Speed Test (Problem: AMD Phenom 9750)

General questions about Indigo, the scene format, rendering etc...
User avatar
zsouthboy
Posts: 1395
Joined: Fri Oct 13, 2006 5:12 am

Post by zsouthboy » Fri Aug 29, 2008 9:11 am

Isigrim:

Actually, here's what I think is happening: when at full speed, the CPU is automatically being reduced in its clockspeed, due to power requirements. At lower clocks the power requirement is lower, so the scaling is normal. The higher clocks require more power - the first thread on the first core happens at normal speed, the second reduces the speed of both cores, and downward it goes all the way to 4.

So my guess:

Code: Select all

Cores Perf per core
1       1
2       .75 (downclocking to 75% speed)
3       .65
4       .60
Something like that.

Tys
Posts: 18
Joined: Sun Dec 10, 2006 4:53 pm

Post by Tys » Fri Aug 29, 2008 12:09 pm

Another possibility would be inadequate cooling of the cpu. Newer AMDs should be capable of thermal throttling, right?

Deus
Posts: 336
Joined: Sun Feb 04, 2007 3:47 am

Post by Deus » Fri Aug 29, 2008 2:26 pm

Yeah I had some SERIOUS issues with "spontaneous" downclocking of my old AMD cpu.

But what I meant was:

If 4 cores access frame buffer and BSP/Scene at the same time it requires x4 times the memory bandwith as 1 core... If scene is complex enough so that misses on caches on the different cores are frequent you will start to see the dreaded mem bandwith problem. To test if it is heat or bandwith related do some other benchmarks that are multithreaded with cores on/off and see if you get correct scaling. If you do get correct scaling it is a bandwith issue, if other benchmark doesnt scale linearly its a heat/clock problem.

On a related interesting side note: Some problems like L/R factorization of certain matrix sizes on for example a 4 core machine can actually perform x6 as fast as one core because the VIRTUAL cache size is larger (each core process a sub matrix and because sub matrix size = base size/ncores might fit in the cache it doesnt need to access main memory and when running same matrix on one core the submatrix is entire matrix i.e. *ncores larger than the sumbatrix). Its pretty obvious once you start thinking about it. This is sort of the inverse of the mem bandwith problem with many cores. Aslong as cache misses arent frequent and the memory bus can provide enough data to the cores caches it scales properly.
Last edited by Deus on Fri Aug 29, 2008 2:36 pm, edited 1 time in total.

Deus
Posts: 336
Joined: Sun Feb 04, 2007 3:47 am

Post by Deus » Fri Aug 29, 2008 2:32 pm

Let me add more on the heat issue.

Depending on manufacturer it clocks down either by FSB, Multiplier or other means. It can be VERY hard to detect. You need the correct program. I noticed it when gaming when suddenly my FPS dropped 30% after a bit of gaming. Typically try to run a benchmark on a "cold" machine and note down score.. Then you stress test i.e. run CPU/GPU intensive stuff for 30 mins. Then you go back and run the benchmark again. If you get the same score you dont have a heat problem. This is not completely fool proof because it might be that your AMD cpu/HSF combo has some problem (loose, dust, electrical) that it throttles immediately under load.

Isigrim
Posts: 33
Joined: Thu Nov 16, 2006 7:19 pm

Post by Isigrim » Fri Aug 29, 2008 5:01 pm

I can exclude the thermal throttling almost for sure. The cpu doesnt go above 59° Celsius.
But: it seems that the dropping samples/second suggest some kind of throttling. Or is it normal that rendering starts at 370 000 samples per second and drops to 200 000?
If it's in fact a bug of the Phenom cashes, Deneb should really bring some advantage here, because it has 6 MB L3 Cache which is associated 48-fold in contrast to the 32-fold association of 2 MB L3 in the Agena core.

Another thing that makes me think it's caused by the Mainboard is the fact that the mainboard was cheeping and pepping when i did the Test without Cool'n'Quiet. (Forgot to write that in the previous post).

Isigrim
Posts: 33
Joined: Thu Nov 16, 2006 7:19 pm

Post by Isigrim » Fri Aug 29, 2008 5:14 pm

thanks for the replies!!

I will post some more stuff once i have the new Mainboard. Then it will be clear if it's a mainboard or cache issue.
(Let's hope it's the mainboard)

alex22
Posts: 171
Joined: Thu Apr 12, 2007 12:07 pm
Location: Germany

Post by alex22 » Fri Aug 29, 2008 9:30 pm

Actually 4,5 times scaling is not impossible. If you render an image, there are different tasks done besides rendering, for example tonemapping and saving the image. If you assume this takes 33% of the rendertime on a single core, then a dualcore would be not twice, but 2.66 times as fast. This is because tonemapping only has to be done once, no matter how many cores you have. If you go on you will get 4.5 times speed increase for quadcores.

User avatar
fused
Developer
Posts: 3648
Joined: Fri Sep 22, 2006 7:19 am
Location: Berlin, Germany
3D Software: Cinema 4D

Post by fused » Fri Aug 29, 2008 9:44 pm

alex22 wrote:If you assume this takes 33% of the rendertime on a single core
woot 33%. if you set a reasonable update period, it shouldnt take more than 2-5%... (for example update period 15m, tonemapping takes 30s (which is quite extreme), thats 3.3% of the rendertime)

Isigrim
Posts: 33
Joined: Thu Nov 16, 2006 7:19 pm

Post by Isigrim » Wed Sep 03, 2008 12:04 am

Good news and information for people with Phenom processors.

I got my new Mainboard with AMD 790GX chipset (Asus M3A78-T) today and everything works fine now.

Code: Select all

With old Mainboard (Asus M3N78-EH)
4 Threads       5 m, 58 s        215136,98957 samples / second 

With new Mainboard (Asus M3A78-T)
4 Threads       3m, 26 s         362482,72795 samples / second
so my machine is more than 60% faster now, only because of the new mainboard.

So everyone out there who wants to buy a Phenom: Buy a good mainboard for it too. My suggestion is going for the new 790gx chipset, because it already has a sufficient onboard graphics core for render-machines.

Post Reply
24 posts

Who is online

Users browsing this forum: No registered users and 27 guests