HT/HTT Overclocking

[douirc]

ok...so I've been overclocking for almost a year now on the socket 939 procs and have always stuck with the rule of keeping the HTT at 1000 or less. But recently I was reading up on a review (THG or Anandtech...can't remember) and they left the HT multi at 5x when overclocking!? funny enough, they even mentioned the whole downclocking the HTT equal/under 1000 and said they've never underclocked. So I had to give it a shot. I typically run 250x10 with 4x multi on the HT for a 250x4=1000 HTT. I set the HT to 5x for a HTT=1250 and ran a dual SPi 32M successfully. I bumped it up a little higher to 260x5=1300HTT and tried again, only to successfully run another dual SPi 32M. To gain complete confidence, I ran dual Prime95 and played BF2 for about half hour and no problem!

Even more interesting...little things seemed to run faster. Stupid things like menu loading and application access, but maybe I'm just seeing things. I tested SPi 32M results at 4x and 3x multi vs 5x multi and they all seem to provide the same results, so maybe there's no advantage, but what I thought was interested was the stability at HTT greater than 1000.

Anyway, this mean we can just leave HT at 5x from now on?

p.s. sorry if i screwed up the Ht/HTT nomenclature...but I think you get what i'm saying.

[Absolute_0]

placebo effect, you think it might be faster so it seems faster.

Not actually faster though, there have been tests. I always run 3x and my benches are normal.

[[timko]]

From DFI-Street's A64 OCing guide...



Speeding HTT out of spec has about the same significance as 8 cars traveling the same direction on a 16 lane highway and you suddenly make the highway 20 lanes wide. It does not speed up anything. All 8 cars already had their own lane and adding lanes did nothing to speed up anything.




HTT which is a data bus has no problems toting the amount of data it is called on to tote. So it does not matter really how slow or fast you run it except that it will not bootup if too high or too low.

The memory does NOT transfer thru or over that bus so it has no bearing on Sandra memory bandwidth testing either. The memory talks directly to the cpu as the memory controller is there in the cpu. So for all practical purposes any discussion of HTT at the level of the 754 or 939 is mute and only needs be in a 'range' and that range must include enough to boot and not so high it does not boot.

[largon]

DFI-Street - OC'ing for dummies :p:

But yeah, RGone's explanation has a point. HTT is not worth OC'ing.


Overclocking the HyperTransport bus (like any other bus) has 2 effects:

- increased I/O bandwidth
- decreased I/O latency

Decreased latency of course means that more data transfers (clocks) are squeezed in to a single second:
Stock HTT = 2000MT/s (1000MHz), raise it to 2600MT/s (1300MHz) and one data transfer takes 23% less time. 0.77ns instead of 1ns to be exact.
But because HTT bus saturation is rarely (if ever) a problem on K8 systems, real performance gain is not anything like 23%. It's more like 1-3% due to the lower latency. And because HTT is a I/O bus, gains are only visible on extremely I/O reliant applications like high-end 3d graphics.