Comments Locked

30 Comments

Back to Article

  • happyfirst - Friday, October 12, 2018 - link

    I'm in the process of building a new machine (i9-9900k). From what I've read I want to look for the lowest (CL * 2000)/(Speed). I want 64gb. My current machine I average 40-50gb of usage. I'm a software developer and run a lot of vms. I was all set on some G.Skill 3200 CL14 which would be (14*2000)/3200 = 8.75. I can't really find any other memory in the same price range that yields similar or lower numbers. Their current 4600 is CL 19 which would be 8.25, great, but only comes in 8gb sticks so I would max out at 32gb. I do see they say this will come in 32gb pairs. (19*2000)/4800 = 7.91. Comparing numbers (8.75 vs 7.91) does that mean I would really be another 10% faster?

    Is it worth waiting?
    Is 3200 CL14 the current sweet spot?
    How hard is it to overclock memory to 4800? Would I need even better cooling? Right now, I was just going to get a Kraken x72 and I've never worried about cooling memory.
  • SirMaster - Friday, October 12, 2018 - link

    3200 CL14 is the best because it means the RAM is Samsung B-Die which is the best ram available.

    I simply bought 3200 CL14 and overclocked it easily to 4000 CL16 at the stock 1.35v

    I'm sure if I pushed it to 1.5v (like these kits) I could get it to or close to 4500 or maybe 4800 CL19.
  • A5 - Friday, October 12, 2018 - link

    You should probably be on an HEDT platform if you're running enough VMs to use 50GB of RAM regularly and count on it for your livelihood. ECC is your friend.
  • happyfirst - Friday, October 12, 2018 - link

    They're just dev vms. I usually average 3 vms, sometimes up to 6. some of them are just playgrounds. I work with large databases and some queries can use up a lot of memory. I haven't had any problems so far. I've thought about hedt, but I would rather have 8 real cores running at 5 in a cool silent machine then lots more slower cores needing more cooling and fans. What hedt cpu would you choose closest to a 9900k in speed, cores, tdp, and cost?
  • A5 - Friday, October 12, 2018 - link

    Cost is where you're going to lose, obviously. But again, if you're doing this for Actual Work, you should be able to either deduct it or get someone else to pay for it.

    The upcoming 9800X costs about $90 more than the 9900K, but you get quad-channel memory out of the deal, which will do a lot more for you than some CL formula you found online.
  • happyfirst - Friday, October 12, 2018 - link

    I've researched dual vs quad before and never found any major benefits to it. So I've researched it again and still not really seeing anything. Maybe a 'slight' improvement in zipping a file. Could you share any links that truly show quad channel really being worth it?

    Speaking of ECC memory, which would you get? If I went HEDT, assuming I'm doing this right, newegg only shows one single available, a samsung 32gb 2133MHz at $489, so to get quad, I'd need 4 of them, right?
  • UltraWide - Friday, October 12, 2018 - link

    with the HEDT CPU (9800X) you get the full 44 PCIe lanes to the cpu. Optane PCIe storage hooked directly to the CPU might be of interest if you do high I/O work that benefits from low latency.

    Samsung b-die is usually 3200 CAS14 or 3600 CAS15/16. The timings are the clue to getting a higher chance that your DDR4 will be b-die.
  • jon7189 - Sunday, October 14, 2018 - link

    Forgive me for saying it, but you seem so hyper focused on one small attribute that you seem to not be thoughtfully considering some of the excellent advice here. You are running a "server" class workload (VMs and large DBs), and looking a consumer benchmarks (zip files). If you run 4-6 memory intensive DB VMs, then quad channel will double your performance. Also, presumably that 50GB of RAM has to be copied from somewhere, so disk performance will likely be a bottleneck at some point. Get the fastest NVMe drive you can possibly find and connect it to CPU PCIe lanes - don't use the PCH lanes if you want max performance. This is the kind of stuff where the cost of HEDT makes sense.
  • UkeNeverSeme - Friday, October 12, 2018 - link

