HyperTransport Link Testing

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.

http://valid.canardpc.com/cache/screenshot/939145.png

Versus

http://valid.canardpc.com/cache/screenshot/939165.png


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

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

http://www.overclock.net/gallery/dat...rk05ht2100.PNG

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

http://www.overclock.net/gallery/dat...rk05ht3300.PNG


3DMark03 Results

HT@2100 GPU's Stock - 135,456

http://www.overclock.net/gallery/dat...rk03HT2100.PNG

HT@3300 GPU's stock - 135,701

http://www.overclock.net/gallery/dat...rk03HT3300.PNG

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

http://www.overclock.net/gallery/dat...markht2100.PNG

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

http://www.overclock.net/gallery/dat...markht3300.PNG

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

reserved

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

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

The loading of windows and Desktop seemed faster:shrug: 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.

Quote:

---

Originally Posted by Titan7171

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.

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 -

http://www.overclock.net/gallery/dat...003dmark03.PNG

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

http://www.overclock.net/gallery/dat...003dmark03.PNG


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

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

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!

reran 03 with the changing in HT Links!

rerun @ 2100 -

http://www.overclock.net/gallery/dat...ark03rerun.PNG

rerun @ 3300 -

http://www.overclock.net/gallery/dat...t3300rerun.PNG


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!

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

2100Mhz HT -

http://www.overclock.net/gallery/dat...4800ht2100.PNG

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

http://www.overclock.net/gallery/dat...4800ht3300.PNG

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.... :p:

Very interesting results! :up:

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.

Quote:

---

Originally Posted by 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.... :p:

Very interesting results! :up:

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.

Quote:

---

Originally Posted by gOtVoltage

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

The loading of windows and Desktop seemed faster:shrug: 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?

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.

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

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

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

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

Quote:

---

Originally Posted by Lightman

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.

Great findings. Ive always wondered about HT overclocking.

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

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

Quote:

---

Originally Posted by NCspecV81

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:up:

Good job on your tests and thanks for your time:yepp:

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.

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 -

http://www.overclock.net/gallery/dat...le03ht2100.PNG

3300Mhz -

http://www.overclock.net/gallery/dat...le03ht3300.PNG

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

Hey guys, I'm not gonna post any screenshots but I just ran some quick tests in a real game called "TrackMania Nations"...it's got a built-in benchmark in the options.

1680x1050, settings maxed out, Forced Motion Blur, Forced HDR, Dynamic Shadows, 8xMSAA + 8xAF. Here's a resized screenshot of the benchmark: http://img710.imageshack.us/img710/9...onsforever.png

Phenom II 965BE C3 @ 4050 Mhz, NB 2.7 Ghz (18x225)
4GB DDR2-1200 5-5-5-15
Gigabyte GA-MA790X-UD4P
ASUS "TOP" ATI HD4850 @ 800/1125 (2250 effective GDDR3)
Two shortstroked WD 640GB Caviar Black's in a 320GB RAID 0 Array
Windows Vista Home Premium 32-bit

Total avg FPS:
52.0 FPS - 2025 Mhz HT Link (4050 Mhz effective)
52.1 FPS - 2250 Mhz HT Link (4500 Mhz effective)
52.3 FPS - 2475 Mhz HT Link (4950 Mhz effective)
54.1 FPS - 2700 Mhz HT Link (5400 Mhz effective)

I don't know whats up with the last result but I ran the same test 4 times, and re-ran the last one 3 times.

BeepBeep2 - what is your NB frequency? It could have something to do with it being synced with the HT. So in essence, it may not be the case of a higher HT link but it being in sync.

I am going to look at Min FPS on benches. That is what I care most about.

2.7 Ghz... I'm sort of wondering about this.

I saw (really) minor gains as I upped the HT Link but when the two were snyced I gained a whole frame.

Nice Spec!!!

Numbers look good.... maybe I should Run this again and try and tighten up those timings...

http://img695.imageshack.us/img695/6103/53801341.jpg

I did a little test with Crysis benchmark. Below you will see the detailed rsults. I gained 1FPS in Min and 3FPS on Average. Not exciting, but this is Crysis and it wasn't hard to up the HT link.:p:

64bit
HD 5850 Single stock
965BE 3.4ghz / 2800NB
HT2000mhz vs 2600mhz

2000 HT

Quote:

