Fixed (TLB was the issue): Phenom 9600 w/ TLB Disabled slow in image rotate (90deg)

[dman]

So I upgraded my PC a few mo's ago from an Opteron 165 to a Phenom 9600 (New MB, all else equal). I had the Opteron board and CPU on standby and just upgraded my wifes pc with those old parts today.

During that time I noticed that when browsing my (SLR) photo's that it was taking 7-8s for some to display (I thought it was random at first). The delay was the same whether I used irfanview, XNView, and one other free image viewer. I finally found out that the delay was when the programs were using the exif info to auto-rotate the images from my camera. Once i disabled that feature the pics browsing was normal again. I read a while back that Memory performance is impacted with TLB and programs like WinRAR show rather large slowdowns. So, I tried the same tests with AMD Overdrive installed and the patch disabled and the performance was the same = slow. Up until today I thought that I just had never used the auto-rotate feature... but today I setup the old board with the same software (Vista 64bit, etc) and it rotates the pics with no noticeable delay.

So, is the Phenom really supposed to that much slower than an Opteron in these kinds of tasks? Is it perhaps a setting in the MB?

Original System
MSI Neo4 w/ Modded Bios
Opteron 165 (dual core) on Nforce4 @ ~1.83GHz w/ 4GB DDR
Vista 64bit (Ultimate)

New System
Asus M3A
Phenom 9600 @ Default settings [2.3Ghz] (and tested w/ TLB Patch disabled --red dot to green -- via AMD Overclock)
8GB DDR2
Vista 64bit (Ultimate)

At the time I had an ATI x1900 in the Opty, now it's in the Phenom and the Opty is using a Gforce 8600GTS.

I may try removing some RAM to see if perhaps it has something to do with 8GB vs 4GB.

If anyone feels like testing the performance of rotating 10MP images in irfanview and letting me know if you have a similar 5-10s delay.

[update 01/08/09]
So, it turns out that TLB was not disabled as I had been led to believe. I have Vista SP1 which I already know enables the TLB Fix (overriding the bios), but I had previously read that disabling it via AMD Overdrive utility would work. It turns out that is not true. I found and installed a utility from another site, which did disable the TLB Fix and speeds are noticeably improved, image rotation is essentially instantaneous again.

Unfortunately since I use sleep mode, every time the computer wakes up I'd have to run this utility again. On the positive side, I don't do that much image rotation and I haven't noticed the slow down in other areas so I can live with it for now. At least I have found the cause for the problem, and it's the same 'ol TLB. I apologize for rehashing a very old issue. [/Closed]

[Update 2]
I called AMD to see if they had a TLB patch that worked for Vista, but the Tech referred me to the Bios option as the only 'supported' method. He didn't think that the issue above was TLB related, if anyone else wants to test it out, if you have a 9x00 series Phenom (with Vista SP1) open up some images (10MP+/-) in irfanview or xnview or similar program and do a 'rotate' "R" key in irfanview. If it takes more than a few seconds to rotate, you can then search/install the TLB disable program and test again and the difference should be very noticible. I'd be interested in at least one other person confirming this, just to know that it's not a 'cache' problem with my cpu specifically. [/edit]

Tool Link: Xtreme Systems Forum

Requires CrystalCPUID also linked at same forum. For Vista 64bit I downloaded BOTH crystal CPUID's and extracted to the same folder, and then installed the files in 1.05 into that folder as well. I then ran the TLB disable program (in the folder) and tested image rotate and it worked much better (normal).

[Update 3]
02/10/08 - I've had a 9850 in the same system now and everything runs as expected. No isssues at all with the image rotation / display. Once again I can confirm that the performance hit was TLB related.

 

[SickBeast]

There are situations where a Phenom will lose to an A64.

The case you describe sounds like it would only exploit one core at a time, therefore it is quite possible that your Opteron is just better at that particular task.

 

[Flipped Gazelle]

It's true that Phenom's can loose to higher-clocked X2's, but given your Phenom's Mhz advantage over your Opteron, this shouldn't be the case.

