Ok, after searching through Google and all sorts of articles, I've concluded that each "module" in the 8350 has 2mb l2 cache. Now, my question is, if i disable every other core, would the one core left have 2mb of cache for itself? Or is there a hard limit on how much each core is allowed to draw per module? I've been thinking about doing this lately, one since I don't do anything that really needs all 8 cores, and 2 because I could maybe actually have performance gains with the additional overhead I would have to overclock (half the cores, half the heat in theory). Any idea on if this works/ how to verify it does?
The core to that module would have an undivided L2 cache. In don't think the cache is actually partitioned though when both cores are active. However you should get better per core performance if you disable one core per module. In a few beaches I've seen, gains are typically 10-15% but up to 20% gains can be achieved. However those benches were for bulldozer, not pile driver, so your milage may vary.
That would be the case, but it will not be of a benefit, especially considering the loss of half of the processing ability of the CPU.
With the FX series, AMD went with more cores, and to fit them, they use multiple slower cores. For example, an 8350 needs around a 600 MHz advantage over a Phenom II in ordr to match the single threaded performance.
The lower IPC is one of the trade-offs in addition to the shared components of the core modules
Here is an article detailing the core module performance impact.
Few clarifications which may not be stated clearly. the benchmark application was restricted to 4 threads, so in all cases, it would only allocate 4 threads (up to date windows 7 and 8 are core module aware, and will attempt to use 1 core per module before resorting to using an additional fore in each module.
This benchmark allows us to see how much of an impact the sharing has.
While there is an impact, it is not worth disabling any of the cores.
The reason I was thinking about disabling the other 4 is because that would lower the TDP, meaning I could crank the frequencies higher to compensate for the lack of cores. Also, with what I do I really don't "need" 8 cores, yeah it's nice to have but I think 4 really fast cores would do the trick just fine too. That article is nice to show what kind of gains at a 1:1 clock ratio, but I would be raising the speed so performance should go up and potentially take more advantage of the extra resources available.
Well the key word in this is resource sharing. Sharing is a double-edged sword when it comes to performance. When working on two independent tasks, sharing does hinder performance if both tasks are assigned to the same module while leaving an available module is idle. People tend to focus on that, and it was a major source of Bulldozer's disappointing performance. An one core per module config would be beneficial in that situation. However, if working with a thread that has some common data with another thread, and assigned properly to the same module, it prevents redundant data loading, and increases performance as well as efficiency.
That's why disabling one core per module can be a bit of a mixed bag, especially with the refined Piledriver architecture and appropriate software/OS updates.
Honestly, I haven't seen a lot of benchmarks on this issue, so if you could run some benchmarks for us, that would be awesome.
If you enable "One core per compute unit" in your bios, it will make your CPU a stronger 4-core. You will have less cores, but each core will be more powerful.
I had an 8150 for years, and did plenty of testing with it.
well, nothing keeps you from testing it!
just benchmark, disable the cores, raise clocks and benchmark again.
I was actually planning on doing some personal-use testing but I don't mind posting the results here. Any suggestions on tests? I was going to use the following
3D Mark 11
Metro: Last Light
General Use/multitasking performance
any ideas on other tests that seem reasonable?