---

1/10/2010 4:25:35 PM - Vista 64
Beginning Run #1 on Map-island, Demo-benchmark_gpu
DX10 1400x960, AA=No AA, Vsync=Disabled, 32 bit test, FullScreen
Demo Loops=3, Time Of Day= 9
Global Game Quality: Custom
Custom Quality Values:
VolumetricEffects=High
Texture=High
ObjectDetail=VeryHigh
Sound=High
Shadows=High
Water=High
Physics=High
Particles=High
Shading=High
PostProcessing=High
GameEffects=High
================================================== ============
TimeDemo Play Started , (Total Frames: 2000, Recorded Time: 111.86s)
!TimeDemo Run 0 Finished.
Play Time: 30.51s, Average FPS: 65.55
Min FPS: 43.91 at frame 182, Max FPS: 95.98 at frame 1606
Average Tri/Sec: -67619304, Tri/Frame: -1031505
Recorded/Played Tris ratio: -0.89
!TimeDemo Run 1 Finished.
Play Time: 26.31s, Average FPS: 76.03
Min FPS: 43.91 at frame 182, Max FPS: 99.74 at frame 1760
Average Tri/Sec: -77447576, Tri/Frame: -1018687
Recorded/Played Tris ratio: -0.90
!TimeDemo Run 2 Finished.
Play Time: 26.26s, Average FPS: 76.17
Min FPS: 43.91 at frame 182, Max FPS: 102.71 at frame 131
Average Tri/Sec: -77495736, Tri/Frame: -1017396
Recorded/Played Tris ratio: -0.90
TimeDemo Play Ended, (3 Runs Performed)
================================================== ============

Completed All Tests

<><><><><><><><><><><><><>>--SUMMARY--<<><><><><><><><><><><><><>

1/10/2010 4:25:35 PM - Vista 64

Run #1- DX10 1400x960 AA=No AA, 32 bit test, Quality: Custom ~~ Overall Average FPS: 76.1

---

2600ht

Quote:

---

1/10/2010 4:28:27 PM - Vista 64
Beginning Run #1 on Map-island, Demo-benchmark_gpu
DX10 1400x960, AA=No AA, Vsync=Disabled, 32 bit test, FullScreen
Demo Loops=3, Time Of Day= 9
Global Game Quality: Custom
Custom Quality Values:
VolumetricEffects=High
Texture=High
ObjectDetail=VeryHigh
Sound=High
Shadows=High
Water=High
Physics=High
Particles=High
Shading=High
PostProcessing=High
GameEffects=High
================================================== ============
TimeDemo Play Started , (Total Frames: 2000, Recorded Time: 111.86s)
!TimeDemo Run 0 Finished.
Play Time: 30.42s, Average FPS: 65.74
Min FPS: 44.99 at frame 183, Max FPS: 97.47 at frame 1621
Average Tri/Sec: -67774880, Tri/Frame: -1030965
Recorded/Played Tris ratio: -0.89
!TimeDemo Run 1 Finished.
Play Time: 25.98s, Average FPS: 76.99
Min FPS: 44.99 at frame 183, Max FPS: 104.86 at frame 1790
Average Tri/Sec: -80075168, Tri/Frame: -1040085
Recorded/Played Tris ratio: -0.88
!TimeDemo Run 2 Finished.
Play Time: 24.61s, Average FPS: 81.25
Min FPS: 44.99 at frame 183, Max FPS: 112.43 at frame 98
Average Tri/Sec: 82464824, Tri/Frame: 1014932
Recorded/Played Tris ratio: 0.90
TimeDemo Play Ended, (3 Runs Performed)
================================================== ============

Completed All Tests

<><><><><><><><><><><><><>>--SUMMARY--<<><><><><><><><><><><><><>

1/10/2010 4:28:27 PM - Vista 64

Run #1- DX10 1400x960 AA=No AA, 32 bit test, Quality: Custom ~~ Overall Average FPS: 79.12

---

tried 2800 ht link? so it's 1:1 with the NB.

I didn't want to reboot to get 2800 :p:. AOD would only go to 2600HT.:rolleyes:

Man, I either has a sucky IMC or my water just isn't good enough...

Anything over 2650 on the IMC requires massive V's >1.35v and scaling above that is pretty much worthless. :(

I can push 2700 but it takes well over 1.4v, anything above that seems to be a lost cause....
There comes a point (at least in my book) where the voltage just doesn't justify the performance/heat involved... :rolleyes:

Thanks for the numbers NCSpec !
I've always kept it matched or at least close to the IMC...
At least now I'll think twice before I blame the HTLink for OC problems. :up:

Only 3dmark I have installed/downloaded is 06 so I tried that. Got ~60 points out of 524mhz more HT link. Wasn't really expecting anything considering I'm only on a single 4850 though. Now to see if this HT link is stable. I recall having issues with higher HT link in the past but I may be wrong.

I'd like to see some more results from you guys. Make sure it's not some anomaly with my setup.

Quote:

---

Originally Posted by Daveburt714

Man, I either has a sucky IMC or my water just isn't good enough...

