HyperTransport Link Testing

[NCspecV81]

Well I've read and read and read and read, and it seems that almost everyone agrees that keeping the HT Link close or at stock is the best performance. I've been testing this out and the PC seems different with HT being overclocked. I first noticed during LinX testing that GFlops would increase with HT overclocking... SO I'm just going to do a plethora of benches @ my 24/7 overclock of 4.65ghz @ 300bus and manipulate between 2100mhz HT and 3300mhz HT.

It will be an ongoing bench b/c I don't plan on doing all of these in one sitting...so these will be continued and I'll update the OP. Please post some suggestions for benches 3D or 2D.

Also I do have some results so far but I didn't grab screenshots for them b/c I didn't expect to be doing a thread so I have 3dmark06 and 3dmark Vantage with just orbs. I will plan to capture screens from here on out.



Versus




Rest of the test setup -

CPU - Phenom II 965 C3
MoBo - Gigabyte 790FXTA-UD5
RAM - Corsair Dominators GT @ 1600mhz 6-6-6-15-11 1T
GPU - Trifire HD5870's
NB Freq - 3300Mhz
OS - Win7 32bit

[NCspecV81]

3DMark Vantage Results (sorry for the screencapless posts for now)

HT@2100 w/ ALL GPU's @ 1050/1300 - P27,278
http://service.futuremark.com/compare?3dmv=1772928

HT@3300 w/ ALL GPU's @ 1050/1300 - P27,577
http://service.futuremark.com/compare?3dmv=1774184


3DMark06 Results

HT@2100 with GPU's Stock - 27,626
http://service.futuremark.com/compare?3dm06=12918855

HT@3300 with GPU's Stock - 27,971
http://service.futuremark.com/compare?3dm06=12918894


3DMark05 Results

HT@2100 with GPU's Stock - 34,626



HT@3300 with GPU's Stock - 34,809




3DMark03 Results

HT@2100 GPU's Stock - 135,456



HT@3300 GPU's stock - 135,701



I'm kinda confused about this one. HT Link @ 2100Mhz was faster in 3 of the 4 tests by a good margin, but ended up losing the final results. I guess the GT2 win for 3300Mhz was very influential!



Furmark Results -

HT@2100 with GPU's Stock - Score 23,186



HT@3300 with GPU's Stock - Score 23,454



Everything is higher here with the min, average, and max. Just more graphical throughput with the faster HT Link.

[NCspecV81]

reserved

[Titan7171]

its basically a 300 point increase in performance for 1200mhz increase........not really worth it especially since most people have trouble just running 2800mhz on the HT link

This thread over on extreme overclocking forums has some similar discussion and tests http://forums.extremeoverclocking.co...d.php?t=329541 its from October

[gOtVoltage]

The only thing ive every noticed with High HT Above 2500+(5200HT)

The loading of windows and Desktop seemed faster Other than that i nerver knew HT could go over 6000mhz until now lol. Never gave it much attention..

Interesting though . it seems to scale with NB when kept at same speed.

[NCspecV81]

This thread over on extreme overclocking forums has some similar discussion and tests http://forums.extremeoverclocking.co...d.php?t=329541 its from October

not to be a douche bag but that is nothing like what I'm doing at all. The only variable changing between the two tests I'm performing is the HT Link speeds. The variable between tests compared in the thread you linked is bus speed with no mention of HT Link.

But 300 pts by just changing a multiplier in the bios or in windows? How easy can 300 pts get? I'd say it's well worth the 2 seconds it takes.

[NCspecV81]

Okay - 03 is like a gold mine of points once you hike up the clocks and keep up the HT Link! I ran these tests TWICE for each. 136xxx for 2100 twice in a row.. 139,416 and 139,6xx for 3300 twice in a row... Hiked the cpu clock to 4800mhz and upped the gpu's to 1035/1300

2100Mhz -



3300Mhz - GT2 literally goes off the scale compared to 2100Mhz




I'm going to re-run these one more to confirm these. Quite a significant variance once uncorking it some.

[Zeus]

Interesting results, especially for 03. Keep 'm coming.

[Lightman]

Very interesting findings!

I was expecting that 3DM03 will show biggest gains due to being the best scaling benchmark for multi-GPU.

Either lower communication latency between mobo and GPU is helping your results or pure bandwidth helps. If your mobo could change HT link width to 8-bit we could pinpoint which one.