    That's not how memory works though. Basically for every 400Mhz you increase speed, Cas Latency can go up by 2 while providing the same actual latency. This means that a 3600hz CL16 kit is operating at the same latency as a 3200Mhz CL3200 kit, but at higher bandwidth. Så basically higher clocked memory should always result in better performance as long as Cas Latency does not increase by more than 2 per 400mhz. This is obviously a simplification since some workloads do not scale with memory bandwidth and you would be better off getting even tighter timings. I would recommend finding a good B-die kit (3200 cl14 or 3600 cl16), maybe the Dark Pro Xtreem, that's good value for the money, and seeing how tight timings you can get at 3200, then how high bandwidth you can achieve without sacrificing timings too much, and then testing each setting for your workloads.
    If you have infinite money just get the that 4600mhz CL19 kit and see whatever numbers you can make it do. I'm pretty sure the value proposition is never gonna be there, but hey, if you're rich why not :b
  • MrSpadge - Saturday, October 13, 2018 - link

    That's exactly what's in his formula. It actually gives CAS latency in ns.
  • Tamz_msc - Friday, October 12, 2018 - link

    You won't ever realistically achieve these speeds because the probability of getting a binned chip capable of running these speeds is very low. These speeds are validated after a number of trials in silicon lottery and are only tested on one or two motherboards.

    Just get 3200CL14 and be done with it, or if you want to go higher then maybe 3600CL15 at most.
  • JoeyJoJo123 - Friday, October 12, 2018 - link

    Hi.

    >I'm in the process of building a new machine (i9-9900k)
    Sounds great!

    >From what I've read I want to look for the lowest (CL * 2000)/(Speed). I want 64gb.
    Sorry, but this is going to be really impractical to get. In the memory world, you either go for capacity or speed, but here you want both. To get 64GB of RAM on a board that has 4 DIMM slots, you need 16GB per DIMM sticks. These aren’t easily available on the mainstream platform, and even on enthusiast platforms, these sets of RAM are pricey. Anandtech literally just reported on a very exclusive set of memory that will hit the market that may only work with very specific motherboards. It’s not even known whether a motherboard would even work with 4 sticks of these.
    https://www.anandtech.com/show/13458/gskill-unveil... You’ll strongly want to consider moving to an enthusiast platform (Upcoming Intel Skylake-X Refresh: https://www.anandtech.com/show/13402/intel-basin-f... OR AMD Threadripper 2 platform)

    Additionally, in terms of selecting memory, since you’re doing development work, you may just benefit from opting to use ECC memory rather than trying to min/max memory speeds by getting high speed kits. ECC for system stability for your host and all your VMs, and not losing work to a random blue screen or program crash. However, if you want to calculate true latency for memory, this page on Crucial’s site (http://www.crucial.com/usa/en/memory-performance-s... can help you understand how DRAM MT/s speeds compare with CAS Latency to give you a better picture of true latency for memory access in nanoseconds.

    If you’re dead set on having high speed memory and lots of memory capacity, you’ll probably want to stick to an enthusiast platform (access to 8 DIMM slots) and using 8x8 high speed sticks for your 64GB of desired memory. However, “overclocking” memory between so many sticks will be iffy, and getting timings that stick will also be challenging, so you’ll want to pick out memory kits that are on a vendor’s QVL (qualified vendor list) and are tested and pre-approved to work on your selected motherboard.

    >I was all set on some G.Skill 3200 CL14 which would be (14*2000)/3200 = 8.75. I can't really find any other memory in the same price range that yields similar or lower numbers. Their current 4600 is CL 19 which would be 8.25, great, but only comes in 8gb sticks so I would max out at 32gb. I do see they say this will come in 32gb pairs. (19*2000)/4800 = 7.91. Comparing numbers (8.75 vs 7.91) does that mean I would really be another 10% faster?

    In real world terms, memory with a true latency that’s 10% lower would not be 10% faster, so no. Memory transfer speed and latency is hardly ever the bottleneck for most applications. Some applications may also be more sensitive to DRAM speed as opposed to latency, even if the “true latency” would appear slower (weird, I know). There’s no real way to tell how big or small of a performance gain you’d get from splurging extra on better memory, but in general, if you have extra budget, you should aim for your REAL compute units (CPU/GPU, whatever is doing the computation load) to have more cores or have a higher clock speed. Getting faster memory should really only be a priority up to a point, since after ~3200MHz (at reasonably good CAS latency) any real gain you get diminishes really quickly in terms of how much $$$ you’re spending.

    >Is it worth waiting?
    This is a question only you can really answer.

    >Is 3200 CL14 the current sweet spot?
    For most applications, something better than 3200 CL14 would hardly have a perceptible difference to 3200 CL14. And in some applications, you may not even see a real difference with this kind of high speed kit. It’s entire application specific what gains you’d see, but yeah, I think most people would consider the 3000 or 3200Mhz point to be a point of diminishing returns.

    >How hard is it to overclock memory to 4800? Would I need even better cooling?
    You generally don’t want to overclock memory first. Your first real priority should be finding a stable CPU overclock, the highest you can muster at an acceptable temperature/noise level, and then hoping the integrated memory controller (IMC) of your processor is stable enough at that same overclock and temperature at load to ALSO support boosting the memory speed a bit after that. In some cases, you might find that you can find a good CPU overclock but can’t run your memory kit’s XMP profile anymore after hitting a good overclock, but the XMP profile worked fine prior to the CPU overclock. You’d just have to get lucky with the silicon lottery on your CPU that it’s ALSO happy to push memory speeds while maintaining an overclock.

    >Right now, I was just going to get a Kraken x72 and I've never worried about cooling memory.
    You generally don’t really need to worry about cooling memory, so long as you have some passive airflow going past them. Since it sounds like you’d be opting for a CLC liquid cooler, you may just want to make sure there’s a case fan in the nearest slot to your memory modules or CPU (near the CPU power phases) (ex: the top two fan slots on a case, pointing down) so that there’s passive airflow for your memory to utilize.

    The other thing to keep in mind is that DDR4 normally operates at lower voltages (sub 1.5v). Your CPU’s IMC may degrade over time if you’re pushing a higher voltage (~1.6v) to the memory 24/7. The memory kits are good. They can take the voltage just fine, but the IMC’s stability will degrade over time when pushing that kind of voltage, and you may find it’s not able to support those higher speeds (even XMP settings for high-speed kits) if the IMC degrades enough. This does take several months (if not years) for degradation to happen though, but you’ll want to do reading on what voltages are safe for 24/7 memory overclocks on whatever platform you pick.

    Hopefully this was of some help to you.
  • happyfirst - Friday, October 12, 2018 - link

    Yes, thanks!! Thanks everyone for the help.

    So if I'm understanding this right and I could choose between

    3200CL14 = 8.75 or 3666CL16 = 8.72, I shouldn't focus so much on the formula in that both these scenarios will now have the same 'effective?' latency, but the 3666 memory is running faster?

    I am dizzy... Also I've researched HEDT chips again but don't see anything close to a 9900k unless I want to buy older chips or wait to see what they release soon and then I'm getting into 150 to 200+ TDP chips and would definitely need better cooling. I truly value silent machines and like how with the 6700,7700,8800, etc series people can build dead silent machines.
  • JoeyJoJo123 - Friday, October 12, 2018 - link

    >3200CL14 = 8.75 or 3666CL16 = 8.72, I shouldn't focus so much on the formula in that both these scenarios will now have the same 'effective?' latency, but the 3666 memory is running faster?

    In both scenarios, most people would just say to buy whichever set is cheaper, and given that higher MT/s DRAM is usually priced higher, you might find that 3200CL14 is a much better deal. At the end of the day, a 10% drop in memory latency doesn't mean your software would compile 10% faster, since given Ahmdal's Law, the relative "speed up" you'll see would be a fraction of that. https://en.wikipedia.org/wiki/Amdahl%27s_law.

    In layman's terms, if your CPU is crunching numbers 90% of the time, and 10% of the time waiting for memory to give it (the CPU) more new numbers to crunch, then a 10% speedup of memory latency would only result in an overall 1% improvement (10% * 10%). However, a 10% speedup/overclock of the CPU would likely result in an overall 9% improvement (10% * 90%). Because most workloads aren't memory starved, the relative improvement you'll see by chasing down progressively faster and lower latency memory kits hits diminishing returns quickly. This is why people allocate the majority of their PC budget to CPU and GPU as opposed to faster kits of memory. Sure, moving up to 3200MHz DDR4 will yield improvements over 2400MHz memory for a relatively affordable cost increase, but beyond that, there's not much to improve.

    >I am dizzy... Also I've researched HEDT chips again but don't see anything close to a 9900k unless I want to buy older chips or wait to see what they release soon and then I'm getting into 150 to 200+ TDP chips and would definitely need better cooling.

