Athlon XP faster than Athlon 64?

[CP5670]

I recently upgraded from an AXP configuration to an A64 one and although most things have gotten a big speed boost, a few programs have gotten significantly slower for some reason.

XP system: Athlon XP 3200 (200x11mhz, barton, 512k L2), Asus A7N8X, 1x Corsair VS 512mb and 2x Samsung generic 256mb (200mhz 3-3-3-8 1T)
64 system: Athlon 64 3200 (250x9mhz, some kind of 90nm, 512k L2), Asus A8V deluxe, 2x Corsair VS 512mb (208mhz 2.5-3-3-8 1T)
(everything else is either the same or doesn't affect performance)

Dosbox runs much slower than it used to. A lot of stuff that ran perfectly on the XP setup at 20000 cycles and no frameskipping now either needs under 10000 cycles or a 1x frameskip or does not work at all without heavy stuttering. This isn't a huge deal as I only start up these old dos games occasionally, but it's surprising to see such a large difference in favor of the XP. Mathematica also does a variety of numerical calculations a bit more slowly (as measured by its timing function). The differences are smaller here but still consistent across the board.

These results are quite strange considering that the 64 and its setup seems to be better in every way, including the clockspeed. The 64 is an engineering sample with an unusual label but I doubt it's all that different from retail versions since (windows) games and sandra benchmarks all work just as fast as they should. In the games there is no comparison; my framerates shot up pretty dramatically in most places even at video card intensive settings. By the way, CnQ is turned off and the windows settings (including background services) are almost identical between the two setups.

Anyone gotten similar results or can explain what is going on here?

 

[ribbon13]

Ever post on VOGONS?

Are you running a RC 64-bit XP?

 

[Visual]

I dont have time to lookup your board, A8V, but from the name I guess it's with via chipset. I don't know how bad that can lag behind nforce, maybe it's not a big difference, but its still a weird choice.

Why are you running 9x multi on your CPU? And also a divider for the ram?
250x9 is 2250, pretty low OC for a 90nm A64. To equal it at the max 10x multi, you'd need to reach just 225MHz HTT. At that frequency you can probably have your RAM 1:1 too, with the same latencies. (Hopefully the board you got has some decent vdimm options)

For an A64 3200+ I'd expect at least 240x10 to be easily reachable, and running your corsair modules at 240 should be posible too, even if you have to increase the latencies a bit.

Anyway, I realise that your system should be faster than the AthlonXP even at stock settings, so there must be another problem. The thing that comes to mind is to check if you're running in dual-channel - maybe you put the ram modules in the wrong slots and it's working single-channel? Your mobo manual should explain how the ram slots on your board are paired for dual-channel.

If Dosbox is the only app with such a weird performance drop, I'd assume its some issue with Dosbox itself. See if you're using the same versions, same settings, etc. Or get the newest version in case there was some A64-related bug that got fixed recently...

Also I'm curious about the engineering sample that you've got Where did you get it, and could it just be too old and crappy (especially if it can't OC above 2250MHz, i doesn't sound good)

 

[DAPUNISHER]

It is possible that ES chip is defective, more likely it is software specific though as Visual proposed.

 

[trickytraka]

might sound silly... but you do use a FRESH windows XP installation with the new system eh ?

 

[CP5670]

Ever post on VOGONS?

Are you running a RC 64-bit XP?

No, this is just 32-bit winXP SP1. I have been to that forum many times but never posted; maybe I'll ask about this over there.

I dont have time to lookup your board, A8V, but from the name I guess it's with via chipset. I don't know how bad that can lag behind nforce, maybe it's not a big difference, but its still a weird choice.

It uses that K8T800 pro chipset. It's a well known board though and several people around here have it. It seems to be competitive with and sometimes even a hair faster than the nforce3 boards, but it doesn't overclock as well and the VIA drivers can be finicky. I was going to get the MSI neo2 but changed my mind at the last minute due to all the bad experiences I saw about it.

Why are you running 9x multi on your CPU? And also a divider for the ram?
250x9 is 2250, pretty low OC for a 90nm A64. To equal it at the max 10x multi, you'd need to reach just 225MHz HTT. At that frequency you can probably have your RAM 1:1 too, with the same latencies. (Hopefully the board you got has some decent vdimm options)

For an A64 3200+ I'd expect at least 240x10 to be easily reachable, and running your corsair modules at 240 should be posible too, even if you have to increase the latencies a bit.

Anyway, I realise that your system should be faster than the AthlonXP even at stock settings, so there must be another problem. The thing that comes to mind is to check if you're running in dual-channel - maybe you put the ram modules in the wrong slots and it's working single-channel? Your mobo manual should explain how the ram slots on your board are paired for dual-channel.

2250 is about as high as I can get with this processor. I could use 225x10 but it only goes up to 219 or so at the standard timings and 2.7V (it's just value memory, not high end stuff). The tighter timing seems to do a bit more than bandwidth and dividers don't hurt performance anyway according to Zebo's charts. It is running in dual channel though. I get the dual channel message on startup and the sandra memory bandwidth is about where it should be.

If Dosbox is the only app with such a weird performance drop, I'd assume its some issue with Dosbox itself. See if you're using the same versions, same settings, etc. Or get the newest version in case there was some A64-related bug that got fixed recently...

It could well be some oddity with Dosbox (0.63, the latest one, was used in both cases). But why should Mathematica do the same thing? This is mostly FPU and cache dependent and the two processors are pretty similar there. If anything the A64 should have an edge.

Also I'm curious about the engineering sample that you've got Where did you get it, and could it just be too old and crappy (especially if it can't OC above 2250MHz, i doesn't sound good)

It definitely is old and crappy, a week 15 unit that starts to lose stability around 2300 (stock is the usual 2000). I got it for free from one of my dad's colleagues whose husband is an AMD scientist, so at that price I'm not complaining. Might switch to a venice later if those come out as the lower end A64 models.

might sound silly... but you do use a FRESH windows XP installation with the new system eh ?

yeah, it's a fresh install.

 

[11427]

I noticed the same type of thing with three MB's and two chips (a 754 A64 2800+, and a 939 A64 3000+) programs like Seti run much faster on the XP machines. Its not untill the A64's were OC'd to at least 2.2ghz that they surpassed the XP's.

 

[CP5670]

Maybe the 64 is actually weaker in FP performance or the cache is higher latency (Mathematica tends to be unusually dependent on cache) but the superior memory subsystem makes up for that in the majority of cases.

 

[ribbon13]

Originally posted by: CP5670
It definitely is old and crappy, a week 15 unit that starts to lose stability around 2300 (stock is the usual 2000). I got it for free from one of my dad's colleagues whose husband is an AMD scientist, so at that price I'm not complaining. Might switch to a venice later if those come out as the lower end A64 models.

Yep...