ES Xeon Discussion

[scouzi]

no a joke, warranty by the seller is usual, mostly one year(taobao). just read the listing or ask the seller.
even $400usd is too much if the processore dies after 4 weeks.

Compared to $10k plus for retail - buying another as spare is worth it.

 

[DHamov]

even $400usd is too much if the processor dies after 4 weeks.

Strongly agree. And i am still a bit worried about that. Trying to make a Taobao account to see more, but google translate does not do the Chinese captcha, and i need to many tries to pass it.

Compared to $10k plus for retail - buying another as spare is worth it.

I get what you mean, but it is really only true if you would know the probability of failure or at least an upper bound of it.
If the other one also fails after a week, and the next and the next. you can't make a good investment on these, no matter how many one would buy, as a back up. It's a bandit problem (finding out what slot machine gives the best returns). If i would have the money to buy a spare, and if i would be rational i would probably by one of another type of cpu compatible with the motherboard. Reconsidering that i should have probably bought a Motherboard, that had the better compatibility with later steppings. But i got a good deal Asus W790-ACE for 500 Euro.

Even though, we only have very small sample statistics available on the web and here about this topic, it seems that reliability is not such a big issue in practice as it is in theory. Because almost nobody is complaining, and like with reviews usually people with negative experiences are more likely to post reviews than, people with positive experiences. So my heuristic is to calculate with an cpu failure probability of 10% per year for these ES, it still makes it worth while, for my purposes. Calculating everything else with a pessimistic 2% per year. Also buying the real deal official intel cpu of $10K is not possible for me, also because loses its value very strongly, in 3 years its maybe worth $500 or so. Changing perspective 180 degrees, why are all these QYXX being sold, instead of used for cheap cloud services, probably because the sellers can get or do something better for the money, maybe E steppings? Or if something else, anyway i probably do not have access to it.

 

[scouzi]

Changing perspective 180 degrees, why are all these QYXX being sold, instead of used for cheap cloud services, probably because the sellers can get or do something better for the money, maybe E steppings? Or if something else, anyway i probably do not have access to it.

Unless you have some requirement of the special SPR features, the overall system TCO is not cheaper than AMD when you consider the rest of the components. You can get get more bang for the buck buying used recycled Hyperscaler Milan Epycs with DDR4 if all you want a powerful workstation. AMD TR and Epyc ecosystem seems cheaper. There are also some Genoa Epyc QS systems out there for sale in the wild.

I don't quite understand why LGA4677 MB are hard to find in North America. I suspect the large majority of the SPR market is for OEMs and Hyperscalers and there is almost 0 demand for individual builds. No one keeps these in stock.

I did receive my ES and it does appear to be legit. It was used however judging by slight cooler marks. Maybe for testing the CPU itself before selling it or some Intel partner/supplier work. I don't know. Just waiting for the rest of the pieces to trickle in.

 

[DHamov]

You can get get more bang for the buck buying used recycled Hyperscaler Milan Epycs with DDR4 if all you want a powerful workstation.

In general that is true. Also for ddr5 systems AMD threadripper and Ryzen can give more bang for the buck, than official intel products, in terms of hardware specs. However, for Math Kernel Library (MKL) applications which for particular applications is still the best mathematical library in the world and many softwares use it by default. Unfortunately also older versions. In dependence of MKL AMD can be slow or extra headache, needing to recompile things, with other flags, there are fewer users, smaller community, and therefore its more difficult to find helpfull support. The library was developed by intel, and they made it in a way to have an advantage over competitors. So i think for now, MKL+ Intel CPU is still the 'better' choice for a lot of scientific computation.
Not necessarily because it's really better, but extra programming and debugging efforts are also cost that go in the overall consideration of what hardware to buy. Similarly with NVIDIA GPU's and CUDA, all big libraries Pytorch, Tensoflow are faster and easier with NVIDIA gpus. Even if their products are or will become less hardware cost-efficient, they will not lose the lead so fast. And unfortunately for users, (not investors) that will probably make the price difference larger, w.r.t competitors larger. By buying their (intel and Nvidia) stuff eventually on the long run we will be paying for that growing monopoly gap. Except of course when buying ES cpu's. So maybe its even ethically a well-justifiable choice. Also considered early 12900 cpu, those with the round intel logo have surprisingly have AVX512 if e-cores are disabled. But honestly, they remain just 8 p-cores, memory and pcie channels made me look towards ES Xeons.

