Faster than fast Athlon Mobile-Beats A64

[zxlr8]

I thiough that would get your attention....
Heres the story. I have the 2 systems overclocked in my sig-1 A64 clawhammer(1mb l2 cache cg revision) and the HTPC Server (Iqyha0351 athlon xp) I ran some sandra tests on the HTPC Server and according to sisoft arithemetic, My Htpc is a bit faster than my A64 overclocked. Here are the specefics:
Clawhammer@2400mghz= 10408 mips(240x10)prime not tested(don't care)
Athlon Xp@2500mghz= 10644 mips(218x11.5)24 hour prime stable

Now, i am really glad about the HTPC getting great results, but the best sisoft I ever got on my other cpu(Aqyha@2690mhz) was 10234 mips. I am wondering if this program is bugged. I have ran it a few times with the same results.
Now does this seem right? I want it to be, but it looks a little optimistic. I am getting around 19000 3dmarks at this level and if I overclock the video 20k....

 

[Gautam]

Sandra just calculates raw IPC, and disregards most other aspects of the architecture. It's why Prescotts and A64's do relatively poorly in it. The A64's pipeline is actually longer than the AXP's, so it's actually rather surprising that beats out the AXP. The only reason it manages to is the huge cache, and probably some other optimizations. But the A64's biggest punch comes from it's on-die memory controller, and extremely quick memory latency, which Sandra can't measure.

For the record, my highest Sandra is 10930, at 2520MHz with my A64. I got 10525 with my AXP T-Bred at 2.7GHz. Cache and raw clock speeds are really only what it's about in this test.

I'd love to see a 3Dmark with that 9800 and the a64. Hint: It'll be above 22k, no doubt, probably 23k.

 

[zxlr8]

Yes I am installing the x800xt in my laptop as we speak....j/k..I wish. I think that I am happy for a change and will quit messing with my HTPC. It is blazing for anything I run and the 3d performance is great. There is something just super sweet when you can get such great performance out of a $750 machine. What would my HTPC do on superpi? Oh and my best sandra is 10834....2645mghz

 

[zxlr8]

The machine I am building this weekend will be the 9800pro and the a64 like yours. I will let you know where it lands....

 

[kiyoshilionz]

What settings are you running that computer at? I got my computer to 240 x 11 = 2.6ghz @ 2V and it only got 10054 Not to mention it needed that whopping 2V....

What's your experience with that other AQYHA you have? They seem to be pretty crappy compared to the IQYHA's.

 

[zxlr8]

Just use the latest Sandra. I am running 218x11.5 all the time. I got 10587 just now which is the lowest so far. I am running @ 1.76 volts.

 

[Gautam]

Nice, good luck with that, you'll love it. I think you'd get somewhere around 37-38sec in SuperPI, but I'm not really sure. Why don't you try it?

 

[zxlr8]

Can anybody else give me some sisoft comparisons? There is something wierd going on....

 

[macklin01]

I've also noticed strange things in SiSoft, and not just with AMD chips. Back in the day, I was testing the regular PIII tualatins at about 1200 MHz vs the PIII-S at 1266. The only major difference between the two chips was the size of the L2 cache (256 KB vs. 512 KB). In the SiSoft tests, they scored identically when set to the same clockspeeds. In fact, in checking father, I noticed that even Celerons with just 128 KB of L2 could match it if they were clocked to the same speeds. Also, when I looked at the benches they had for the P4, I noticed that P4 celerons and regular P4's of the same speed scored about the same.

What I took out of this is that whatever tests SiSoft are doing, they can fit either entirely within the L1 cache, or within a 128 KB L2 cache. This also means that their CPU arithmetic tests are pretty simplistic. I'm not sure how much I trust those numbers as really testing the entire CPU design. I think I like how Gautam put it: the SiSoft arithmetic test only rates the IPC of the architecture. -- Paul

 

[zxlr8]

agreed...

 

[kiyoshilionz]

Muahah I just got 10,954 with the updated Sandra. Much better now, but if only I could keep my stupid processor cooler, I'm over 50 degrees now IDLE. It was much cooler yesterday so I could run everything a bit faster. This horrible stepping requires 2V for 2.6ghz I'm toying with the idea or pawning this off on eBay or classifieds and going for an IQYHA, but then I might not get my high FSB......

 

[zxlr8]

I will try tonight to run it at higher fsb. I should be able to catch you....

 

[kiyoshilionz]

Well I just got 10963 running it again after freeing up some RAM, but my computer just rebooted on me cuz I haven't given it enough voltage. At least you're stable....I'm just doing this to see how high I can go for now.

 

[Gautam]

FSB doesn't play into this, not one bit. Rather, see if you can get a higher clock speed by lowering the FSB.

The only major difference between the two chips was the size of the L2 cache (256 KB vs. 512 KB).

Hrmmm, that's interesting, because 512k Bartons do quite a bit better than 256k T-Breds, IIRC.

The only T-Bred I have with me is a 1700+ T-Bred A . It will not overclock. Can anyone post a result of a Barton at 11x133? I'll put what the 1700+ gets and we can take a look. I don't remember what version of Sanrda I ran that 2.7GHz bench with is the thing.

 

[zxlr8]

FSB definitely plays a role in the new sandra. 200x12.5=9900mips Don't ask me why...

 

[Gautam]

Weird. I'll try running at 280x9, see if that makes any difference.

Edit: No difference.

10978, I'd appreciate any increase.


EDIT: Bah!!! I was running an old version. 11,624.

 

[macklin01]

Ahh, I had forgotten about the versioning. I think I was also talking about the older Sandra. The newer one may well be different. -- Paul

 

[Gautam]

Even the older versions reflect the architectures. The Bartons just always do better than the T-Breds, rather noticeably. Maybe somehow the Tualatin's architecture deals with the lack of cache better?

 

[zxlr8]

Ok I just got 11030 mips. That is the best I have seen out of an athlon xp,,,,,

 

[macklin01]

Even the older versions reflect the architectures. The Bartons just always do better than the T-Breds, rather noticeably. Maybe somehow the Tualatin's architecture deals with the lack of cache better?

Hmm, good question. I'm not entirely sure. That's a very good possibility, however.

Also, perhaps there were some architectural changes between the Tbreds and Bartons other than cache? Was the L1 cache size changed? I'm afraid I'm not completely familiar with their details. (IIRC, they're both a .13u process, with a difference of 256 KB in their L2 sizes.) -- Paul

*edit*
I found this interesting link:

http://www.cybercpu.net/review/amd/barton/index.asp

Here's where I wrote a bit on Tualatin-S vs. Tualatin:

http://www.overclockers.com/articles684/

*/edit*