Skylake-X and the 128-GB-of-RAM CAD translation workstation

As some of you probably know, being a one-person IT department can be a blessing and a curse. For me, it means being a jack of many trades and master of woefully few. Suffice it to say that I try, but that it can be difficult to keep up with everything. One part of the job that I had to concede to others years ago was building my users’ PCs. Nowadays, almost everyone is content with a laptop and calls for a desktop are few and far between.

Every once in a while, though, there’s a need for something special. When those needs come along, I can’t resist the call to build them myself. I mean, how else are you going to get exactly what you want? You’re not going to get it from Dell, dude. About four years ago, that need led to the assembly and use of “Godzilla”—an Ivy Bridge-E workstation built with PC enthusiast parts. Godzilla had one job: to be a shared resource for translating customer-supplied CAD data from various formats into native SolidWorks data. He worked well enough—for a time.

The user complaints were sporadic at first, but soon coalesced into a clear message. Godzilla, while still very capable, just didn’t cut the mustard anymore. After considering my options and consulting with the rest of the TR staff, I drew up plans for a system that would tackle Godzilla’s weaknesses. “Gipsy Danger” was born.

75% of the cost showed up in the first box from Newegg

Let’s have a look at her specs.

We’ll talk about those choices more in a moment, but first let’s talk about the logic behind this shared workstation concept and why it leads to building a machine like this. In short, the software is way more expensive than the hardware. As long as my users can play nice together in the same sandbox (which can be a big ask), it’s vastly more cost-effective to build a $4,500 PC than to buy everyone a seat of our preferred translation tool. Not to mention how asking everybody’s own laptops to crunch the numbers would slow down their day-to-day workflow.

I love it when a plan comes together.

We know that doing all the heavy lifting in one place works best, but 128 GB of RAM—really? Yes, really. The translation work is mostly single-threaded, with some exceptions for complex assemblies of multiple parts that the software can spin off to their own threads. That fact still calls for an Intel chip. Gipsy’s 14-core i9-7940X is faster than Godzilla’s quad-core i7-4820K any way you slice it, but let’s face it, not that much faster per-core. No, Gipsy was built to do more things at the same time than Godzilla, pure and simple. Considering that I’ve seen just two simultaneously processing parts use three-quarters of Godzilla’s 64 GB of memory, the 128 GB that Gipsy sports may actually be a little on the low side. Oh, and I got the 3200 MT/s good stuff because it was on sale.

Ain’t she a beaut?

Beyond what I just explained, there isn’t a lot more to say about the i9-7940X choice. It seemed like more, slightly lower-clocked cores would be pointless considering the memory limitations (I can’t believe I just said that about 128 GB of RAM). When I enabled XMP on the Asus motherboard (eight sticks at 15-15-15-35, baby!) I also disabled Multi-Core Enhancement. While I was testing with Prime95, the CPU parked itself at a steady 3.8 GHz on all 14 cores—nice and safe for a workhorse. Speaking of the motherboard, I choose Asus’ Strix X299-E because it wasn’t too extravagant, had decent looking cooling on the VRMs, and because the memory kit was certified for use with it. So far, so good.

This SSD is forever lost under the massive ROG heatsink.

Sometimes the best compliment you can give something is that it needs no explanation. I feel that’s the case for both the Samsung 960 Evo SSD and the EVGA Supernova PSU. Those things are as solid as they come, and I’ve worked with them before. Similarly, the H115i Pro was an easy choice, and it’s keeping the CPU around 50º C under full load. As for the PNY Quadro P4000, well, sometimes you have to pay the piper. A P1000 series card would probably be sufficient, but it just didn’t feel right not putting something a little higher-end into such a nice rig. This way, I have it. At the very least, I needed something with certified drivers so I could get support from software vendors if needed.