 

[scouzi]

In general that is true. Also for ddr5 systems AMD threadripper and Ryzen can give more bang for the buck, than official intel products, in terms of hardware specs. However, for Math Kernel Library (MKL) applications which for particular applications is still the best mathematical library in the world and many softwares use it by default. Unfortunately also older versions. In dependence of MKL AMD can be slow or extra headache, needing to recompile things, with other flags, there are fewer users, smaller community, and therefore its more difficult to find helpfull support. The library was developed by intel, and they made it in a way to have an advantage over competitors. So i think for now, MKL+ Intel CPU is still the 'better' choice for a lot of scientific computation.
Not necessarily because it's really better, but extra programming and debugging efforts are also cost that go in the overall consideration of what hardware to buy. Similarly with NVIDIA GPU's and CUDA, all big libraries Pytorch, Tensoflow are faster and easier with NVIDIA gpus. Even if their products are or will become less hardware cost-efficient, they will not lose the lead so fast. And unfortunately for users, (not investors) that will probably make the price difference larger, w.r.t competitors larger. By buying their (intel and Nvidia) stuff eventually on the long run we will be paying for that growing monopoly gap. Except of course when buying ES cpu's. So maybe its even ethically a well-justifiable choice. Also considered early 12900 cpu, those with the round intel logo have surprisingly have AVX512 if e-cores are disabled. But honestly, they remain just 8 p-cores, memory and pcie channels made me look towards ES Xeons.

Also for CUDA - NVDIA - everyone develops AI tooling with CUDA as the default. If you want to use AMD or Intel GPUs, you have to refactor and reconfig away from CUDA with other librairies or packages to override it. In can be a PITA. A lot of people don't realize that CUDA is 18 years old and also sill supports Fortran. It's not just about GPU - but the maturity of CUDA is also what is hard to disrupt.

 

[DHamov]

I hope it is ok if i move the discussion from the other thread to here, because it is more general.
Its a bit of a mess with quotes of quotes, i apologize and hope the mess is not to big.

HWiNFO64 report

View attachment 34108

MS03-CEO or MS73 with 4th/5th gen BIOS(R0x) supports 5600 memory, no overclocking.
if CPU supports 5600 or mem. overclocking 5600 is working.

Thanks its clear now, that it works.
But now from HWiNFO64 reports we can know that overclocking works for that cpu, any other pointers how one could know that before owning the cpu?...

1)

if CPU supports 5600 or mem. overclocking 5600 is working.

2)

if they write 4800 then this is max. except unlocked X/K models

At intel they write
Up to DDR5 4800 MT/s 1DPC
Up to DDR5 4400 MT/s 2DPC
for the official version of this cpu (an 8461V i presume).
So 2) applies and its not an X/K model, but the report you showed (for a 8461V ES showed it supports memory overclocking so 1) also applies, and of course 1) overrides 2).

So if i understand it correctly 8461V Es would run on MS03-CEO or MS73 with 5600 jedec, 1.1v RDIMMS really at 5600MHz speeds.
Now how to generalize this, using 1 and 2? And apply it to specific examples.
For example Q0KG or 8490H ES Q03J E0 stepping, would they run also run memory (with jedec 5600 and 1.1v) with 5600 on MS03-CEO or MS73?
And since they are supposed to run on any C741 board... Or am i now insinuating something that is not true?

 

[RolloZ170]

At intel they write
Up to DDR5 4800 MT/s 1DPC
Up to DDR5 4400 MT/s 2DPC
for the official version of this cpu.
So 2) applies and its not an X/K model, but the report you showed was for a 8461V ES i presume, so 1) also applies, and of course 1) overrides 2).

correct: Q16Z 8461V EV-QS supports mem.overclocking, the production unit (retail) not.
in theory...BIOS prgrammer limits RAM clock to 4800 on SPR.

 

[scouzi]

Does 6400Mhz DDR5 XMP make any sense with a QYFS on ACE or MS03-CE or is it overkill?

 

[RolloZ170]

Does 6400Mhz DDR5 XMP make any sense with a QYFS on ACE or MS03-CE or is it overkill?

works only on ASUS W790 Ace, but QYFS supports mem.oc but no guarantee how high.
gigabyte C741 = no XMP, means only jedec speed.

 

[RolloZ170]

For example Q0KG or 8490H ES Q03J E0 stepping, would they run also run memory (with jedec 5600 and 1.1v) with 5600 on MS03-CEO or MS73?

there are some more E0 stepping: maybe anyone have and can check the memory frequency support ?
Q03J E0 60C 1.60Ghz similar plat 8490H
Q0KG E0 52C 2.00Ghz similar plat 8470
Q0KL E0 52C 2.10Ghz similar plat 8470
Q0KH E0 48C 2.20Ghz similar plat 8468
Q0KJ E0 48C 2.10Ghz similar plat 8469C
Q0KS E0 36C 2.00Ghz similar plat 8452y

btw: not all Q0__ are E0 stepping:
Q0MN E3(EV-QS) 48C 2.10Ghz similar plat 8468

 

[DHamov]

there are some more E0 stepping:...

Nice list. i wrote two ebay sellers yesterday night. One of them also sold a Supermicro dual socket board so i asked for compatibility and hwinfo report about memory overclock. But unfortunately no useful feedback yet.

 

[scouzi]

works only on ASUS W790 Ace, but QYFS supports mem.oc but no guarantee how high.
gigabyte C741 = no XMP, means only jedec speed.

