Announcement

Collapse
No announcement yet.

DDR4 Memory Scaling & DDR4-3600 Testing With AMD Threadripper On Linux

Collapse
X
 
  • Filter
  • Time
  • Show
Clear All
new posts

  • linuxgeex
    replied
    Originally posted by pal666 View Post
    if they are starved due to infinity fabric then faster memory isn't going to help
    Higher memory clock lowers the infinity fabric overhead. That's been proven on various review sites. For Ryzen the faster the memory clock the lower the NUMA performance impact. For Intel where the ring bus operates on its own internal clock improving the speed of the RAM clock doesn't help with NUMA performance... but the ring bus is still faster than Infinity Fabric, even with the fastest modules money can buy for Ryzen, and this is why some of the HPC benchmarks that are memory-starved due to NUMA timings are more than twice as fast on Xeon. Faster memory does help here.

    Leave a comment:


  • pal666
    replied
    Originally posted by davidbepo View Post
    ryzen processors are memory starved due to infinity fabric. no surprises here
    if they are starved due to infinity fabric then faster memory isn't going to help

    Leave a comment:


  • Anty
    replied
    Wut? Where did you got this info from? Of course they can.

    Leave a comment:


  • Apopas
    replied
    Is it still valid that Ryzen processors can not boot if all RAM slots are filled?

    Leave a comment:


  • starshipeleven
    replied
    Originally posted by tiwake View Post
    Think I should try overclocking my ECC ram?

    Leave a comment:


  • Anty
    replied
    Originally posted by linuxgeex View Post

    G.SKILL 4 x 8GB TridentZ RGB F4-3200C16Q-32GTZR module kit has CAS-16 (16-18-18-38) timings at DDR4-3200
    G.SKILL 4 x 4GB Ripjaws V Series F4-3600C17Q-16GVK module kit has CAS-17 (17-18-18-38) timings at DDR4-3600
    and the Ripjaws has CAS-15 (15-16-16-32) @ 3200, CAS-13 (13-14-14-28) @ 2666, and CAS-11 (11-11-11-28) @ 2133

    If the timings were left the same between 3600 and 2133 then 35% of the potential performance of the memory was left on the table due to sheer laziness.
    I have to agree - XMP profile + auto settings for secondary and tertiary subtimings on this platform are bad. Manual tight settings give both better bandwidth and lower latency.
    And it is a pity Michael has only sh...y hynix based RAM for testing
    This 3600@17-18-18-38 with auto subtimings will be worse than properly tuned b-dies at 3200....

    Leave a comment:


  • Del_
    replied
    Originally posted by tiwake View Post
    Think I should try overclocking my ram?
    No. If you need ECC ram, you need stability. Then I recommend using supported frequency and timing for the chips. If you want to overclock, you probably do not need ECC ram, in which case you can by ram rated to 3600MHz from factory. Overclocking is always a game of chance. Anybody telling you differently do not know what they are talking about.

    Leave a comment:


  • Confucij
    replied
    Can you please test Quartus compilation time with Threadreaper? We'r giing to buy workstation for compilation. I prefer AMD but not sure if it will outperform Xeon setup in price/performance.

    Leave a comment:


  • linuxgeex
    replied
    Originally posted by Ehvis View Post
    "The memory timings were not tweaked between the different frequency levels."

    Can you clarify what that means? Because if it means that the timings were kept the same for tests, then the test would be a bit silly.
    G.SKILL 4 x 8GB TridentZ RGB F4-3200C16Q-32GTZR module kit has CAS-16 (16-18-18-38) timings at DDR4-3200
    G.SKILL 4 x 4GB Ripjaws V Series F4-3600C17Q-16GVK module kit has CAS-17 (17-18-18-38) timings at DDR4-3600
    and the Ripjaws has CAS-15 (15-16-16-32) @ 3200, CAS-13 (13-14-14-28) @ 2666, and CAS-11 (11-11-11-28) @ 2133

    If the timings were left the same between 3600 and 2133 then 35% of the potential performance of the memory was left on the table due to sheer laziness.

    It's quality benchmarking like this which sets Phoronix apart from other credible review sites, lol.

    It's also why I will never trust APU benchmarks from this site... because Michael refuses to disclose memory timings for the APU tests, even though I have suggested him how to easily automate that reporting.
    Last edited by linuxgeex; 25 November 2017, 02:07 AM.

    Leave a comment:


  • tiwake
    replied
    Originally posted by coder111 View Post
    Hmm, it's a bit of a shame highest speed ECC memory I can find is DDR4-2666.

    I'd love to have ECC, but if does look like it would be a substantial hit on performance... And probably my wallet as well.
    You can still overclock ECC memory, its just motherboards don't typically allow you to because they are "server" motherboards. I actually literally just got my ryzen 1700x desktop computer to assemble into a computer right now and I have two sticks of 2666mhz 8gig (16 total) crucial ECC memory. I'm very tempted to try overclocking it. With overclocking ECC memory you can easily tell when there are issues (at least with linux).

    Think I should try overclocking my ram?

    Leave a comment:

Working...
X