    I'm not sure how you don't see anything relatively comparable to the 9900k in its price range and core/thread count? When Intel announced the 9900k they also co-announced the 9980XE Skylake Refresh on their HEDT platform. They're both arriving soon(tm). See: https://www.anandtech.com/show/13402/intel-basin-f...

    Yes, these come at higher TDPs, but you could just buy a cooling solution that better fits the higher TDP of HEDT CPUs and still achieve a similar overclock. As someone has stated before, the 9800X will be very comparable to the 9900k, and you'll have more DIMM slots for more memory, and more features that may be useful to your virtualization tasks (such as assigning PCI-e devices exclusively to a VM, such as a GPU).

    You'd still be able to, for example use a Kraken G12 to adapt the Kraken x72 you're currently using for a CPU to cool a GPU, resulting in lower noise levels, so it's not like buying a new/larger CPU cooler to accommodate a higher CPU TDP is a total sunk cost and loss of your Kraken x72. You can utilize both.
  • happyfirst - Friday, October 12, 2018 - link

    I'd seen the 9800X, but slightly higher cost, higher TDP, and only 4.5 vs 5. No idea which will overclock better. I've never needed to assign PCIe devices, I don't game, have 3 monitors and I use cheap fanless graphics cards since all I'm looking at is source code and database results and developing web apps. I wasn't worried about more dimm slots since I can get 64gb in 4 slots. BTW, I found the ECC memory at newegg, under server memory, but they max out at 2666 CAS 17. Is that all really worth it? Seems like to get ECC I have to run slower memory?
  • JoeyJoJo123 - Friday, October 12, 2018 - link

    >but they max out at 2666 CAS 17. Is that all really worth it? Seems like to get ECC I have to run slower memory?

    I don't really think what I'm saying is getting through to you. Why is it that you think enterprise database servers run ECC memory exclusively, despite the loss in memory frequency? Memory latency isn't the bottleneck for these workloads. Spending excess money on faster memory kits will not yield better results than having 100% uptime/availability to your system's resources.

    If you still want to pair high speed memory kits for a PC with a workload that's almost exclusively used for database + development workloads, then please do so. At the end of the day it's your money and your needs and whatever floats your boat is fine with me.
  • happyfirst - Friday, October 12, 2018 - link

    I know why enterprise servers run ECC. I'd rather have a faster machine even if it means giving up .0????01 uptime. I want performance. If my machine crashed, I'd reboot and life goes on. That being said, I don't want it crashing constantly. I care about real world performance. I use to care about benchmarks etc until I finally grasped how often they don't really translate into real world performance. I'm not installing this machine in a datacenter and running live clients off it.
  • Oxford Guy - Friday, October 12, 2018 - link

    Get the ECC RAM if you're a serious professional. You will lose more time on instability than you will from the speed of the RAM being a bit lower.
  • vbigdeli - Friday, October 12, 2018 - link

    Why do you suggest Intel rigs ?? AMD threadripper is better right now..
  • happyfirst - Friday, October 12, 2018 - link

    It's just not that simple to say AMD is better. Many different factors and it's usually an apples to oranges comparison. AMD is about lots of cores, yet in many tests, AMD doesn't beat the intel chips, they just come really close and they are cheaper. And they only come close in the multi-threaded tests due to their higher core counts. Intel pulls away slightly further in single core tests. Not every app is multi-threaded.
  • sa666666 - Friday, October 12, 2018 - link

    But you just said that software development is your main use case, which is where AMD excels with more cores. And being cheaper is nice too.

    Sorry, but it seems you've already made up your mind on exactly what type of CPU and RAM you want, and with every suggestion you kind of miss the point and go back to your original ideas. If you weren't willing to entertain other opinions, why even ask for advice? Just go ahead and buy what you want.

    BTW, I am also a software developer, and have a 64GB RAM development system based on AMD. Probably will upgrade to Threadripper next year. This was the fastest system I could build for this purpose. You should really consider all the alternatives before just blindly buying Intel.
  • Oxford Guy - Friday, October 12, 2018 - link

    "it seems you've already made up your mind on exactly what type of CPU and RAM you want, and with every suggestion you kind of miss the point and go back to your original ideas."