Anyway nice to look at such a scores and improvements from just HT!

[NCspecV81]

reran 03 with the changing in HT Links!

rerun @ 2100 -



rerun @ 3300 -




I'm not sure how much more concrete you can get than that. I ran the 2100mhz one twice that time and posted the best score. 4k difference for 2 seconds of work? Ummm Hell yeah!

[NCspecV81]

06 reran @ 4800 with 1035/1300 on gpu's

2100Mhz HT -



3300Mhz HT - I just feel like these are free points! I haven't noticed any difference in stability what so ever.

[Daveburt714]

Things have always seemed to run smoother for me when I've kept
the HTLink and IMC linked (or HTLink down 1 click). Of course I never took the initiative to test it....

Very interesting results!

What are you running your IMC @, 3300?
I was under the impression HTLink couldn't be faster than the IMC.

If thats correct (and I'm pretty sure it is), theres NO WAY I could get my
HTLink that high anyway.

[NCspecV81]

Things have always seemed to run smoother for me when I've kept
the HTLink and IMC linked (or HTLink down 1 click). Of course I never took the initiative to test it....

Very interesting results!

What are you running your IMC @, 3300?
I was under the impression HTLink couldn't be faster than the IMC.

If thats correct (and I'm pretty sure it is), theres NO WAY I could get my
HTLink that high anyway.

Yes the IMC is @ 3300mhz for both 2100 HT and 3300 HT.

[[XC] Oj101]

The only thing ive every noticed with High HT Above 2500+(5200HT)

The loading of windows and Desktop seemed faster Other than that i nerver knew HT could go over 6000mhz until now lol. Never gave it much attention..

Interesting though . it seems to scale with NB when kept at same speed.

Could you please elaborate on that?

[gOtVoltage]

Its a wierd one . If i put my HT to say 2600 while keeping every thing else the same it seems that when my windows /vista loads and when ever anything is clicked on or moved on the desktop its faster.


Kinda like how you set the Mouse speed faster.
In a way id say :Raising the HT may of made it easier for the mouse and or i/o's to communicate ,making them appear faster ..

You would Notice this more and more the Higher the CPU clock goes up as well.

You could even notice it more on older AMD K8 systems.. 754/939 it was very noticable.

[[XC] Oj101]

I'm GUESSING that would happen only on a USB mouse. Could you please time your startup?

[charged3800z24]

Am I the only one that noticed the furmark run and the minimum FPS? Quite an improvement there....

[afireinside]

Could you test 3d03 real quick with one card? Really interesting. Last test I saw showed it eventually got slower with higher HT.

[massman]

This makes more sense than the reports of people losing FPS when increasing the HT Link, quite frankly.

Either lower communication latency between mobo and GPU is helping your results or pure bandwidth helps. If your mobo could change HT link width to 8-bit we could pinpoint which one.

Good thinking, I'd like to see this put to test.

[Raiderman]

Great findings. Ive always wondered about HT overclocking.

I take it you are on Phase to get those kind of clocks?

[demonkevy666]

could you see how high the pcie speed will go too and if help any when high HTT is used.

[Titan7171]

not to be a douche bag but that is nothing like what I'm doing at all. The only variable changing between the two tests I'm performing is the HT Link speeds. The variable between tests compared in the thread you linked is bus speed with no mention of HT Link.

But 300 pts by just changing a multiplier in the bios or in windows? How easy can 300 pts get? I'd say it's well worth the 2 seconds it takes.

Its no problem, Yea you are right, that was a few months ago and I didnt read through the thread when I posted the link.

Well I guess if its possible to run the ht link speed at 3300mhz without it being a burden to overall stability then what the hell ......The question is though how many people here can run at those speeds. Now I think many of us will give it a shot and see how high we can go. I still think 300 points for 1200mhz is very weak but fck it every point counts when you are running benches

Good job on your tests and thanks for your time

[BeepBeep2]

This is really interesting...but I don't know, it is safe 24/7 to run 5200+ on the HT Link? Alot of people tend to say it's "bad" and can cause weird stuff.

[NCspecV81]

BeepBeep I haven't noticed any difference between 2100 and 3300 other than a perkier system.

Single 03 by request

CPU 4800mhz GPU 1035/1300

2100Mhz -



3300Mhz -

[Don_Dan]

Thanks for testing this, seems like Tony's measurements were off.