I have the EK torque screwdriver. It worked fine for most EK stuff at 0.6Nm, although Intel seems to recommend 0.9Nm for LGA4189 (which I assume is the same for LGA4677). I’ve been looking to buy one of Wera 074700 (0.3Nm–1.2Nm) but it went out of stock a few weeks ago when I needed it… (perfect timing).
If you’re building a first custom loop, I recommend looking into Aquacomputer’s stuff and some external temperature sensors (e.g., an inline water temperature sensors). Aquacomputer Quadro (or Aquacomputer Aquaero) is an excellent PWM controller that can run autonomously based on an external temperature sensor. (The only downside is that they require a license to update Aquasuite software after 18 months or so, though you don’t really need to run the software).
EKWB’s recommended 0.6Nm seemed to be on the low end of all the recommendations I was seeing for the socket too. As far as I can tell all the sockets use nuts made out of the same PEEK material so it’s weird seeing such a wide variety of torque specs (Asrock recommends 0.7-1.35Nm for the socket, which is the widest range I saw).
I ended up getting a Tohnichi RTD120CN (0.2Nm–1.2Nm)
I got Aquacomputer’s dual pump top due to my paranoia of one of the pumps failing:
I’ll see if I can snag some photos this evening or weekend.
Currently the system is shoved into a CaseLabs S8S case and just the cpu is on water. I have a CaseLabs MH10 arriving next week that I might shove the system in for a bit until the parts for the TX10D become available.
I have most everything connected to an Aquaero right now, but not doing much with it atm besides fans and monitoring temperatures because the case is temporary.
The torque spec is interesting to me. The captured nuts on the 4677 pro block from EKWB seem to be plastic so I was very worried about over-tightening them and stripping the insides out.
The nuts are Polyether ether ketone plastic which is basically the most expensive, performant plastic available… That being said now that I’ve properly torqued the nuts down, 0.6Nm feels like alot.
I haven’t been able to complete my loop yet due to running into some unforeseen issues with a reservoirless setup but I should have more fittings by Wednesday to remedy.
On the subject of memory overclocking,
I started looking at the Kingston FURY Renegade Pro DIMMs and I found out that even the 6000MHz SKUs are running the M88DR5RCD01 RCD which is first generation chip which means that it’s overclocking potential will be limited, the M88DR5RCD02 RCD is already out and is binned for 800MHz more nominal clock speed.
More reason to wait for the G.Skill memory to come out.
3495X has arrived. 256GB kit of 5600, Kingston Fury should arrive this evening. Still waiting on EK water block to ship, and until then I’ve the Noctua NH-U14S DX-4677.
@sirn looks like our processors are consecutive serial numbers.
Will you be doing an updated video? Interested what you’ve learned.
Got the epic deliveries this week. TRPro 5995WX arrived yesterday. Both machines are closely spec’d (RAM capacity + Storage + Video), for A/B comparison.
A Japanese reviewer (Jisakuhibi) noticed some limitation with how Windows group the threads in their 3495X. Since Windows always groups the processor into 64 threads groups, for 112 threads, Windows groups these as 64 threads and 48 threads (verifiable via coreinfo).
Workload that are assigned to 48 threads groups are capped at 48 threads even when manually reassigning to 64 threads groups, which can cause some performance issues. Applications need to be fully aware of these groups to utilize the full 112 threads (Cinebench, Blender, and V-Ray does, Unreal only does for shader compilation, Handbrake kind of does, but heavily biased toward one group).
I wonder how much of the real world performance is affected by this…
Nice video on the Falcon Northwest! (Also nice sneak edit in Floatplane ver.)
Looks like Supermicro is doing the reverse of what others are doing. Vertical socket for 3400 board (X13SWA), and horizontal socket for 2400 board (X13SRA).
The 360 rad on top of the Alta F1 is complete overkill for this thing even under load, the cores stay below 60c running at a 320w load from cpu-z’s stress test:
That’s the little nuvoton PWM controller each fan header has; even though the motherboard’s spec sheet claims it can handle 36w… it clearly cannot. Perhaps the 36 watt figure is for strictly PWM controlled fans and not the 3-pin variety that need a variable DC voltage.
Other observations:
There is no difference in performance when benchmarking between the “Balanced” and “Ultimate performance” in Windows, the only difference is in idle power; Ultimate will yield a 125w idle CPU power draw and Balanced will bring that down to 70w.
