I’m currently bench testing my setup with this board and a 5975wx CPU. I run 8 x 16gb ecc reg ddr4/3200.
I noticed something when running prime95 for a few hrs,like overnight and that is although CPU runs avg 45c or there abouts for a brief moment (according to hwmonitor max temp) the CPU reaches 95c. This is with dual 140mm Noctua u14 setup. Strange thing is during passmark ,cinebench r23 etc I do not get this peak it seems to only happen after many many hours .
Without pbo I get 50070 on cinebench and 308w draw, with pbo enabled …so far I get 52700 with Upto 420w draw
hmm I’m on nixos so I hadn’t run benchmarks to compare performance (they’re not easy to set up). I’ll try and get one set up to compare PBO settings. the lack of frequency scaling makes me doubt that there’s going to be a significant difference but who knows.
also, 3733 MT/s ended up being much easier to stabilize with tighter timings. I ended up using a procodt of 43.6 Ohms with VDDP set to 0.85V. I’m also pretty confident at this point that a AddrCmdSetup of 61 is ideal for this board.
timings:
tCL: 18
tRCDRD: 22
tRCDWR: 8
tRP: 22
tRAS: 21
tRC: 58
tRRD_s/tRRD_l/tFAW: 4/4/16
tWTR_s/tWTR_l: 6/14 (6/12 or 6/10 might work but I haven’t tested them yet; this appears to be a threadripper limitation)
tRFC/tRFC2/tRFC4: 696/232/174 (still working on tRFC2/4; I get single bit flips after 3 hours of stress testing with those set tighter)
3866MT/s posts but I get wildly incorrect data on first read a decent chunk of the time. 3800MT/s does not post. I might try 3933MT/s just to see if it works any better since apparently some frequencies work much better for the infinity fabric than others.
VMWare has only added ESXi compatibility up to the AQC112C so far, according to the compatibility matrix. The AQC113/AQC114/AQC115C are newer pci-e gen 4 chips relased by Marvell last year in 2021.
I also found this from last year:
I guess the Marvell chips/cards are too “consumery” for VMW to add support for at this point. Most enterprises use the Intel/Dell/HP cards. Maybe they’ll add support later if more people keep bothering VMW/Marvell.
that sucks, it means if i want a smaller mobo, i have to buy both an M2x4nvme card AND an intel NIC, which puts the card out of a reasonable price adjancy
does the Arctic Freezer 4U SP3 fit on the Asrock WRX80 ? in pictures it seems like the heatsink on the VRM closer to the ports side is very tall, and I wonder if that creates an issue
wow, just turned on 4 numa nodes per socket and memory bandwidth improved by 13-14% and memory latency improved by almost 10%. as far as I can tell, it hasn’t impacted cpu throughput performance. the linux scheduler seems to be pretty good about not moving tasks between cores on different numa nodes. if you’re on linux and on this board, set the numa mode to NP4.
well, yeah, but with 64GB of RAM per numa domain, that doesn’t happen terribly often. I’ve also started creating little numactl wrappers for lightly threaded applications, so they’re running on the appropriate CCD given their I/O needs. but a lot of my work happily scales to 64 threads, each accessing a dedicated chunk of RAM, so the numa nodes are pretty much always a win for me.
mlc results with numa off:
Intel(R) Memory Latency Checker - v3.9a
Measuring idle latencies (in ns)...
Numa node
Numa node 0
0 83.2
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using only one thread from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 173380.8
3:1 Reads-Writes : 167018.7
2:1 Reads-Writes : 169705.3
1:1 Reads-Writes : 171710.7
Stream-triad like: 172038.8
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using only one thread from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0
0 173319.8
Measuring Loaded Latencies for the system
Using only one thread from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 454.40 173211.5
00002 453.72 173107.7
00008 459.06 171844.2
00015 465.98 168112.2
00050 397.91 167074.2
00100 370.15 167004.3
00200 310.15 166536.0
00300 234.81 167069.8
00400 143.10 155939.0
00500 124.26 129519.1
00700 109.46 94400.6
01000 103.33 66974.9
01300 100.77 51912.0
01700 98.96 40026.2
02500 97.50 27578.3
03500 96.99 19946.7
05000 96.30 14200.9
09000 95.85 8210.1
20000 95.28 4062.9
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 19.5
Local Socket L2->L2 HITM latency 21.4
mlc results with numa on:
Intel(R) Memory Latency Checker - v3.9a
Measuring idle latencies (in ns)...
Numa node
Numa node 0 1 2 3
0 74.4 81.1 87.4 89.3
1 83.0 75.6 89.9 87.5
2 92.7 89.7 75.5 80.9
3 90.0 87.6 81.3 74.7
Measuring Peak Injection Memory Bandwidths for the system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using only one thread from each core if Hyper-threading is enabled
Using traffic with the following read-write ratios
ALL Reads : 197584.6
3:1 Reads-Writes : 189204.1
2:1 Reads-Writes : 190143.6
1:1 Reads-Writes : 191169.8
Stream-triad like: 192187.3
Measuring Memory Bandwidths between nodes within system
Bandwidths are in MB/sec (1 MB/sec = 1,000,000 Bytes/sec)
Using only one thread from each core if Hyper-threading is enabled
Using Read-only traffic type
Numa node
Numa node 0 1 2 3
0 50137.1 49701.3 47833.5 46873.1
1 48939.3 49192.4 46601.1 47654.2
2 47569.5 46526.4 49177.8 48944.1
3 46635.9 47479.7 48889.5 49226.4
Measuring Loaded Latencies for the system
Using only one thread from each core if Hyper-threading is enabled
Using Read-only traffic type
Inject Latency Bandwidth
Delay (ns) MB/sec
==========================
00000 390.43 197451.4
00002 389.66 197494.3
00008 387.75 197569.1
00015 391.24 197550.5
00050 373.16 197982.1
00100 351.22 198478.3
00200 289.55 198254.2
00300 144.82 192632.0
00400 117.81 157979.6
00500 108.17 130175.2
00700 97.33 94440.8
01000 92.97 67108.1
01300 91.43 51970.8
01700 89.88 40165.1
02500 88.94 27654.7
03500 87.89 20015.0
05000 87.52 14228.8
09000 86.99 8270.7
20000 86.77 4135.4
Measuring cache-to-cache transfer latency (in ns)...
Local Socket L2->L2 HIT latency 19.0
Local Socket L2->L2 HITM latency 20.8
this is, in the worst case - between the most distant numa nodes - only 5ns worse latency than the latency when numa nodes per socket is set to NPS1. that’s more than reasonable imo for an almost 15% win in memory bandwidth and almost 10% win in memory latency when the system is loaded.
@Kish@tristank im thinking if having only one fan (due to the limitation of the board, will be enough due to the quality of the heatsink + having good in/out fans to the case itself
Are you asking how to get to the BNC interface? If so, look at the DCHP record on your router to get the IP. The OUI will be under “ASRock Rack” for the BNC nic. Or you can get the IP from the BIOS, on the BNC tab.
Keep in mind that the BNC takes a while to boot up. So give it a good 5 mins or so after turning on the power switch on the PSU to fully boot.
Or are you asking how to access the remote viewer of the BNC? Go to Remote Viewer tab once you’re logged into the BNC UI. You can use either the HTML5 client or Java.
Note: the viewer only works if you don’t have an external graphics card installed. Or if you have no monitor connected to your external GPU, if you have one installed. I don’t think you can have both a monitor attached and use the remote viewer, as the BNC is unable to redirect the video.