01-21-2017 12:21 PM - last edited on 03-06-2024 07:24 PM by ROGBot
01-22-2017 08:14 AM
adamr wrote:
Anyone else have/had issues with a Haswell processor and the new led vengeance DDR4 ram. I'm getting random BSODs. Are there specific settings people are using to get these to work properly?
I have 4x8gig sticks at 3000mhz*
01-22-2017 02:05 PM
01-22-2017 02:49 PM
Korth wrote:
The iMCs built into your Haswell-E i7-5960X will natively support up to DDR4-2133. The iMCs built into your Broadwell-E i7-6950X will natively support up to DDR4-2400.
These are currently the fastest JEDEC DDR4 standards, and they're all the Intel parts (processor and chipset/PCH) are officially rated to support. Anything slower is a defect or a misconfiguration, it should run at rated spec or it should be replaced/repaired. Anything faster is a memory overclock, it's running overspec, there are no promises and no guarantees. 32GB of DDR4-3000 (or faster) is still considered somewhat extreme.
Specs for the R5E show support for up to 64GB DDR4-3300. Specs for the R5E10 show support for up to 128GB DDR4-3333.
ASUS/ROG has overengineered and overkilled everything they can on these mobos to provide these capabilities. I'm guessing they've even binned their bulk Intel chipset (PCH) parts so that they can use only the "best" ones in these flagship mobos. And I'm sure they've tested extensively with every DDR4 part they could obtain because they want to proudly advertise how fast and awesome their mobos are, to set records, and to aggressively outsell competitors in the gaming/enthusiast/overclocker "niche" market. Many people have reported running even faster DDR4 on these mobos. But these ASUS/ROG specs are only "official" in the sense that they're the upper limit on what ASUS/ROG feels confident about promising, they're not a real guarantee that every DDR4-3000 kit will actually work on every mobo for every customer.
Corsair and G.Skill do not actually manufacture DDR4 silicon - they use whatever Hynix, Samsung, or Micron silicon they can get which meets their spec. They purchase it in binned bulk and re-bin it themselves, always culling the "best" bits aside to use in their fastest "extreme" DDR4 products. And they always have to balance price and availability, charging as much they think people will pay, selling it while it can still command highest premiums, while also having enough stock on hand to supply demand. They couple these DDR4 memories with (DIMM-embedded) memory address controllers which are binned and re-binned in similar fashion.
They do manufacture the actual DIMM PCBs, using "optimized" signal paths and thicker copper weights and buffers/caches and added components (like decoupling caps) to reduce latencies. They're fiercely competitive about establishing their brands as the fastest DDR4 available, and charging the most they can get for it - but they can only sell what they're able to buy. Silicon fab process yields are always improving, each new product batch is slightly more refined than its predecessors, and incremental performance increases gradually do add up over time, so DDR4 memory just keeps on getting faster and faster. But Intel and JEDEC have not (yet) caught up, they're far more interested in realistically increasing overall statistical performances/yields for their mass-produced deployments (improving their products) than they are in supporting quirky cherry-picked 1-in-a-million parts which are capable of running much faster.
Some processors have "better" silicon than others. Some have stronger iMCs and UnCore/SysAgent circuit blocks than others (and respond differently to changes in voltages to these subcomponents than others). Almost every processor can be overclocked to some extent, and almost every processor can run faster than DDR4-2133/-2400 memory speeds. But you can't really compare one processor on one motherboard running one memory kit against another - every single silicon part is unique (and full of unique flaws), components and settings which work best on one platform will not work best on any other, components and settings which work for one guy won't necessarily work for another, overclocking is about pushing the limits and it's as much art as science.
Some general rules of thumb for X99 DDR4:
- Always get a proper factory-matched quad-channel kit. Extreme DDR4 is already finicky stuff. Mixed or mismatched kits have multiple (even conflicting) timings and incompatible "guard band" tolerances. Yes you might be able to get them all overclocked together, but you're more likely going to struggle with getting them to work together at all.
- Don't exceed the "fastest" DDR4 supported by your mobo. You might be able to break new records, but you're more likely just going to waste money paying for extreme DDR4 performance which can't be realized.
- The primary bottleneck on DDR4 speed is your X99 processor. Some "above average" procs can overclock, some "above average" procs can run superfast memory, few "above average" procs can simultaneously overclock everything all across the board, most "above average" procs have one strength and several weaknesses. And not all procs are "above average" - the vast majority are simply "average" and (by definition) there's just as many "below average" procs as there are "above average" ones.
