@AXm77 seems like it boots fine anyway according to the OP? If the channel orders were wrong you wouldn't have ECC errors, it just wouldn't boot.
@omnibus what's idle power look like?
3060 if you feel brave and know your workload fits in 12GB (most image generation work does with a little finagling), 3090 if you don't feel brave. P40 and P100 are for language model enthusiasts who need all the memory than can get, you'll rarely see the benefit of 24GB for image generation and...
Granite Rapids is what you're waiting for, EMR is just a refresh that supports the next JEDEC memory standard and a few more cores. GNR with 12x MCR DIMM has a ton of theoretical bandwidth (almost a terabyte per second) but it's going to be really hard to hit 1TB/sec in real-world applications...
There are so many better used options that it's hard to justify a new 8124P for home use, especially since DDR5 is *not a feature* at the moment unless you are doing CFD. You can get 4x the cores at the same price with Rome, the same multithreaded performance and slightly less single threaded...
1. Intel and AMD are about the same. Actual performance is going to depend on implementation and firmware settings; Zen 4 and and Golden Cove have comparable IPC and the 7950X and 13900K have comparable 1T turbo speeds.
2. Design choices. No one expected Rome to work as well as it does, but...
If the drives are on a backplane, you could probably bypass the 5V power line and feed it with your own solution, but if I'm not mistaken that's $32K of drives plugged into a $16K NAS - why not find a proper backplane that runs the drives from the 12V rail instead?
This looks like a good deal: Supermicro X11DPH-T LGA 3647 Motherboard with 2x 2U HS Tested 2x Silver 4114 CPU | eBay
You even get a pair of joke CPUs which are still faster than your 2630.
Had a bad time when I did this, would not recommend - some workloads became very slow. You're better off with 8x64GB of slower memory if you're cost constrained - performance will drop a few percent but you'll have no pathologies.
WRX80 is just Epyc with audio and unlocked multipliers, so I imagine the population rules for Epyc also apply. Be warned: AMD platforms can exhibit pathological behavior with anything other than 8 channels.
The torque driver thing is kind of bogus anyway, in practice just tighten the screws until they bottom out but don't wail on them and it'll be fine. The ILM has springs which set the clamping force so it's not like you're relying on tiny deformations in the socket mechanism set by the screw...
aha, spot on. three of the bad DIMMs were missing resistors (?) near the pins and the fourth had a decoupling capacitor ajar.
anyone know what value the 0201 resistors near the edge connector should be or should I just ohm it out?
I finally got around to build up the 8461V @Kizune sent me in August and seem to be in a situation where four (!) of my eight RDIMMs fail to POST on this X13SEI. Two throw "failed component" errors, one says "signal too marginal", and one "fails to train".
I am beginning to suspect this eBay...
Does your application need full lane allocation to each device? The reason why desktop platforms have so few lanes is because lanes cost a lot of money (by my count, DDR5 channels and PCIe 5.0 lanes are something like a dollar a trace!) and you usually don't need full bandwidth to every device.
No. Release stepping ES processors are built from the same wafers as production units. They are likely not subject to the same acceptance testing as production parts, but are built on the same lines and statistically, should be at least as reliable. If you buy a final-stepping ES and its bad...
This site uses cookies to help personalise content, tailor your experience and to keep you logged in if you register.
By continuing to use this site, you are consenting to our use of cookies.