I just opened a 33mb tif in Irfanview, rotated 90 deg, and it was almost instantaneous. Using Phenom 9850 @ 2.75 ghz, 4 gb RAM, Vista x64.

EDIT: I just resampled the image by a factor of 2 in both directions, and rotated 90 deg again. Almost instantaneous.

 

[soccerballtux]

Originally posted by: SickBeast
There are situations where a Phenom will lose to an A64.

The case you describe sounds like it would only exploit one core at a time, therefore it is quite possible that your Opteron is just better at that particular task.

Rotating a picture is very parallelizable. I'd be surprised if the code hasn't been written to use infinity cores for this.

 

[Idontcare]

Originally posted by: Flipped Gazelle
It's true that Phenom's can loose to higher-clocked X2's, but given your Phenom's Mhz advantage over your Opteron, this shouldn't be the case.

I just opened a 33mb tif in Irfanview, rotated 90 deg, and it was almost instantaneous. Using Phenom 9850 @ 2.75 ghz, 4 gb RAM, Vista x64.

EDIT: I just resampled the image by a factor of 2 in both directions, and rotated 90 deg again. Almost instantaneous.

I don't have irfan installed here (been ages since I used it) but it stands to reason to expect this task to be ridiculously fast on modern processors with their GB/s bandwidths and worst case 100ns access penalties for L3$ misses.

Something is dreadfully awry with the OP's system. It almost sounds like that SSD issue Anand debugged on those MLC non-Intel SSD's where the whole system would hang just because some random write somewhere (in the SSD controller in this case) caused some massive system stalls.

(seconds is literally billions upon billions of nanoseconds, to lock up a CPU operating in nanoseconds for billions of clock cycles requires some bad mojo in the system).

The first thing to do to help this guy is have someone attempt to replicate the problem on their 9600 system using irfanview and disable the TLB patch as he did.

 

[dman]

Most items are fast, the image rotation is just one area that was slow, and literally it'll take 5-8 seconds to rotate one image, the vista 'wait' mouse pointer spins 5-10 times. It happens outside of irfanview, had the same results in XNview and fsviewer.

I disabled Antivirus and my install was clean, but I can do some more testing. As was said above, was just looking to see if anyone else has the issue.

Also, the opteron 165 is dual core, so, going from 2-4 core and faster per-core it shouldn't be slower.

 

[dman]

Bump, it was TLB issue, apologize for bringing up old issues. More detail in OP.

 

[SickBeast]

Originally posted by: dman
Bump, it was TLB issue, apologize for bringing up old issues. More detail in OP.

I just read at the Tech Report that it affects even the Phenom II, so don't be so sure that it's the TLB.

Here is a link to a Photoshop benchmark which shows an A64 6400+ beating out even the newer Phenoms and the Phenom II

 

[dman]

For all of my regular computing needs the 9600 is/was/has been fine. I only noticed a major slow down during the image rotation process. That was painfully slow / annoying and seemed way out of place. That performance problem was completely resolved when I finally, successfully, disabled the TLB patch. Thus, I know for certain that my issue was fixed by the TLB tool above (but whether or not the tool does anything else besides disable TLB--well--that I can't confirm). In any event, I've got a 9850BE on the way now to replace the 9600BE so I won't have to worry about running the tool.

The link to the worldbench results is still interesting. I'm going to guess that it's a combination of intel code optimizations (PS generally favoring Intel CPU's) and clock speed (whereas the x2 6400 runs at 3.2GHz, and the other AMD cpu's are running at slower core speeds). I'm also going to further guess (based on history of these kinds of things) that the benchmark / utility / sotware doesn't properly run on Phenom 920 Core and uses alternate / legacy code when running on these CPU's. I doubt there'll be a patch for CS2, but it might be interesting to see what happens with CS3 / CS4 (if they run newer code).

Well, there could be many other explanations, and the bottom line is if those are the tasks you run often go with the intel family if you don't want to deal with the delays. Since I'm not running those tasks (very often) it's not an issue for me, but could be very important to others and good to know.