Memory Latency for Anandtech Skylake Review

[Wall Street]

I am a little bit disappointed with Anandtech's review of the Skylake DDR4 latency.

From the review:

At this point I would also compare the DDR3 to DDR4 results on Skylake above 16MB. It seems that the latency in this region is a lot higher than the others, showing nearly 100 clocks as we move up to 1GB. But it is worth remembering that these tests are against a memory clock of 2133 MHz, whereas the others are at 1866 MHz. As a result, the two lines are more or less equal in terms of absolute time, as we would expect.

This test is measured in CPU clock cycles, so the two latency measures are not more or less equal in terms of absolute time.

Here are the memory timings tested:

Memory DDR Speed Clock Rate CAS Latency Latency Time
DDR3 DDR3-1866 933 Mhz 9 cycles 9.64 ns
DDR4 DDR3-2133 1067 Mhz 15 cycles 14.06 ns

As you can see, the DDR4 at CAS 15 has 45.8% higher latency (14 ns vs. 9.6 ns), as can clearly be seen in the chart in the review.

For this test we took Intel’s most recent high-end i7 processors from the last five generations and set them to 3.0 GHz and with HyperThreading disabled. As each platform uses DDR3, we set the memory across each to DDR3-1866 with a CAS latency of 9. For Skylake we also run at DDR4-2133 C15 as a default speed.

Here you can see them confirming the timings and speeds for the test. Note that the DDR4 is tested is at bone stock and fully supported by Intel and JEDEC. However, note that for the DDR3, they test DDR3-1866 at CAS 9. The Intel CPUs which were tested officially support only DDR3-1600, DDR3-1866 is technically an overclock. Also, JEDEC timings for DDR3-1866 is for CAS latency from 10 - 13 cycles. CAS 9 at DDR3-1866 is out of JEDEC specs.

I know that much faster DDR3 RAM is available and almost nobody runs their DDR3 at stock settings with JEDEC timings. But I can't understand why they would test faster than spec DDR3 without comparing it to faster than spec DDR4-2666 at CAS 14 or DDR4-2400 at CAS 13 which are timings which should be achievable with available kits.

I suspect, and see other threads supporting the idea, that the faster memory can make up for many of the head scratching results in the Skylake review where it performed slower than existing platforms.

 

[Dasa2]

you may find this one interesting i know i did
http://www.hardware.fr/articles/940-5/cpu-ddr4-vs-ddr3-pratique.html

as for skylake performing slower i think you will find that most the time its a driver related gpu performance bug thats getting around or in the case of the anandtech review mostly just margin of error and a gpu bottlneck

http://www.anandtech.com/bench/CPU/1321
Shadow of Mordor on ASUS GTX 980 Strix 4GB
1080p Ultra, Average Frames Per Second

Intel Pentium G3258 3.2 GHz 99.92 FPS
Intel Core i7 6700K 4.0 GHz 99.69 FPS
Intel Core i7 4770K 3.5 GHz 99.62 FPS
Intel Core i7 4790K 4.0 GHz 98.7 FPS

 

[Soulkeeper]

I share your disappointment here, but there are a few other things to consider.

./mem-ns 2133 15
one: 14.06, eight: 17.35
./mem-ns 1866 9
one: 9.65, eight: 13.40
./mem-ns 3200 16
one: 10.00, eight: 12.19

The gap does close on 8 byte transfers, and ddr4 has the ability to "chop" an 8 byte transfer and begin another without blocking. I'd have to look up the specifics on this a bit more, but perhaps this helps it in anything other than simple 1 byte accesses, otherwise the latency disadvantage could hurt more. Maybe someone else has read up more on the specific changes with ddr4

3200 speeds would be a nice oc range in my mind ...
it's a shame that the burst length didn't get a boost this gen

 

[ShintaiDK]

DDR4 is really dissapointing so far. And its quite clear Skylake is memory limited.

Consumer HMC/HBM cant come fast enough.

2011 we had DDR3 1333 CL9. 2015 we got DDR4 2133 CL15.