A single HyperTransport Link Frequency Benchmark

[Jovec]

Playing around with a new system and decided to bench various HT Link speeds. I might update with another benchmark, probably Dawn of War 2 for it's built in benchmark tool. Obviously one can't conclude too much from a single benchmark, but here it is for what it's worth.

Hardware: 1090T @3.6GHz, CPU-NB @2.6GHz, 8GB DDR3 @1600, HT Ref clock @200, 5850 @stock
Software: Win7 64, Cat 10.6, CineBench 11.529 64-bit OpenGL benchmark

5 run average
-----
1000 HT: will not boot into Windows
1200 HT: 65.25
1600 HT: 65.24
1800 HT: 65.03
2000 HT: 63.92
2200 HT: 65.02
2400 HT: 65.04
2600 HT: 64.16


Dawn of War 2 @1680x1050 built-in benchmark 3 run average (min/avg/max)
-----

             Low Settings           Med Settings           High Settings         Ultra Settings
             ------------           ------------           -------------         --------------
1200 HT: 49.5 / 98.1 / 182.7    33.1*/ 72.2 / 145.0    18.0 / 67.3 / 139.0     4.7 / 59.0 / 124.0
2000 HT: 48.2 / 94.0 / 173.0    36.1 / 72.5 / 145.0    18.7 / 67.5 / 139.7    13.7 / 61.4 / 129.7
2600 HT:                                                                      16.0 / 61.7 / 130.7

*There was an outlier result for 1200HT/Med/min.


SuperPI 32M
3 run average
-----
1200 HT: 19m 11s
2000 HT: 19m 11s


7zip - 1.14GB (10,415 files) into a 364MB 7z file
4 runs, discarded first run , average of remaining 3, manually measured
-----
1200 HT: 3m 23s
2000 HT: 3m 22s


8.85GB network copy
3 run average, manually measured
-----
1200 HT: 2m 40.7s
2000 HT: 2m 41.7s

 

[Idontcare]

Is there any reason to suspect a 1.6GHz HT would be the bottleneck in the specific benchmark app you used?

If it isn't, then raising the HT clockspeed wouldn't be expected to result in an improved bench result.

 

[Jovec]

The only thing I've seen regarding the HT Link is from a benchmark on the CPU-NB that concluded it's best to keep the HT Link at or around 2000. I was hoping to see a clear trend (that trend could be "it doesn't really matter what speed the HT Link is at") but now my curiosity will probably get me to run some more to see what shakes out. I'll probably also drop the HT if/where it bottlenecks.

Edit: added 1000 and 1200, waiting for DoW2 to download via Steam.

 

[extra]

Hmmmm, a better test might be running some sort of memory bound benchmark? Maybe something involving WinRAR? (Dunno, I've never been into benchmarking).

 

[DrMrLordX]

If you want something that's memory-bound, there's always SuperPi 32m.

 

[Phynaz]

Memory controller isn't related to HT in a single socket system.

 

[extra]

Hmmm, I thought the memory controller talked to the CPU cores and the CPU cores talked to each other over the HT bus?

 

[Jovec]

Yep, I plan to add a few more later today when I get the time. If nothing else, basically this will confirm (for me at least) that the HT Link speed doesn't matter for typical desktop usage.

 

[DrMrLordX]

It matters when you're using an IGP, but not by much. It might also make a difference on systems loaded down with a ton of USB harddrives and PCI-e devices and crap.

 

[Phynaz]

Hmmm, I thought the memory controller talked to the CPU cores and the CPU cores talked to each other over the HT bus?

No, the cores talk to each other through the crossbar switch. HT is only used for the interface to the southbridge and intercpu communication in multi-cpu systems.

Frankly HT is wasted on desktop systems, but it keeps things simpler for a company such as AMD to only have one interface across product lines.

 

[Jovec]

Updated the OP.

The only result that showed any real difference was Dawn of War 2 minimum frame rates at the Ultra preset. Arguably it's worth testing more games at high settings at various HT speeds to see if it holds, but I don't have anything modern with good benchmarking tools to test.

 

[DrMrLordX]

Interesting. Your results reflect what I saw when trying to speed up my Super Pi runs by jacking up HT link speeds.

 

[Nemesis 1]

Super Pi is single thread only one core does the work.

 

[DrMrLordX]

Super Pi is single thread only one core does the work.

I was doing Pi runs on my Sempron 140 (single core). Same results (no change in performance related to HT speed).