see how much CPU it would require with your fit. This gives you a 'scaling factor' with how close you are to 1:1 scaling.

If I do this for your two tests, I get:
-5.261ln(x) + 33.838
for the first one, which correlates to 5.25 GHz to cut the time in half compared to 3.2 Ghz, this is 1.64 factor, which is okay, but not great.

for the otehr test I get
-2.57ln(x) + 13.65
which corresponds to a 3.37 GHz processor to cut the time in half. This is a 1.05 scaling factor rcompared to a 3.2 GHz CPU, which is pretty good considering the best you can generally expect is 1.0

It's kinda technical, but my job is basically data analysis and statistics. Your expectations seem too high. This is why generally you see professional benchmarking programs stay away from time based benchmarks. They're confusing

 

[hurtstotalktoyou]

It's been a while since my last math course, but I'm pretty sure I'm right.

Double the processing power would indeed mean half time for encoding.

458/2 = 229
(458/229)-1 = 1.000 (which is 100%)

I left out a "- 1" in my equasion above, which I have just fixed

 

[Concillian]

Originally posted by: hurtstotalktoyou
It's been a while since my last math course, but I'm pretty sure I'm right.

Double the processing power would indeed mean half time for encoding.

458/2 = 229
(458/229)-1 = 1.000 (which is 100%)

I left out a "- 1" in my equasion above, which I have just fixed

Both are approximations, because they are assuming a linear relationship of a log function. Using my stats software from work on the log function, it looks like your linear approximation is low and mine is high, the number was actually about 14.5% (almost right in the middle )

The log function looks somewhat like a hyperbole if you remember what that looks like. Your linear approximation takes the tangent and my linear approximation is kind of cutting through the middle of the curve, so mine was optimistic and yours was pessimistic. It's hard to explain without showing the picutres of the functions. If you have excel or open office, plot the data with a scatter plot, do a curve fit with a log function and set it to 'forecast' to about the halfway point and you can see what I mean.

It's not a regular straight linear function because you should never cross zero on either axis... you will never have zero time unless you have infinite MHz, and zero MHz will yield infinite time.

Sorry I've managed to take things way off topic, I'll stop now. Sometimes I get too technical for my own good.

 

[Duvie]

Originally posted by: Concillian

Originally posted by: hurtstotalktoyou

I'm not sure where you're getting the 19% figure. It took 458 sec. to process x min. of audio at stock speed. That means it can process x/458 min. per sec. At 2.13 GHz, it can process x/414 min. per sec.

[(x/414) - (x/458)]/(x/458) = 458/414 = ~.106 = 10.6%

So you're saying that if you double the processor's speed you expect it to take 0 seconds? Because that's the only way to get a 100% increase in performance the way you're showing it.

Time functions are not linear, they are logarithmic. 2x processor speed = 1/2 time. 100% improvement = 50% time.

458 / 2 = 229
458 - 414 / 229 ~= percentage benefit (with some error because you are making a linear approximation of a logarithmic function)

That is a raher roundabout way to do it though it is right...
Look at my link....

I had a 1.8ghz cpu vs a 2.664ghz machine = this represents a 48% increase in mhz....or technically represents a value of 148% the original...also known as a ratio of 1.48 versus cpu A and cpu B

If you look at my times I had a

BESWEET GUI (wav to ac3)(sec)-----------------99----------------------68---------------+45.6% (1.45588)



A 100% increase would not take anything to zero....

a 100% increase in mhz in my case would have been a 1.8ghz versus a 3.6ghz...in reality is represents a 200% or 2x the original value but the differencial is what is 100%

A=1.8ghz
B=3.6ghz

1.8ghz/(A-B) = increased percentage differential

It can be seen both ways...

However in this case a 100% increase to my 1.8ghz versus the 3.6ghz would be to take the 99 and become 45sec....Now it would take a 7.2ghz chip theoretically to reach the 22.5 second range..It will never reach 0 theoretically....

A=cpu speed 1
B=cpu speed 2

B/A = ratio.....Time of A/ratio or Time of A X inverse of ratio.....

Example

A=1.8ghz
B=3.6ghz

Time of A=99sec
Time of B=x

3.6/1.8 = 2.00 (also known as 200% or true double).....99 sec / 2.00 = 45sec


His example

A=1.6ghz
B=2.13ghz

Time of A=458sec
Time of B=x

2.13/1.60 = 1.33.....458sec / 1.33 = 344 sec....That would be true linear...his 414 seconds represnets something far from linear...

the ratio of the 2 speeds shoudl directly correlate the ratio between the 2 times...

458/414 = 1.106 or 10.6% increase in performance for which time represents a decrease in time....


My example

A=1.8ghz
B=2.664ghz

Time of A=99sec
Time of B=x

2.664/1.8 = 1.48.....99sec / 1.48 = 66.89sec...I actually got a 68.....near linear

actually 99sec / Ratio = 68sec...... Ratio = 99sec/68sec or 1.455 ...close to original ratio...

 

[hurtstotalktoyou]

Yup. I thought so.

 

[Hacp]

So no higher overclock huh?