Finally, that brings us to the Define R6. Now, I don’t have a lot of experience building in Fractal cases, so there was a small learning curve. Overall, I can see how it all works now, but I had to put things together and take them apart again more times than I would have liked. There’s clearly a lot of thoughtfulness in the case’s design, and it looks incredible, but I still wish it was a bit taller so my preferred top-mounted radiator location had fewer collision issues. Thankfully, I didn’t need the case’s 3.5″ drive mounts and the radiator mount was removable, avoiding what would have caused a lot of frustration building around the top of the motherboard.


Let’s take a look at some rudimentary performance numbers. At the time of this writing, Gipsy has only been in production for just over 24 hours, so I won’t be making any grand proclamations. However, even just using CPU-Z’s built-in benchmark tells the tale I’d expect to see: somewhere around 20-25% single-threaded performance gain, and over 300% multithreaded. That’s not going to save my users a lot of real-time waiting-around when they need something right now, but all those extra cores allow for many more instances of the translation software to run at the same time without bogging things down. I know, who’da thunk?

The type of parts that use 20–30 GB or more of RAM each don’t come around every day, but when they do, they can take double-digit hours to process. Even with 64 GB of memory, Godzilla wasn’t good for much else while chewing on a couple files like that. Gispy’s 128 GB gives her more flexibility. That big pool of RAM isn’t just for running more large files at once—the system has plenty of extra cores available so that folks with simpler needs can skip the line and get their data converted ASAP. Of course, Godzilla is hardly ready for the scrap heap, and for SolidWorks-only work he’ll make a good sidekick to Gipsy for a long time to come.


I hope you enjoyed this impromptu look at my new favorite computer. Some of you probably noticed the thread about this build that I started earlier in the week. I’ll be circling back to that thread soon with more information about real-world usage and performance of both Gipsy and Godzilla. I’ve got a little rendering face-off to run using PhotoView 360, too. Stay tuned.

Colton Westrate

I post Shortbread, I host BBQs, I tell stories, and I strive to keep folks happy.

Comments closed
    • Takeshi7
    • 2 years ago

    That Task Manager CPU screen drives my OCD crazy because that last row of CPU graphs isn’t full. If it were me, I’d have to either buy a 16 core or a 12 core just to make sure to avoid that.

    • dragontamer5788
    • 2 years ago

    I can’t help but think that…

    [quote<] ...but all those extra cores allow for many more instances of the translation software to run at the same time without bogging things down.[/quote<] This implies EPYC superiority. Intel's Skylake-X's chief advantage is incredibly efficient inter-process communication ("Mesh Network") and single-threaded performance. EPYC has 8-channel RAM (vs 6-channel on Skylake-X) and many, many more cores. for the same cost. The 24-core / 48-thread [url=https://www.newegg.com/Product/Product.aspx?item=N82E16819113466<]EPYC 7401P[/url<] is just $1,149.99 for example, with the 32-core / 64-thread 7551P at $2,299.99. Surely, 8-channel RAM would more than make-up for the fact that AMD's memory controller is individually weaker? 8-channel 2666MT/s > 6-channel x 3400MT/s. Wouldn't 8-channel + higher cores probably lead to better multi-threaded performance?? -------------- With that being said, for your use cases... Skylake-X is still a perfectly acceptable choice. I'm kinda curious how an equivalent EPYC system would do under the same circumstances though. For your uses, it seems like Skylake-X > Threadripper in any case. You'd still need to go bigger (!!) than Threadripper. 6-channel + superior memory controller > 4-channel Threadripper. And Skylake-X is still superior on an individual core-per-core basis.

    • yeeeeman
    • 2 years ago

    Why not a Threadripper 1950X?

      • drfish
      • 2 years ago

      It was considered, however…

      Single threaded performance trumps the extra cores.
      [url=https://www.newegg.com/Product/Product.aspx?Item=N82E16820232642<]Equivalent memory[/url<] would have cancelled out any cost savings. Only a [url=https://gskill.com/en/product/f4-2933c14q2-128gfx#tabs-qvl<]handful of mobos support[/url<] such memory - and I was still worried about compatibility based on what I've read/heard from others.

