Intel Haswell Memory Scaling With Ubuntu 14.04 + Linux 3.13

[phoronix]

Phoronix: Intel Haswell Memory Scaling With Ubuntu 14.04 + Linux 3.13

After the recent tests of AMD's Kaveri APU with DDR3-800MHz to DDR3-2133MHz Linux memory testing and following up with AMD Kaveri DDR3-2400MHz testing on Ubuntu Linux, many Phoronix readers followed up with a request of new memory testing done on the Intel side. In this article are benchmarks of a Core i5 Haswell CPU looking at the CPU and graphics performance impact with memory frequency scaling on Ubuntu 14.04 with the Linux 3.13 kernel.

http://www.phoronix.com/vr.php?view=19787

[mercutio]

cas latency

i can't see anything about what cas latency is being used. i see the ram is 9/11/11 at 2133. i've found 8/9/8 settings at ddr3-1600 to work pretty well on haswell, giving some bump over 9/9/9 or 9/10/9. still being faster than ivy bridge cpu with ddr3-2000 9/11/9. some applications benefit a lot more from faster ram than others though. i wonder if more tests are needed of normal cpu type things.

the timed apache compile i suppose is one thing that was significant. i saw some boost with compile times on i7 with tighter timings iirc.

[darkbasic]

Why didn't you test up to 2400MHz? It doesn't seem improvements are negligible...

+1 for CAS latencies

[mercutio]

Why didn't you test up to 2400MHz? It doesn't seem improvements are negligible...

+1 for CAS latencies

he's using b85 chipset, which doesn't support > ddr3-1600. it does support setting memory timings though, at least my asrock board does. the weird thing is that even on things i thought were mostly memory limited, the haswell system beats ivy bridge, even with slower ram. the improvements in cache speed etc with haswell seem pretty significant. and i suppose a lot of things fit in cache most of the time.

[mercutio]

he's using b85 chipset, which doesn't support > ddr3-1600. it does support setting memory timings though, at least my asrock board does. the weird thing is that even on things i thought were mostly memory limited, the haswell system beats ivy bridge, even with slower ram. the improvements in cache speed etc with haswell seem pretty significant. and i suppose a lot of things fit in cache most of the time.

i think it must be cas9 latency or higher, i'm seeing 15292.32 on a i7-4770 for stream copy with ddr3-1600 @ cas 8 latency. i assume i5 vs i7 shouldn't make much difference with such a test.

[mendieta]

I think the most interesting bits by far have to do with video performance. CPU performance gains with memory bandwidth, but CPU clock is king:

http://www.overclockers.com/3step-guide-to-overclock-intel-haswell

The overclockers.com article is really good for anyone overclocking their Haswell. It'll save you TONS of time.

[mercutio]

I think the most interesting bits by far have to do with video performance. CPU performance gains with memory bandwidth, but CPU clock is king:

http://www.overclockers.com/3step-guide-to-overclock-intel-haswell

The overclockers.com article is really good for anyone overclocking their Haswell. It'll save you TONS of time.

it's a bit different with games which is what most overclocking things are focused on. where it's usually video > cpu > memory. on some things memory can make a difference, but it's harder to test. it matters the most when you're using more memory than fits in cache though. so boosting cache makes memory matter less.

[wizard69]

And people wonder why I see DDR4 and other new technologies as very important.

This is a nice well done test that demonstrates clearly what fast RAM can do for you. At this point most users would be wise to wait for DDR4 solutions to hit the market.

[wizard69]

i can't see anything about what cas latency is being used.

Why would you. The data is there to demonstrate what impact faster RAM has on a system. The test would be far less valid if each run was tweaked for fastest RAM performance. Especially if those tweaks are not universally achievable.

i see the ram is 9/11/11 at 2133. i've found 8/9/8 settings at ddr3-1600 to work pretty well on haswell, giving some bump over 9/9/9 or 9/10/9. still being faster than ivy bridge cpu with ddr3-2000 9/11/9. some applications benefit a lot more from faster ram than others though. i wonder if more tests are needed of normal cpu type things.

The tests demonstrate clearly what faster RAM -can- do for you. There is little benefit it excessive granularity here.

the timed apache compile i suppose is one thing that was significant. i saw some boost with compile times on i7 with tighter timings iirc.

[mercutio]

Why would you. The data is there to demonstrate what impact faster RAM has on a system. The test would be far less valid if each run was tweaked for fastest RAM performance. Especially if those tweaks are not universally achievable.

The tests demonstrate clearly what faster RAM -can- do for you. There is little benefit it excessive granularity here.

Sure not all ram can acheive faster timings than it's designed for. But the ram is specced at ddr3-2133 with cas 9 latency. which means it should be able to do cas 7 latency at ddr-1600 speeds.

I happen to be using ddr3-1333 cas9 at ddr3-1600 cas8 latency, which is overclocked a little. i did do memory tests to make sure it was reliable, but I wouldn't expect all ram to be able to manage that. sure the difference is only 10% or thereabouts, but there's a huge jump in cost to go to a faster cpu than a i7-4770, the only real alternative seems to be a i7-4771 which is 100 mhz faster for all cores, and the same turbo speed. whereas there is plenty of faster ram on the market.

Once you have a fast cpu or cpus, a fast ssd, or ssd raid array, then the two biggest differences come from faster video cards or faster ram. it seems those tests were with onboard ram. I'm wondering how much cpu performance is impacted from using onboard video. But I imagine it's a lot less than it used to be, and with faster ram even less. So if being cpu bound is your major concern, and you want to spend a little bit more money rather than a lot more, getting faster ram isn't terribly expensive currently. cheaper than the difference between z87 and b85 chipset motherboards generally, so you're still bound by ddr3-1600.