Exploring Hyper-threading Performance - Test Systems and Benchmarks
| Xeon Test System | Pentium 4 Test System | |
| CPU(s) | Intel Xeon 2.8GHz (x2) | Intel Pentium 4 3.2GHz |
| Motherboard | MSI E7505 Master 2-F |
Shuttle FB75 |
| Chipset | Intel E7505 | Intel i875P |
| Memory | 4x 256MB Crucial PC2700 | 2x 512MB Kingston PC4000 |
| Storage | Western Digital 800JB | Western Digital 400JB |
| OS | Windows XP, SP1 | Windows XP, SP1 |
The memory used in the Pentium 4 test system was supplied graciously by Kingston. I've used memory from every major player in the industry over the years and this stuff ranks right at the top. It's been fantastic for me.
This is generally the point of an article or review where I'm rubbing my hands together with glee preparing to compare two or more impressive systems. In this case, we'll try not to compare the performance of the Xeons with that of the Pentium 4 test system as we're interested in the trends (if any) in our benchmarking suite when hyper-threading is enabled and disabled. You never know, maybe we'll encounter something interesting.
What benchmarks are we going to look at today?
-
Sisoft Sandra CPU
- Sisoft Sandra CPU Multi-Media
- Sisoft Sandra Memory
- Cinebench 2003 Rendering
- DVD Encoding performance
- PCMark 2004
I would have liked to add some real game benchmarks to our suite for this article but unfortunately the only video card I have available to me right now is an old Elsa Quadro SDR. I doubt anyone really wants to see it struggle along.
All the benchmarks, save our DVD encoding benchmark, are readily available so that you can all try this on your own with your Xeons and P4Cs. Let's get started.
Sandra's CPU benchmark is obviously quite optimized for hyperthreading at this point, and the numbers certainly show that. We see an average improvement of ~52.5% when hyper-threading is enabled on the P4 and an average improvement of ~28% with the Xeons. In the Dhrystone ALU, performance seemingly lags behind a little bit when HT is enabled but I could probably chalk that small percentage up to human error.
As expected another win for hyper-threading. We see an average improvement of ~36% on the P4 and ~34% on the Xeons. Next!
Given that we've already talked about how hyper-threading is a constant struggle for access to resources (execution units, cache and time on the bus to hit system memory) a memory benchmark might prove useful in our analysis. As such, here we are! Unfortunately the results don't really jump out at us. Hyper-threading doesn't seem to have much of an impact on overall memory throughput, as it shouldn't.
Let's move on to a few more interesting results.