        • brucethemoose
        • 2 years ago

        I suppose that rules out EPYC too.

        You get 16 DIMMs of space, but not enough per-thread performance.

          • mczak
          • 2 years ago

          Epyc is in fact significantly worse than Threadripper if you need high performance for a few threads, since it only boosts to a max of around 3Ghz (exact number depending on model but none is significantly higher) whereas Threadripper has a boost clock of around 4Ghz. (Of course, Threadripper throws energy efficiency completely out of the window with such loads of few threads, but for enthusiast and not server parts that’s quite expected.)

    • highlandr
    • 2 years ago

    Mmm, dat sweet L2…

    • tipoo
    • 2 years ago

    My first thought was CAD = Canadian dollar.

    And it always just translates to sadness.

      • K-L-Waster
      • 2 years ago

      We’re sorry.

    • setaG_lliB
    • 2 years ago

    That CPU-Z screenshot prompted me to bench all of the machines I have lying around. After all, it’s Friday night and I have no life.

      • Mikael33
      • 2 years ago

      I did too, but I only have 1 machine assembled right now haha.
      [url<]https://valid.x86.fr/t8w5pj[/url<] Edit: I should have checked the article date ...

    • martin0641
    • 2 years ago

    I just configured a similar system from Dell, a T640 – except that I put 768GB of ECC RAM and 56 cores in it. Nice little tower, I can likely run a VM or two – though I did go over your cost envelope in doing so…

      • Beahmont
      • 2 years ago

      I suppose the only question left is how many dozens of times did you go over that cost envelope?

      No, seriously. I morbidly curious about just how far over you went.

      • Srsly_Bro
      • 2 years ago

      a VM or two before or after Spectre and Meltdown patches?

    • crystall
    • 2 years ago

    I find that much non-ECC RAM in a single system quite scary. Especially if that system is going to be on all the time and used for professional purposes.

      • blastdoor
      • 2 years ago

      I’d love to see some discussion around this point.

      If you don’t have ECC RAM, what is the error rate, and what is the consequences of those errors?

        • drfish
        • 2 years ago

        Fair question. We make packaging for parts, not the actual parts. The stakes are very low.

          • blastdoor
          • 2 years ago

          I used to think that I had to have ECC RAM because, of course, errors sound like a bad thing.

          But for the last 9 years I have done similar work on machines with and without it and I have not been able to perceive a difference in stability or the results of my work.

          I’ve asked about ECC in many places and the answers range from OMG HOW COULD YOU NOT! to “unless you are running a nuclear power plant don’t bother”

          That’s a pretty broad range of responses so I’m not sure what to make of it.

          I’m kind of leaning towards thinking the nuclear answer might be more right though.

            • tipoo
            • 2 years ago

            Looks like around sub 1% of all DRAM sticks have a bit flip per year, and most bit flips are relatively harmless. This error rate also goes down every year and is outdated. ECC by 2013 had nearly zero error rate, non ECC had 0.83%

            [url<]https://www.pugetsystems.com/pic_disp.php?id=25541&width=800&height=800[/url<] So there's probably uses less serious than nuclear power plants that could use it, but they're not exaggerating too much that very few use cases truly /need/ it anymore, as RAM gets better and better even without it. Thing is though many industries like geological surveying and data science won't blink at the cost increase and are in the "may as well" boat on ECC, just in /case/ it prevents a costly error.

          • DPete27
          • 2 years ago

          Wow! How big are the file sizes if they’re using 20GB of RAM? Seems like the package guy only cares about the outline/shell of the part and not the 10 million little pieces that make up the part.
          Not only would that reduce system load, but could drastically increase conversion time.