Anything over 2650 on the IMC requires massive V's >1.35v and scaling above that is pretty much worthless. :(

I can push 2700 but it takes well over 1.4v, anything above that seems to be a lost cause....
There comes a point (at least in my book) where the voltage just doesn't justify the performance/heat involved... :rolleyes:

Thanks for the numbers NCSpec !
I've always kept it matched or at least close to the IMC...
At least now I'll think twice before I blame the HTLink for OC problems. :up:

---

I remember you saying you seen nice things keeping them in sync (NB/HT). I ran test a whil back on one of my Phenoms, I don't think I posted them. The gains were about what I saw to day. I think I migh do some experimenting this week with some other games. I think there might be a spot for HT with different CPU clocks.
On the NB, I had 1.35v on that for 2800, I only ran 4 instances of Memtest for an hour to test it. Not sure if it would pass Prime95. 3000mhz needs around 1.5+ just to do a bench.

3Dmark 2003
single HD4850 @ 675/1050
PCIE 110mhz

GT3 seems to steadily increase with HT link on my setup. GT2 didn't skyrocket like yours nor did GT4 decrease. I'll try again in a few days when my 5850 arrives but I think the trifire you're running is part of the reason your gains are so massive. Still, there is performance to be had (both increased and decreased it seems...) with higher HT link

http://img.photobucket.com/albums/v4...m2/1834HTL.jpg

http://img.photobucket.com/albums/v4...m2/2358HTL.jpg

http://img.photobucket.com/albums/v4...m2/2620HTL.jpg

Tested near stock, linked, and one multiplier below linked to see if I had odd results like beepbeep. That didn't happen though. I don't think I can take my NB any higher :(

Quote:

---

Originally Posted by charged3800z24

I didn't want to reboot to get 2800 :p:. AOD would only go to 2600HT.:rolleyes:

---

In order to get 1:1 with the northbridge you have to set the NB multi to 12, and use the bus speed to reach 2700+ on the NB clock. HT multi is locked at 12.
I am sure you already knew this, just thought I would post for other peoples benefit.

It's board dependant.

Not all boards gain from HT, also we have seen what appears to be a higher max fps but a lower min.

http://www.cpuid.com/perfmonitor.php


There is an option to monitor HT bandwidth.It doesn't get saturated with a single cpu,there will be a speed up with 3-4
cards plugged in.Thanks for testing,nice work :up:

I mean 13x multi, my math was off :rolleyes:

Quote:

---

Originally Posted by Daveburt714

Man, I either has a sucky IMC or my water just isn't good enough...

Anything over 2650 on the IMC requires massive V's >1.35v and scaling above that is pretty much worthless. :(

I can push 2700 but it takes well over 1.4v, anything above that seems to be a lost cause....
There comes a point (at least in my book) where the voltage just doesn't justify the performance/heat involved... :rolleyes:

Thanks for the numbers NCSpec !
I've always kept it matched or at least close to the IMC...
At least now I'll think twice before I blame the HTLink for OC problems. :up:

---

I'm gonna guess sucky IMC...my air cooling isn't strong enough to run 4 Ghz 24/7, yet I can bench 3DMark06 @ 4.1 Ghz with a 2.8 Ghz NB.

Now, would you like to try this with a 2D bench? :)

Also guys/gals..when your testing what works and dosnt for HT..

Make sure the PCI-E is Locked @ 100;)

Simply keep it it locked not on AUTO(set it manualy)... Some people use 100 or 101 and even up to 110...I always use 100(stability)


This will affect the Bandwidth of the Videocard/s if you do not keep this value the same for every run you do while changing HT.

Its something thats overlooked when comparing HT freqencies besides using the Same Ram and CPU and NB settings.

Not to mention I have seen upwards of a 500 point diff between bench runs at same settings in the same session just by a simple reboot.

200 points is well within bench variance. It's only the diff between a "bad" and "good" run.

The first run after a reboot is always better than those run afterwards, keep this in mind when using 03 as a proof for your theory.

Can you please rerun your game benchmarks at the lowest settings possible?

no aa, low low low, 1024x768 :)

This will be a sign that it is an improvement for the CPU-GPU-memory transfers if it improves significantly.

Quote:

---

Originally Posted by Don_Dan

The first run after a reboot is always better than those run afterwards, keep this in mind when using 03 as a proof for your theory.

---

I ran them half a dozen times on the same boot for each. I'd say that factors that out. I'm also aware of this phenomenon.

TBH i have just kept my HT in sync with my NB speed through my overclocking of this chip and ram, i did try dropping the HT 2 see if it made a stability difference and there was none to b seen (at least for me) so for at least 95% of the time my NB is in sync with HT speed or vice versa depending on how ever you want to look at it

Dragon

The other phenomenon is some cpu's at max clocks can't hack high HT ;)

Quote:

---

Originally Posted by Raiderman

I mean 13x multi, my math was off :rolleyes:

---

I knew what ya ment;), But ya, I would have to go into the bios and change my mem divider to use the HT ref OCing, and I was too lazy to do so:p: