memory for budget machine

[Mark Rose]

Originally Posted by wildrabbitt

Does this motherboard look good for an i5-4590 (that's what I've gone and bought) :


http://www.newegg.com/Product/Produc...=afc-Skimlinks


?

If you're concerned about price, there are motherboards that cost a little more than half that.

[wildrabbitt]

This page says it has linux drivers

http://www.tomshardware.co.uk/forum/...n-chipset.html

The board looks like a good price for the quality. That's why I thought it might be good.

EDIT : posted a bit too late for the previous post.

What boards then?

[wildrabbitt]

Found a nice looking board. Says it supports XMP memory.
The speeds it mentions are 1333 and 1600 but they seem to be for normal memory.

Will it support 1600 XMP memory?

plus, is there any point buying 32GB of memory?

[VBCurtis]

Originally Posted by wildrabbitt

Found a nice looking board. Says it supports XMP memory.
The speeds it mentions are 1333 and 1600 but they seem to be for normal memory.

Will it support 1600 XMP memory?

plus, is there any point buying 32GB of memory?

Since you didn't give us a brand or model, our reading comprehension is no better than yours- it says "supports XMP" and "1600", so sure, it should support 1600 out of the box. XMP is a protocol for memory to identify to the board what speed it can support, so you don't have to mess with BIOS settings. If a stick's XMP speed/latency is not supported by the board, the board defaults to some slower setting (which you can often manually improve- just not automagically).

"any point".... for you, or for anyone? Obviously, there are use cases. Do any of the use cases apply to your interests? Do you have machines that page to disk when you wish they wouldn't? Two examples from this forum are large ECM bounds, and the matrix step of NFS factorizations.

[wildrabbitt]

Cheers for the reply.

The model I'm thinking about buying is this one :

http://www.gigabyte.com/products/pro...px?pid=4731#sp

Perhaps that can help you tell me whether the 1600 XMP will work?

Do you have machines that page to disk when you wish they wouldn't? Two examples from this forum are large ECM bounds, and the matrix step of NFS factorizations.

I take it from the first question that paging to disk is slow and XMP gets around it.

Can I ask : If XMP comes into it's own when doing 'large ECM bounds' and NFS factorisations, what type of mersenne.org work
involves these two things?

[VBCurtis]

Originally Posted by wildrabbitt

I take it from the first question that paging to disk is slow and XMP gets around it.

Can I ask : If XMP comes into it's own when doing 'large ECM bounds' and NFS factorisations, what type of mersenne.org work
involves these two things?

Sorry, that's my failure to multi-quote. I was replying to your second question "any point in 32GB of memory", totally independent of XMP profiles. There are large ECM problems and large NFS factorizations (see the factoring part of the forum for NFS) that require more than 16GB of memory.

As for LL testing, your previous questions & answers in this thread covered that little memory is needed for Prime95/mprime itself, but memory speed is needed. Again, XMP is merely a convenience to save you the hassle of manually setting memory speed/latencies; it speeds nothing up on its own. If you buy a combination where the XMP isn't recognised, you just have to set the mhz speed and latency in BIOS yourself.

If you dig through some older posts in this subforum, you'll find folks discovering that the haswell-generation (any model number starting with 4) quad cores saturate even 1866 mhz memory when running 4 LL tests. If you run 1600mhz memory, you may find that 3 LL tests and a 4th other task (see the rest of the forum for ideas!) may get more overall work done than 4 LL tests, because the 4th test will have to wait for memory access alongside the other 3. Again, this has been discussed with actual numbers in other threads, as opposed to my hand-waving here. I can say that I believe memory bandwidth needs rise with the exponent/FFT size, so reports discussing tests at 60M may not be pessimistic enough for your planned 80M tests.

[Prime95]

Originally Posted by wildrabbitt

Found a nice looking board. Says it supports XMP memory.
The speeds it mentions are 1333 and 1600 but they seem to be for normal memory.

Will it support 1600 XMP memory?

plus, is there any point buying 32GB of memory?

It should support 1600 XMP memory if they the motherboard maker has gone to the trouble of listing that speed. You may need to go into the BIOS settings to have the board use XMP memory speeds rather than "normal" memory speeds.

As I understand your intended use, there is no need for 32GB of memory.

[kracker]

Originally Posted by VBCurtis

If you dig through some older posts in this subforum, you'll find folks discovering that the haswell-generation (any model number starting with 4) quad cores saturate even 1866 mhz memory when running 4 LL tests..

Bottlenecked with 2400MHz after 2 threads for me (quad Haswell)

1792K FFT
1 Worker = 7.2 ms/iter
2 Workers = 7.5 ms/iter
3 Workers = 8.4 ms/iter
4 Workers = 9.8 ms/iter

Also... I've found that dual 1600 is perfect for my dual core Ivy Bridge rigs.

[wildrabbitt]

Where can I learn about the FFT's?

I don't understand what the K number means before FFT.

[retina]

Originally Posted by kracker

Bottlenecked with 2400MHz after 2 threads for me (quad Haswell)

1792K FFT
1 Worker = 7.2 ms/iter
2 Workers = 7.5 ms/iter
3 Workers = 8.4 ms/iter
4 Workers = 9.8 ms/iter

At the very least you see a slowdown which is to be expected. Probably due to increased latency having to wait for other data transfers to complete. But I don't see a bottleneck there, you are still getting more throughput from the machine as you increase the core count.

[VBCurtis]

Originally Posted by wildrabbitt

Where can I learn about the FFT's?

I don't understand what the K number means before FFT.

If you mean "which FFT is used for what size candidate in LL tests?" head to the benchmark page at mersenne.org.

If you mean "where can I learn how fast fourier transformations work mathematically?", try google/wiki/your friendly local upper-division mathematics curriculum.

"K" means either metric thousand, or binary thousand (1024). The number is the size of each transform, and thus is related to the size of each memory transfer in a single iteration. The memory access is *not* equal to the size of the transform, but the numbers are related (I believe linearly). So, a 4096K FFT will take twice as long per iteration as a 2048K transform, and involve at least twice the memory-data-transfer.