 

[hurtstotalktoyou]

Well, I had it going at 2.24 GHz for a few minutes, but I don't want to push my luck.

I've noticed some weird problems in AVIpreview that deal with "lost memory" or something like that. I'm concerned that my generic 3-4-4-8 RAM isn't entirely stable at asynchronous frequencies.

Are my fears legitimate, or am I just being overly cautious?

EDIT: I'm now at 2.28/855/190. Let's see if this can stay stable.

 

[hurtstotalktoyou]

Like the 32-bit 2600+ Xbit Labs test, my Sempron 64 2800+ reached 2.28 GHz (855/190) with no problem. Raising the voltage from 1.4v to 1.5v allowed me to take it to 2.4 GHz (900/200) on a retail cooler.

I'm amazed. I'm floored. I bought a CPU for $76, spent $20 extra on an overclocking board and $50 extra on a powerful PSU, and now I'm able to run at Athlon 64 3300+ speeds! This is much more than I ever expected!

 

[Duvie]

I think you should be figuring out why yuou have not seen more of a linear performance chart with your ocing...most things scale linear unless they are bottlenecked by vid card or IO limited...

Look at my review here...I had a 44+% OC and look at how many things scaled close to perfect...granted I had some higher memory speeds but at loosened timings cause I could run 2-2-2-6 at 400ddr...

http://forums.anandtech.com/messageview...hreadid=1531414&enterthread=y&arctab=y

 

[hurtstotalktoyou]

It might be my 5400 RPM PATA/100 HDD. My SATA/150 drive hasn't arrived in the mail, yet. Or it could be my crappy 3-4-4-8 RAM. I'm just not sure.

 

[SickBeast]

You should read Anand's review of the 3100+ Sempron. He said that even overclocked to 2400mhz (20%), in Doom 3, it only gave a 4% boost. From the sounds of it, the 128kb of L2 cache is the main culprit.

I'm actually seriously turned off of the Sempron line at this point. The 2800+ A64 isn't that expensive anyways.

 

[hurtstotalktoyou]

Originally posted by: SickBeast
You should read Anand's review of the 3100+ Sempron. He said that even overclocked to 2400mhz (20%), in Doom 3, it only gave a 4% boost. From the sounds of it, the 128kb of L2 cache is the main culprit.

I'm actually seriously turned off of the Sempron line at this point. The 2800+ A64 isn't that expensive anyways.

First of all, it was a Sempron or nothing. I simply didn't have the money for an A64, even the oldest, slowest 2800+ Hammer. Also, the 1.8 GHz A64 doesn't overclock well. The Sempron 2800+ @ 2.4 GHz is probably faster than the Athlon 64 2800+ @ 2.1 or even 2.2 GHz. At 2.4 GHz, my Sempron is basically working slightly faster (900 MHz HT) than a stock speed Athlon 64 3300+.

Anand's review of the Sempron 3100+ compares it to an Athlon 64 3000+, both working at 1.8 GHz. Although there are other differences between the CPUs, it's a good starting point to compare the effects of cache size. Anand's audio/video and multitasking tests (the only ones I'm concerned about) show a 0-14% performance gap, with most benchmarks showing less than a 10% difference. While that's certainly not wonderful news, it's on par with a 200 MHz clock boost.

In the end, my setup is running slightly faster than the speed of an Athlon 64 3300+. I'm very happy with that.

 

[monster64]

Originally posted by: SickBeast
You should read Anand's review of the 3100+ Sempron. He said that even overclocked to 2400mhz (20%), in Doom 3, it only gave a 4% boost. From the sounds of it, the 128kb of L2 cache is the main culprit.

I'm actually seriously turned off of the Sempron line at this point. The 2800+ A64 isn't that expensive anyways.

Are you joking? Thats like overclocking my 6800 GT to ultra speeds then wondering why it didn't lower my video encoding time.

 

[charloscarlies]

Originally posted by: hurtstotalktoyou
Also, the 1.8 GHz A64 doesn't overclock well.

Sorry to take this a little off topic, but this just isn't true. I've been through many A64 2800+ chips that overclock very very well. These were all Newcastles though...not the crippled Clawhammers.

 

[Acanthus]

My 1.8ghz in sig seems to do alright

 

[Furen]

Originally posted by: Acanthus
My 1.8ghz in sig seems to do alright

They're talking about 1.8GHz s754 Newcastles...

 

[Acanthus]

Originally posted by: Furen

Originally posted by: Acanthus
My 1.8ghz in sig seems to do alright

They're talking about 1.8GHz s754 Newcastles...

when building a new rig, why would you limit yourself to S754 when S939 boards have come down to similar prices (unless youre buying a POS)

 

[hurtstotalktoyou]

Originally posted by: Acanthus

Originally posted by: Furen

Originally posted by: Acanthus
My 1.8ghz in sig seems to do alright

They're talking about 1.8GHz s754 Newcastles...

when building a new rig, why would you limit yourself to S754 when S939 boards have come down to similar prices (unless youre buying a POS)

I got my board + retail CPU for just under $145. Sure, Venices are nicer, but they cost a *lot* more, especially considering the s939 motherboard premium. I only had $150 with which to work.