            • drfish
            • 2 years ago

            Funnily enough, it’s the surface data that causes the problem much of the time. The “little pieces” aren’t always such a big deal, but the complex curves/contours on the outside of a part that are drawn can get quite intense (especially depending on what kind of data conversion the part has already been subjected to). It’s not quite the same, but I think if it akin to [url=https://techreport.com/review/21404/crysis-2-tessellation-too-much-of-a-good-thing<]this[/url<]. Typically, we're working from 50MB-500MB files, but we occasionally have to processing multi-gig monsters just for the most basic of dimensions or to display our product with the part inside of it. /not a CAD expert, I just build computers

        • crystall
        • 2 years ago

        You can find a very good paper on memory errors here [url<]http://www.cs.toronto.edu/~bianca/papers/sigmetrics09.pdf[/url<]. What they found out was that the error rate for DRAMs was higher than what they had anticipated and in their testing a machine would usually see at least one memory error per week. This was done on machines with significantly less memory than this one and using older memories (DDR and DDR2). On my personal machine (which has DDR3 ECC) I see a lower rate, in a year I can usually find half a dozen errors being logged. However my machine is not on all the time. The impact of the errors is going to be variable: worst case one error might end up in an area of memory which will be written to the disk permanently corrupting that data, but it might also end up in a totally irrelevant area (such as an image being held in memory) so the user won't even notice. Personally the biggest advantage to having ECC is protection from permanent errors. It already happened to me that a DIMM would start exhibiting permanent single-bit error(s) because of age. Identifying the issue w/o ECC usually means running a memory test, in the meantime you might experience slight instability to your machine - or data corruption - without realizing what is happening. The last time it happened it was on my laptop, it took me three days to realize what was happening and I realized it only because I was running large compilation jobs that would use nearly all memory and thus fail rather often. With a lighter workload I might have lived with that error for weeks before realizing what was going on. Again, the biggest issue in that scenario is permanent data corruption. IMHO the price of ECC memory on a machine as large and powerful as the one in this article is really not much compared to the guarantee that you won't run into trouble with memory errors and it's probably worth a lot less than the time usually required to identify and fix them.

          • tipoo
          • 2 years ago

          As you said that’s with DDR1 and DDR2, and RAM error rates have fallen every year:

          [url<]https://www.pugetsystems.com/labs/articles/Advantages-of-ECC-Memory-520/[/url<] By 2013 ECC was posting almost zero errors, and regular was at 0.83%, which made it safer than ECC of years past.

            • Waco
            • 2 years ago

            Safer, but still not “safe”. My NAS logs a half dozen or so bit flips a year. 16 GB (2 sticks) DDR3 ECC at ~7500 ft elevation.

            • UberGerbil
            • 2 years ago

            [quote<] ~7500 ft elevation.[/quote<] Well, there's (most of) your problem. Those of us at sea level are careful to keep most of Earth's atmosphere between us and cosmic rays. You aspiring astronauts have to take additional measures.

            • Waco
            • 2 years ago

            We’re adding cosmic ray detectors to our machine rooms to measure the effect.

            My home NAS sees faults like this far more often than I see in our datacenters (with 100K+ DIMMs), it’s entirely possible I just have a flaky board, memory controller, or stick of RAM. We occasionally get the odd stuck bit that a power cycle will fix at work.

            • tipoo
            • 2 years ago

            I wonder if Thanos’s arrival set off bit flips in RAM around the world. A lesser but serious problem.

            • Srsly_Bro
            • 2 years ago

            I’m sorry for only being able to give you one up vote. You deserve more.

      • Krogoth
      • 2 years ago

      It isn’t really that bad as long as your workload doesn’t require precision.

    • Kretschmer
    • 2 years ago

    But will it run 14 instances of Crysis?

      • Srsly_Bro
      • 2 years ago

      14 instances and 14 up votes? Half life 4 confirmed.

        • highlandr
        • 2 years ago

        I’d upvote you but the 4 you’ve got already backup your discovery.

