Which CPU to Handle Memory Reads

[scaryfast]

I have a Q6600, at 3.2, that I run at about 80% usage with up and down constant spikes of about between 40% and 98% (often 100 but not flatlining). I'd push it even more if I could.

Which CPU would best handle virtually constant memory reads?

One program will read the memory of the another program every so often. The program whose memory is being read will be using about 600 MB of memory. And then I use multiple users too.

Would an "on CPU die" memory controller help reduce the CPU load?

Would Intel or AMD chips be better at this task. I would look at benchmarks, but I don't know of any that would really mimic this scenario.

According to Anand, Clarkdale and Arrandale have a little more latency because,
................................................................

"""A side effect of the Clarkdale/Arrandale architecture is that the memory controller is now located on the GPU and not the CPU, although both are still on package and should still be quite low latency."""
http://www.anandtech.com/cpuchipsets/intel/showdoc.aspx?i=3513&p=5

....................................................................................................

I'm not sure if latency is what uses some of the CPU's ability.

What CPU's do you think would best handle my job? (Desktop Class)

Thanks for your help, your input is greatly appreciated!

 

[veri745]

I'd say i7 would probably be fastest, followed closely by Phenom II. I don't have specific benchmarks to refer to, but that's what I recall based on memory performance reviews I've read.

Both have integrated memory controllers, which is what you'll definitely want.

 

[DrMrLordX]

Any LGA1366 processor on the market will give you the best available memory performance at stock speeds (or overclocked speeds) among desktop x86 processors. So that means Core i7 (not to be confused with LGA1156 Core i7 CPUs).

You will probably not need to worry too much about CPU core speeds, and instead worry about system memory, memory controller, and cache performance. So, if you want to do that, you will want to get the highest uncore speed you can safely achieve along with the fastest RAM your budget can accommodate (and then you will want to OC that RAM as well . . . just be careful with the vdimm).

That should get you what you want.

Otherwise, OCing something like a i7-920 to performance levels greater than your Q6600 @ 3.2 ghz should be trivial. That combined with a nice uncore/memory overclock should give you considerable improvement with memory read performance.

 

[BD231]

Phenom II sports the lowest latency implementation but you get more bandwidth with the i7 ... which actually doesn't mean a thing in your case.

I'm not sure what you're privy to paying for your hardware, but I don't think you'd find a better value than in the Phenom II. The i7 is a bit expensive.

 

[Accord99]

I'm not sure if latency is what uses some of the CPU's ability.

What CPU's do you think would best handle my job? (Desktop Class)

Thanks for your help, your input is greatly appreciated!

Can you setup your tasks so you can have an objective way to measure the amount of work done over time? If so, I would suggest you do a test on the Q6600 first, to see what factors affect the performance of your work the most. In this test, I would set the Q6600 to 9x333, to make things simpler, and run the memory at both DDR2-667 and DDR2-800 (or even higher if your setup supports it) and compare the performance gains. Next, if I would try to run in single channel mode to see how dependent it is on bandwidth.

 

[hans007]

Phenom II sports the lowest latency implementation but you get more bandwidth with the i7 ... which actually doesn't mean a thing in your case.

I'm not sure what you're privy to paying for your hardware, but I don't think you'd find a better value than in the Phenom II. The i7 is a bit expensive.

it depends on if you are reading huge blocks of memory , o rjust randomly accessing some area in the 600MB.

if its just lots of random access then you go for latency. if its for bandwidth you'd probably want a core i7 lga 1366 with triple channel (or i guess the latest AMD 12 core has 4 channels so that might be even better)

also if the 2 programs in question are ones you wrote, i have read that if you can eliminate as many mutex calls as possible it will greatly speed up the whole thing because of less user space kernel space calls which are expensive.

 

[Ben90]

Can you setup your tasks so you can have an objective way to measure the amount of work done over time? If so, I would suggest you do a test on the Q6600 first, to see what factors affect the performance of your work the most. In this test, I would set the Q6600 to 9x333, to make things simpler, and run the memory at both DDR2-667 and DDR2-800 (or even higher if your setup supports it) and compare the performance gains. Next, if I would try to run in single channel mode to see how dependent it is on bandwidth.

I agree with this, except I would try first stock: 266x9, then 300x8, then 343x7 (if you can) to see if your FSB is giving you any issues.

I'm not too familiar with LGA775 overclocking, but I'm sure they have RAM multipliers you can play with to see what exactly you are limited on.