The bog standard Hynix HMCG78MEBRA DIMM I got will happily overclock. I had the wrong impression earlier thinking I needed XMP RDIMMs in order to run fast memory. I can get some vanilla RDIMMs and overclock them myself it turns out.
What kind of OC were you able to achieve on your non-XMP advertised M-die RDIMM? Trying to see if its worth it to purchase the Kingston XMP certified kits or just OEM Hynix sticks, thanks!
I bumped up the frequency from 4800MHz to 5600MHz without even adjusting voltage and it passes stress test. I didn’t mess with the timings and they were automatically scaled 16% looser with the 16% frequency scaling. I’m sure they could be tightened up, especially with some added voltage, after all it is only running at 1.1v.
—This is only one 16GB stick so I’m not really going to test what I can get out of it since it is a temporary setup, but I get the impression it could go higher.
I’m going to hopefully get x8 64GB sticks of Hynix A-die, they are available on the market right now. Not sure which RCD to get though, the Hynix A-die DIMMs can be had with either a first generation Renesas RCD (more available) or a first generation Montage RCD (hard to find/expensive). It’s known that the first generation Montage RCD can clock above 5600MHz since that is what Kingston is using for at least some of their kits. It’s also known that the first gen Rambus RCD can clock higher too based off of the testing I did with my current setup.
Not sure how much the PMIC brand matters for overclocking.
Edit: PMIC choice does matter for DIMM overclocking, their are two basic JEDEC flavors of PMIC chips for 12v DIMMs, there is a PMIC5000 and a PMIC5010; the 5010 is a low power (15.3w) version and the 5000 is a high power (25.5w) version. Obviously it is preferable to choose the high power version.
Here’s a handy part number suffix decoder for SK Hynix RAM showing what parts other than the actual DRAM dies are being used:
Update on the overheating onboard fan controller: I swapped out to AP184 fans which are PWM controlled and the temps on the controller chip have come down by over 40C.
I got 8 sticks of HMCG94AEBRA102 in my system now. Overclocking is not nearly as easy as I thought it would be, upping VDD and VDDQ voltage isn’t increasing memory stability, I think I’m missing something.
Here’s where I’m at so far with AIDA64 mem test @ 5800MHz with fairly loose timings:
sort of disappointed to only be getting 220GB/s read speed, not sure if its a problem with aida64 not being optimized for the architecture or cut down xeon sku somehow having reduced memory read performance (maybe less mesh network pathways?).
@sirn@jamiej what kind of read speeds can you get on the AIDA64 memory test?
I saw benchmarks of others with W9-3495x’s getting above 300GB/s.
I’m still running on 6x Kingston KSM48R40BD4TMM-64HMR 4800MT/s. My number looks something like this. This is with stock everything. (I’ll have the last 2 sticks sometime this month and will retest then.)
Per-core L1 looks quite slower than your result (569.5 GB/s × 16 cores vs. 365.1 GB/s × 56 cores). I guess this is proportional to the lower CPU clock? I should possibly try running this again next time around with some OC.
Thanks for the sanity check! It looks like the memory read speed does scale somewhat with core count (or it takes more than 16 cores to saturate the 4 memory controller’s read performance), but not for write speed.
One thing I’ve noticed with my RAM is that with all the DIMMs packed right next to each other, the 2 inner DIMMs run 10-15c hotter than the DIMMs on the edge. I’m going above 80C on the inner DIMMs, you’ll likely be immune from that problem since your DIMMs are spread out further from each other.
Here’s what AIDA54 mem looks like with some BCLK (105MHz), LLC Mesh (2700MHz) and super light core clock (+200MHz) overclocking:
The cpu-z compute performance is about exactly a 7950x’s (although SPR-WS is clocked lower than Zen 4), and as expected the memory is 3-4 times higher than consumer Zen 4; I think this platform is going to compare favorably to Storm Peak, especially since Storm Peak isn’t’ getting Genoa’s 12 channel memory controller setup.
On HSF, waiting for pump, used my last pump building TRPro before I had water block for this build. The Noctua HSF is huge, and RAM runs in the upper-60C under memory intensive loads due to less than ideal air flow. Some air flow is delivered by 3x40mm fans cooling VRM
