Thanks a lot. Following ur instruction, the bios have been found that there is no 06F3, only 2 06F8(different microcode id) and 1 06F2.
If i want to support the ES as much as possible.(I have QYFS, Q0KG, Q30G) what should I do?
ES Xeon Discussion
- Thread starter britinpdx
- Start date
Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
If i want to support the ES as much as possible.(I have QYFS, Q0KG, Q30G) what should I do?
Thank a lot again. I am waiting my BIOS IC programmer to arrive. I will try immediately after.
no way. Q30H is Emerald Rapids and not a WS processor.
SPR-SP D0 only work because there are SPR-WS ES stepping D0(step 3) processors.
Last edited:
Anyone have IPEX benchmarks for 4th or 5th gen ES, QYFR or SRMJ? 111gb/s on full 16 DIMMs on int8 7b model but only getting 2 tokens/sec. AVX/AMX getting blocked? Tried node assigning too, no luck.
SRMJ on HBM saturated I am reaching about 5 token/sec and made sure have memory bound HBM, but am not even saturating bandwidth. Though do see prompt generated improvements compared to DDR5 for 9480.
Something BIOS am missing on both platforms? AVX throttling? ES woe?
Maybe not best place to ask this in particular.
SRMJ on HBM saturated I am reaching about 5 token/sec and made sure have memory bound HBM, but am not even saturating bandwidth. Though do see prompt generated improvements compared to DDR5 for 9480.
Something BIOS am missing on both platforms? AVX throttling? ES woe?
Maybe not best place to ask this in particular.
In any case, you need to disable NUMA virtualization. Under Windows OS, you should not wait for normal speeds. You can also use the ik_llama.cpp fork for model interference. If you have a two-socket system, then you will not be able to fully load it due to UPI speed limits.
I previously used a two-socket system in conjunction with several GPUs. Through KTransformers, one model can be loaded onto each socket, thereby significantly increasing the speed of access from GPU to RAM.
Yes, I run dual socket configs and assign by socket, 2,3,6 and 1,4,5 already with breakout boards broadening that up to 9+8 GPUs at x4. This was only a test I was looking at with CPUs on other results some people have on non ES examples. But I figured it out! I apparently was not running int8 and was on bf/fp16.
Non-ES info but I've noticed recently market had price adjustment (or new-obsolete stock release) for SPR 8473C (for Microsoft) and 8487C (not sure to whom those were intended for - AWS, GCP, AliBaba, IBM, etc.) at some ok-ish prices, I think both models are designed for multi-socket configurations (UPIs unlocked),
None of sellers on ebay are first-hand owning it - therefore they're only resellers (and just charge some margin on input price), so don't expect material discounts (I think around 5% max), which probably still makes those CPUs a bit overpriced (subject to how much value you can extract for your use-scenarios of course), at these prices I would get those 3-4 years ago (and was actually submitting offers in this range) but not now anymore.
None of sellers on ebay are first-hand owning it - therefore they're only resellers (and just charge some margin on input price), so don't expect material discounts (I think around 5% max), which probably still makes those CPUs a bit overpriced (subject to how much value you can extract for your use-scenarios of course), at these prices I would get those 3-4 years ago (and was actually submitting offers in this range) but not now anymore.
Last edited:
i guess these are 806F6 stepping E3 EV-QS (early validation QS)
8487C EV-QS
8473C EV-QS Q172
correction: non QS at same low prices(ebay)
Last edited:
this does makes sense, ebay images for 8473C show L240.. (L= Malaysia, 2=year 2022, week 40) - period they started extending SKU range offering,
(and for 8487C = L211.. , year 2022 week 11), although as you have explained previously - seller have no troubles re-print anything they need on lids.
they usualy show production units to explain what's coming and to hide the source of the QS.
QS: you will get a processor with "Intel Confidential" printed on it not what pictured.
example: recently sold 8470 E3 QS:
ebay listing 8470
i asked for QDF, seller replied: " this is Q16R "
sell ES/QS is just not allowed(for the original ES customer), reprint to sell as retail is illegal/crime act, even in china.
Last edited:
Any tips or thoughts on how to reduce idle power consumption on these setups? I'm not overly concerned with it, but my idle power consumption according to BMC is around 360W.
System specs:
OS: Unraid 7.1.4
CPU: Xeon Q2SR (1/2 cores disabled)
MB: Gigabyte MS03-CE0
Memory: 256GB 4800MHz ECC
GPU: Intel Arc Pro A40
I do have 10x3.5" HDDs, a few 2.5" SSDs, 2xM.2 drives, and an intel x710 NIC taking some power
When I had my W680 setup with a 14900k the idle power was around 200W. I don't expect that but if I'm missing something obvious it would be good to know. Overall I love the build and being able to dedicate "real" cores to my VMs.
System specs:
OS: Unraid 7.1.4
CPU: Xeon Q2SR (1/2 cores disabled)
MB: Gigabyte MS03-CE0
Memory: 256GB 4800MHz ECC
GPU: Intel Arc Pro A40
I do have 10x3.5" HDDs, a few 2.5" SSDs, 2xM.2 drives, and an intel x710 NIC taking some power
When I had my W680 setup with a 14900k the idle power was around 200W. I don't expect that but if I'm missing something obvious it would be good to know. Overall I love the build and being able to dedicate "real" cores to my VMs.
sorry but this is real off topic here.
there are many similar discussions with tips and todos.
your mesh is full engaged 2400mhz, not IDLE. high perform.plan never idles....
around 55Watts.
high perf.plan. min.processor state 25%
- Processor idle demote threshold 95
- Processor idle promote threshold 98
I'll have to see if there are any settings in the bios I can modify to lower the power consumption of the CPU itself. Unraid does not have as many tunables as windows.
you should first check power consumption of all single parts. or you tweak BIOS not knowing CPU is the special part.