Manual of MS03-CE0 states max memory freq is 4800 for gen4 and 5600 for gen5. Is this accurate? Will it downclock faster jedec memory to 4800? No reason to buy higher than 4800 unless for eventual upgrade to Gen5 Xeon?

 

[RolloZ170]

--

Manual of MS03-CE0 states max memory freq is 4800 for gen4 and 5600 for gen5. Is this accurate? Will it downclock faster jedec memory to 4800? No reason to buy higher than 4800 unless for eventual upgrade to Gen5 Xeon?

you can buy JEDEC 5600 modules, they have profiles for 4800.
note gigabyte c741 & W790 do not support XMP, you need real jedec.
i.e. the Kingston Renegade Pro DDR5-5600( KF556R36RB-16 ) is jedec 4800, runs only @5600 if you select XMP profile #1.
• Default (JEDEC): DDR5-4800 CL40-39-39 @1.1V
• XMP Profile #1: DDR5-5600 CL36-38-38 @1.25V
• XMP Profile #2: DDR5-5200 CL36-40-40 @1.25V
• XMP Profile #3: DDR5-4800 CL36-38-38 @1.1V

 

[DHamov]

Manual of MS03-CE0 states max memory freq is 4800 for gen4 and 5600 for gen5. Is this accurate? Will it downclock faster jedec memory to 4800? No reason to buy higher than 4800 unless for eventual upgrade to Gen5 Xeon?

The interesting thing is that some D0 stepping cpu's like QYFS and some others support memory overclocking, while the official cpu versions do not. So those 4th gen cpu's could maybe still run at 5600MT with jedec 5600 memory, even at not overclocking board. I hope some one with such a cpu a GIGABYTE Board and Jedec 5600 memory can test this and confirm with a screenshot.

Do QYFP and/or QYFQ also support memory overclocking?

 

[RolloZ170]

Do QYFP and/or QYFQ also support memory overclocking?

yes, correct.

 

[scouzi]

The interesting thing is that some D0 stepping cpu's like QYFS and some others support memory overclocking, while the official cpu versions do not. So those 4th gen cpu's could maybe still run at 5600MT with jedec 5600 memory, even at not overclocking board. I hope some one with such a cpu a GIGABYTE Board and Jedec 5600 memory can test this and confirm with a screenshot.

Do QYFP and/or QYFQ also support memory overclocking?

I may end up with that confirmation soon. I just got a tracking number for my MS03-CE0 board. Debating on buying 4800 vs 5600 now on a pure cost basis since I won't benefit from 5600 it seems unless I upgrade to Gen5 in the future.

Question for the board. What is the most impactful performance wise for memory? Using the most channels or running higher spec memory? (tighter timings, higher freqs etc, Tuned kits etc)

 

[DHamov]

populate 8 channels, dual ranked if you can get.

Sorry to ask or note, but is dual ranked really better?
i thought that each extra rank is considered as a additional RDIMM for that channel and because
For many cpu they write
Up to DDR5 4800 MT/s 1DPC
Up to DDR5 4400 MT/s 2DPC
Is the extra rank from the dual rank considered equivalent as an extra DIMM or not?
Because i would be afraid it would clock things down, thats why i was looking for single rank.
But i am happy to find out all the time that i am wrong.

 

[RolloZ170]

Sorry to ask or note, but is dual ranked really better?
i thought that each extra rank is considered as a additional RDIMM for that channel and because

2 DIMM per channel are 2 loads
1 DIMM dual ranked is one load ( address lines, data lines are high impedance )

For many cpu they write
Up to DDR5 4800 MT/s 1DPC
Up to DDR5 4400 MT/s 2DPC

Up to DDR5 4400 MT/s 2DPC can be overruled by disabling "Enforce POR"

Is the extra rank from the dual rank considered equivalent as an extra DIMM or not?

no. dual ranked = one address puffer.
you buy a dual ranked 4800 and it runs at 4800.
the mem.controller can do rank interleaving, prepare next rank while current operation.

 

[scouzi]

you buy a dual ranked 4800 and it runs at 4800.
the mem.controller can do rank interleaving, prepare next rank while current operation.

My uderstanding too. DR is better because of interleaving. Even though only one rank can accessed at a time, the other can be used while one is busy.

 

[scouzi]

The interesting thing is that some D0 stepping cpu's like QYFS and some others support memory overclocking, while the official cpu versions do not. So those 4th gen cpu's could maybe still run at 5600MT with jedec 5600 memory, even at not overclocking board. I hope some one with such a cpu a GIGABYTE Board and Jedec 5600 memory can test this and confirm with a screenshot.

Do QYFP and/or QYFQ also support memory overclocking?

Here one analysis on Gen5. Should be close on Gen4 if you could boost it.

Intel 5th Gen Xeon Performance Benchmarks With DDR5-4800 vs. DDR5-5600 Review - Phoronix

www.phoronix.com