Howdy, Stranger!

It looks like you're new here. If you want to get involved, click one of these buttons!

Anand's Memory Test

RidelynnRidelynn Member EpicPosts: 7,383

http://www.anandtech.com/show/6372/memory-performance-16gb-ddr31333-to-ddr32400-on-ivy-bridge-igp-with-gskill/

Interesting results. We've often said that "faster" memory doesn't mean a whole lot on a PC (unless your using integrated graphics). Anand just published their memory speed benchmarks with a Ivy Bridge CPU (in what was basically a big advertisement for G.Skill). I have to admit, it did surprise me some.

This type of test has been done a lot in the past, and usually we see it crop up when shifting memory types (like from DDR2 to DDR3). Almost always, the difference in memory-specific benchmarks is amazing, but the real-world difference is negligible.

Well, I have to say that this article pretty well sums up what we have held as common wisdom. I would have liked to see similar benchmarks with an AMD CPU as well, but maybe we'll get that in another article. The author does a very good job of talking about timing versus clock, and how it all affects performance.

Integrapted graphics get a nice boost by memory speed - no surprises there. Even the sub-par Intel Integrated get a nice lil boost by upping the clock speed (the sweet spot for price at the moment seems to be 1866, there is a steep diminishing return past 2133). Graphic boosts of upwards of 20% were seen when using integrated graphics.

On the general computing side - most tests were around 5% going from 1333 to 2400. A few tests saw some surprising gains (Maya rendering, WinRAR), but just as many didn't see any measurable change at all.

Anand ends up recommending 1866 9-10-9 for general purpose (a 4x4Gb kit at $95). For $20 over the standard 1333, I can't see really arguing against that except for extreme budget builds - $5 a DIMM isn't really a premium, there are some measurable gains to be had, they just aren't very big. Jumping up to 2133 adds an additional $35, and 2400 $50 over the price of the 1866 kit - those just don't seem to really add much to the party *unless you plan on running IGP*, with the 2400 kit showing very small gains over the 2133 kit in almost every test.

Comments

  • QuizzicalQuizzical Member LegendaryPosts: 25,347

    I think all of those kits are ridiculous for most people.  Most people don't have any plausible need for more than 8 GB of system memory.  Even those who do need 16 GB should maybe think about getting two 8 GB modules rather than four 4 GB modules.  For example:

    http://www.newegg.com/Product/Product.aspx?Item=N82E16820231489&name=Desktop-Memory

    http://www.newegg.com/Product/Product.aspx?Item=N82E16820226331&name=Desktop-Memory

    You do want four modules rather than two for a Sandy Bridge-E system.  But Sandy Bridge-E is itself ridiculous for most people.

    That's not to say that the measurements are wrong.  They're still useful information.  They just don't lead to good recommendations on which of the tested kits one should buy, as the correct answer is "none of the above".

  • CleffyCleffy Member RarePosts: 6,412
    One thing I don't like is marketing BS on the speed.  I have some DDR3-1600 at 9-9-9.  When I bump it from 1333 to 1600 it can't handle the overclock and resets on post after I shut down the system.  My board was designed to take 2133 mhz.
  • RidelynnRidelynn Member EpicPosts: 7,383


    Originally posted by Cleffy
    One thing I don't like is marketing BS on the speed.  I have some DDR3-1600 at 9-9-9.  When I bump it from 1333 to 1600 it can't handle the overclock and resets on post after I shut down the system.  My board was designed to take 2133 mhz.

    Well also keep in mind, that to run your memory at those speeds, your not only clocking the memory, but the memory controller. In all modern CPU's, the memory controller is part of the CPU. Intel only supported DDR3-1333 officially until Ivy Bridge, and even then only bumped up tp DDR3-1600. Even the latest AMD CPU's (A10) only officially supports DDR3-1866. So to go past that, your not only clocking the RAM, but also the controller.

  • RidelynnRidelynn Member EpicPosts: 7,383


    Originally posted by Quizzical
    That's not to say that the measurements are wrong.  They're still useful information.  They just don't lead to good recommendations on which of the tested kits one should buy, as the correct answer is "none of the above".

    I only recommended a speed/timing, which is valid no matter what kit you get in whatever size (sure the pricing may be off, but the price differential will remain proportional) - Anand was basically doing a advertisement for G.Skill, and can recommend whatever kits they want based on that.

  • QuizzicalQuizzical Member LegendaryPosts: 25,347

    Another thing to remember is that he's testing with a heavily overclocked Core i7-3770K.  Revert to stock speeds or drop hyperthreading and suddenly the processor can put quite a bit less memory bandwidth to good use.  Trying to find the most demanding test that they can makes sense for writing the article, but it's not representative of the situation that most people will face with their own hardware, even if we restrict to people who have an Ivy Bridge quad core processor.

    1600 MHz memory makes sense for most desktops now that it really isn't any more expensive than 1333 MHz memory.  Maybe 1866 MHz memory makes sense if you're trying to feed integrated graphics from it or need to feed a lot of processor cores, such as the upcoming 8-core Vishera or the Ivy Bridge Core i7 tested.  But for most people, even paying just an extra $10 or so for 1866 MHz memory doesn't make a whole lot of sense.

    Another thing I'd like to point out is that smaller latency gains alone could explain diminishing returns as the clock speeds get higher.  The 1333 and 1600 MHz kits have exactly the same latency timings.  Since memory latency timings are in number of clock cycles, the 1600 MHz kit would have a significant advantage over the 1333 MHz kit even bandwidth isn't a factor at all.  Adding 266 MHz at a time and the timings increase by 0-0-0-0. 0-1-0-4, 0-1-1-0, and then 1-1-2-3.  If you measure latency times in ns rather than clock cycles, each faster kit will have an advantage over the one before it, but it gets smaller as you go faster, and the 2400 MHz kit has roughly the same latency as the 2133 MHz kit.  That might well be why the 2400 MHz kit isn't any faster than the 2133 MHz one.

Sign In or Register to comment.