- Populating all the DIMM banks on your mobo will put more strain on the iMCs. There's more capacity, more ranks, more density, more signal latencies, more chances for things to go wrong, less chance of every single component and subcomponent being capable of achieving or sustaining extreme performance thresholds. The fastest possible DDR4 performance will adhere strictly to Intel spec: a matched kit of four single-sided DIMMs, for a Haswell-E on an R5E/R5E10 this would ideally be 4x16GB SS DDR4-3300/-3333+. Those fancy heatspreaders and LEDs almost always hide double-sided DIMMs: twice the number of DDR4 chips (each with half the density) on each stick, twice the latency (in worst-case situations) to address two memory "ranks", and about twice the profits from the perspective of the memory manufacturer.
- Don't expect the factory-programmed profile on your extreme DDR4 to actually work "out of the box". You might have to manually tweak timings and settings. And you'll probably have to (reluctantly) bump your VCCIO/SA voltages in BIOS (this is almost a certainty when using the ASUS OC Socket vs the proc's FIVRs).
01-22-2017 04:33 PM
01-22-2017 04:57 PM
01-22-2017 05:16 PM
Korth wrote:
Specific answers to specific questions:
You should probably update your R5E firmware to BIOS 2201, generally considered the best BIOS version for Haswell-E processors. Backup all data first, especially if you have a RAID, lol. I'd usually advise not mucking around with what works unless it promises to actually fix something that's broken or it promises to add more of something you want. But in this instance it looks like it'll be the last and best BIOS ever available for your processor so you might as well enjoy whatever vaguely documented promises of "increased performance" and "increased stability" ASUS has offered.
Not sure which voltages you should adjust, if any, lol it works now so why tempt fate? I learned almost everything I know about overclocking my R5E from reading the overclocking guides on the ROG forums.
Just because someone else got their Corsair DDR4 to run at 3400 or 3600 doesn't guarantee you can do the same, even on the same hardware configuration.
Early-release products usually have a wider (and more random) overclockable performance range beyond the top end. I have an ancient fresh-onto-the-market PNY GTX980 reference card which can run much, much faster than my last-gen factory-overclocked EVGA Classified GTX980 cards. Luck of the silicon lottery, unaffected by the factory binning processes which always develop after enough of the product exists on market that variances can be accurately measured.
Today Corsair bins DDR4 rated for 3000, 3200, 3400, etc. It's pretty much assured that a 3200-rated stick won't perform at 3400 if a 3400-rated stick also exists - if it could, then why would Corsair cut it down? Don't believe the hype about how QA testing on each stick is too costly and time intensive to perform, that was true with later-gen DDR3 but today it is well worth the time because each rung up the binning ladder brings in an ever-increasing profit. People may have done statistical batch checking with manually-pulled random samples from each lot on silicon memory bins a few years ago ... today they run thousands of raw chips (as in, not just the sticks but the actual chips on them) through automated processes every hour and get neatly-tabulated lists of performance metrics for every piece as a matter of routine. I'm not even sure if human labour is still involved at any stage of the process, the chips are peeled off a roll of tape by robotic tools and they're passed from machine to machine at every step (to prevent contamination) until they're put into final customer packaging.
01-23-2017 12:08 AM
01-23-2017 05:38 AM
adamr wrote:
Thanks @Brighttail. Your posts have been very informative. Thing is in my situation running my 32gig kit at 1.2v and 2133mhz ( jdec), I'll still get a random memory management bsod. I've heard that sticks not run at their spec will give issues ... Is this true ? (Even if I run at 2133mhz )* *
Secondly with what you stated above and getting a slight insight in the manufacture process finding a faulty dram stick seems highly unlikely as the automated process would definitely have picked this up . At present I'm busy testing stick at a time
01-23-2017 11:31 AM
Brighttail wrote:
These days getting a faulty stick is a rarity but it does happen. The only way you can test that is the old one stick at a time and testing them.
Honestly, 2133 and 2400 speeds should give you zero issues. I would suggest going into BIOS and resetting it to optimal defaults, test at 2133 or 2400 (whichever your kits base freq is) and see how it goes. The problem in the end may be something with your Overclock.