DDR4 Memory Speed Tests With The Core i7 6800K On Ubuntu Linux

[phoronix]

Phoronix: DDR4 Memory Speed Tests With The Core i7 6800K On Ubuntu Linux

A few days ago I posted my first Core i7 6800K benchmarks under Ubuntu Linux compared to various other CPUs. Out of requests from some premium members, here are some straight-forward memory clocking tests of the i7-6800K on Ubuntu 16.04 LTS with the Linux 4.8 kernel. Tests were done from dual-channel DDR4 up through quad-channel DDR4 at DDR4-3000MHz.

http://www.phoronix.com/vr.php?view=23566

[schmidtbag]

Did the timings change when messing with the clock speeds? Some of these results don't really make sense.

I'd also be interested to see results of real-world tests. Synthetic memory benchmarks are designed to take advantage of multiple channels.

[TumultuousUnicorn]

Did the timings change when messing with the clock speeds? Some of these results don't really make sense.

I'd also be interested to see results of real-world tests. Synthetic memory benchmarks are designed to take advantage of multiple channels.

Results seem legit: the memory controller support DDR4-2400 (see here), that's why there isn't performance gap after 2400MHz.

[bitman]

Err wait a minute... ii-6800K ark page says it supports DDR4 2400/2133 memory. Are you telling me that page is inaccurate and said CPU runs up to DDR4 3000 MHZ?

[Veto]

Interesting benchmark! I would have liked a single channel benchmark as well though.
I am thinking of making a low power (NAS) build with a Skylake Mini-ITX board and could probably save a bit of power by going single channel. However, I have yet to see a benchmark quantifying the performance difference.

[torsionbar28]

I'd also be interested to see results of real-world tests. Synthetic memory benchmarks are designed to take advantage of multiple channels.

x2, synthetics are a curiosity, but real world application performance is more meaningful. I suspect the difference is far more subtle in real usage, as was the case with DDR3.

[Passso]

x2, synthetics are a curiosity, but real world application performance is more meaningful. I suspect the difference is far more subtle in real usage, as was the case with DDR3.

Many other websites made this kind of tests. Results are all the same : in real world applications, the difference is close to 0 (generally <1%).
Only some benchmarks and a few memory stressing programs can show something.

Difference between DDR3 and DDR4, timings etc. show no difference too. I was so puzzled that I tried myself and... < 1% in all cases!

[Azpegath]

I agree with schmidtbag, it would be interesting to see you do a similar test with a selection of perhaps 4 of these setups with real games (and some compilations?)! All of them probably aren't needed, since we only need to see if there is a correlation between the synthetic tests and a real-world scenario.

[zeroepoch]

I've always wondered what the speed up from dual to quad channel was, but never went looking for benchmark results. I found the article interesting. Maybe others already knew these results, but I didn't.

[coder]

Many other websites made this kind of tests. Results are all the same : in real world applications, the difference is close to 0 (generally <1%).

That's highly dependent on the workload. For some workloads (mostly threaded, data-dependent), memory bandwidth matters a great deal. For others, it's as you say.

perhaps 4 of these setups with real games (and some compilations?)!

Unless you're using an iGPU (which this particular CPU lacks), games generally aren't a good way to stress main memory bandwidth.

Deep learning tends to be memory-bottlenecked (on GPUs, at least).