    This.
  • happyfirst - Saturday, October 13, 2018 - link

    My mind isn't made up until it's shipping. Brand loyalty is for idiots. I don't care what brand I have. No chip is the clear winner in every single test. I've been googling intel vs amd for a long while , looked at tons of comparisons with ryzen etc, and sorry, I don't see either as the clear winner. It's not that I'm sticking with my original idea or am not listening to advice. That is just some people are hell bent on their priorites becoming everybody else's priorities and then get all upset when people don't drop everything they've researched and just accept their advice of which there is ALWAYS conflicting advice.

    Since the comments, I am REresearching ryzen again and of course I'm back to where I was this summer, i7-8700k, 1700x, i9-7900x, and now the newer i9-9900k, 2700x, 2950x, i7-9800x or just #@(#@# it and go for something like an i9-9940x or higher. As far as AMD is better, that all depends on what one's PRIMARY priority is. If your definition of apples to apples is to compare similarily priced processors, AMD is better. If you compare core for core, Intel is better. Cost is not 'really' an issue for me. That doesn't mean I will just blindly get a new 9980XE. I have to ask myself is it worth it for me? NOT was it worth it for you. I want the fastest overall machine that can remain dead silent when pushed but I don't want to get into custom water cooling. I'm not watching netflix, playing a game, encoding a video, rendering an image, photoshopping, and compiling an app all at the same time. Very little info on the newer chips. And then this whole ECC for a desktop. Am I lucky that I've never had a problem using non-ECC memory? And then 8 memory slots, what's better 8x8 or 4x16? Why is all the desktop memory fancy with heatsinks and the ecc memory a plain old board? I'm not building a flashy machine, everything going in a fractal R6 solid case that will be hidden away so that it's that much harder to hear so I don't care what the memory looks like, just is it worth it. On a 2950X, is the real world performance difference between slower 2666 ECC and 3200CL14 non-ECC 1%, 10%, 25% ?

    I DO appreciate all the advice and now back on the fence of which way to go. It seems to all boil down to what is the right balance of single vs multi threaded performance for the types of apps I run and what all am I really doing at the same time.
  • Peter2k - Monday, October 15, 2018 - link

    If your workload can use more cores/threads then its AMD
    If you need higher clocks and better boost for single core apps, then Intel

    That said, Threadripper offers the most pcie lanes for lots of M2 drives, something Intel HEDT costs more

    Also the 9900K is soldered, the HEDT chips might not be (the older ones were not) making it easier to tame them cooling wise
  • vbigdeli - Saturday, October 13, 2018 - link

    Yes I have two Intel workstations right now but I'll buy AMD in future due to more security and less costs..I just hope to see some models in Zen and threadripper series with onboard graphics..I don't use them for any graphic or game related programs.
  • RSAUser - Sunday, October 14, 2018 - link

    Intel currently has better virtualization.
    We'll need to wait for the 7nm shrink and see if the click speed improvement will help out there.

    Price/performance, AMD wins and I'll probably get an AMD machine for my next rig as I render on the side, but pure performance Intel wins, so if your work can pay the difference off quickly, it's worth.
    Saving just over a minute or two a day is a workday a year, that easily pays the difference for me.
  • Ej24 - Friday, October 12, 2018 - link

    If you want a crap ton of super fast memory go HEDT. Quad channel memory even at like 2400mhz will trounce dual channel at 4000mhz. I never understood why my heavily overclocked i7 with OC'ed memory at home got destroyed on most workloads by an old haswell era Xeon e5 quad core workstation at work. Its because the Xeon has quad channel memory.

    On dual channel 4000mhz memory you may get 47GB per second bandwidth rather than 39GB per second at 2933mhz. But quad channel will get 80 or 90GB per second at 2400mhz, that's before overclocking.

    Go HEDT, overclock the cpu and the memory and you'll get even more performance gains.
  • Peter2k - Monday, October 15, 2018 - link

    Could be, depends on workload

    But his seems to be more professional

    New HEDT is hex channel btw (I think to have read that)
  • TitovVN1974 - Friday, October 12, 2018 - link

    Having an elder 5820K,
    I would personally go with new Skylake-X Refresh (OR some Xeon, if I could afford it)
    p.s. You may even consider lesser EPYC with 8 channel RAM (make sure first that it suits the workload)
  • TitovVN1974 - Saturday, October 13, 2018 - link

    p.s. Maybe it is worth searching for benchmarks of VMs with your apps and guest OSes, in your host OS, or asking other professionals in Your field ?
    If typical runtime is short, ECC may be less essential.

Log in

Don't have an account? Sign up now