          • drfish
          • 2 years ago

          You got three from me for that observation, and everyone knows what that means…

    • uni-mitation
    • 2 years ago

    Give us a warning before you put such hotness on display. NSFW.

    uni-mitation

    • uni-mitation
    • 2 years ago

    Duplicate>>>>>>> I need to get a better mouse

    • the
    • 2 years ago

    This reads like a use-case where the throughput is limited by memory capacity. As such, was a dual socket Broadwell-EP system considered? Depending on the board, that’d permit 256 GB of non-ECC memory. Registered ECC would also be an option which can be up to 128 GB per DIMM. Only downside would be paying the Xeon premium, especially in high clocked models for the necessary single threaded performance.

      • drfish
      • 2 years ago

      Nothing that exotic was really on the radar. 128GB was about all the RAM I wanted to pay for anyway, and I was shooting for the highest clocks possible. Depending on how things go, my next step would probably be trying out chuck’s Optane idea if necessary.

        • GTVic
        • 2 years ago

        Did you consider a configuration similar to the Lenovo P520 with a W-Series Xeon W-2145, that is only 8 cores but starts at 3.7GHz up to 4.5GHz.

          • drfish
          • 2 years ago

          I did not. I know I used “workstation” in the headline but this was always destined to be a Skylake-X or Threadripper “HEDT” system. I knew what I could get away with is this case, but would totally switch gears to Xeon/Eypc/ECC under different circumstances. Plus, everyone in the office is digging the RGB LEDs. 😉

          • Srsly_Bro
          • 2 years ago

          8 core CPUs don’t have enough cores.

          • GTVic
          • 2 years ago

          Thanks, W-Series motherboards definitely are not easy to find for people building a system, definitely not with LEDs 🙂

          When they first were announced it looked like some potential for high end enthusiast setups so I’ve been keeping an eye on them and we will be getting some P520s in at work shortly.

          I guess now you get a thumbs down for even asking questions …

    • chuckula
    • 2 years ago

    [quote<] Considering that I've seen just two simultaneously processing parts use three-quarters of Godzilla's 64 GB of memory, the 128 GB that Gipsy sports may actually be a little on the low side.[/quote<] For all the people who moan about Optane you just found a great use-case for picking up one of the M.2 variants: Throw it in only as a swap ("virtual memory" in Windows-speak) device that's just there to augment your already maxed-out RAM. Even the 118GB model is [url=https://www.newegg.com/Product/Product.aspx?Item=N82E16820167453&cm_re=optane-_-20-167-453-_-Product<]under $200[/url<] which is not much at all for this type of system and less than 1/8th the price of the RAM at its current list price.

      • davidbowser
      • 2 years ago

      I would think this would be a great use case and relatively easy to benchmark.

      • drfish
      • 2 years ago

      That’s a solid idea, and I’ve got a spare M.2 slot to play with. We’ll see how fast “guys, just run more instances” starts to cause problems.

      • Srsly_Bro
      • 2 years ago

      I think this would have been great, also.

      [url<]https://www.intel.com/content/www/us/en/solid-state-drives/intel-ssd-software-defined-memory-with-vm.html[/url<] I think this is Xeon only, tho...

      • Mr Bill
      • 2 years ago

      An excellent suggestion, which time has finally come.

    • Mr Bill
    • 2 years ago

    This is the really cool take-away for multicore systems.
    [quote<]...but all those extra cores allow for many more instances of the translation software to run at the same time without bogging things down.[/quote<]

    • chuckula
    • 2 years ago

    [quote<]About four years ago, that need led to the assembly and use of "Godzilla"—an Ivy Bridge-E workstation built with PC enthusiast parts. [/quote<] We did something like that at the end of 2013 too. Why are you in such a hurry to change the hardware? -- Apple

      • alloyD
      • 2 years ago

      Why are you waiting so long to change the hardware?
      — Apple (Mobile Products Division)

      • tipoo
      • 2 years ago

      Le sigh. They said it’s not launching till 2019 so that’s an astounding 6 year gap between “Pro” platforms. That’s a crazy amount of time in silicon years.

      The iMac Pro is pretty cool but not for all needs.

Pin It on Pinterest

Share This