ES Xeon Discussion

Notice: Page may contain affiliate links for which we may earn a small commission through services like Amazon Affiliates or Skimlinks.
S

SDletmk

New Member
Dec 30, 2023
9
0
1
RolloZ170 said:
they sell ES from intels city. i had NEVER issues with customs, i live in EU, DHL express customs are bandits,
they KNOW they are ES but even intel doesn't care.
apart from that you CAN BE OWNER of a ES/QS, it is intels decision.
the original customer of the ES signed a contract, if he sells the ES HE brakes the rule, the buyer is fine.
in fakt you can not buy a ES, you pay for the service to get the next USER of it.
Click to expand...
Yes, but I'm not going to try to argue that point to the person running the program that's building this.

Does anyone know if ES processors have the same functionality as consumer in terms of optimization? And if Ice Lake does support this optimization, but just to a lesser extent? Accelerating Stable Diffusion Inference on Intel CPUs (huggingface.co)

Also, Techpowerup claims that Bronze 3408U supports PCIE 5.0, which disagrees with the data from Intel's Ark page. Has anyone confirmed either way what this CPU supports? Its low power and cost would be ideal if it only worked with the 5.0 lanes.
Intel Xeon Bronze 3408U Specs | TechPowerUp CPU Database
Intel® Xeon® Bronze 3408U Processor

Basically, my choice flow at this rate is:
If Ice Lake supports Stable Diffusion acceleration, get 4189 consumer CPU and motherboard with many 4.0 x16 slots.
If not, does ES Sapphire Rapids support SD acceleration? If so, get a low-power, low cost Gold or Silver ES if a suitable motherboard can be found.
If not, does 3408U support PCIE 5.0? If so, then get consumer 3408U (or if there's no difference between it and an available ES 3408U) and lowest-cost MB with at least 3 x16 slots.

Also, please feel free to tell me to get off this thread and make my own topic regarding this if you feel I'm getting too off-track from the thread's main subject.
 
B

bayleyw

Active Member
Jan 8, 2014
315
103
43
SDletmk said:
Yes, but I'm not going to try to argue that point to the person running the program that's building this.

Does anyone know if ES processors have the same functionality as consumer in terms of optimization? And if Ice Lake does support this optimization, but just to a lesser extent? Accelerating Stable Diffusion Inference on Intel CPUs (huggingface.co)

Also, Techpowerup claims that Bronze 3408U supports PCIE 5.0, which disagrees with the data from Intel's Ark page. Has anyone confirmed either way what this CPU supports? Its low power and cost would be ideal if it only worked with the 5.0 lanes.
Intel Xeon Bronze 3408U Specs | TechPowerUp CPU Database
Intel® Xeon® Bronze 3408U Processor

Basically, my choice flow at this rate is:
If Ice Lake supports Stable Diffusion acceleration, get 4189 consumer CPU and motherboard with many 4.0 x16 slots.
If not, does ES Sapphire Rapids support SD acceleration? If so, get a low-power, low cost Gold or Silver ES if a suitable motherboard can be found.
If not, does 3408U support PCIE 5.0? If so, then get consumer 3408U (or if there's no difference between it and an available ES 3408U) and lowest-cost MB with at least 3 x16 slots.

Also, please feel free to tell me to get off this thread and make my own topic regarding this if you feel I'm getting too off-track from the thread's main subject.
Click to expand...
It sounds like you are asking for consulting help. What is your workload? Saying "I need a GPU host for work but my boss won't let me use engineering samples unless they ship from the US" doesn't really help us help you (it sounds like you are trying to generate images, but that should run on the GPUs - why are you looking for a CPU that can run Stable Diffusion?)
In fact, if your boss were smart he or she probably shouldn't go into production with engineering samples regardless of where they ship from; all of the unsorted bugs and errata which are of inconsequential impact in single user bare metal workstation really add up in a multi-user virtualized environment.
Anyway, assuming my assumptions are right and you are hosting SD1.5 or 2.1 on a bunch of GPUs just get a 4028GR and fill it with used blower 2080 Ti. You'll burn more power over a bare bones mining rig, but unlike the mining right your inference server will be supported and reliable.
 
R

RolloZ170

Well-Known Member
Apr 24, 2016
5,475
1,659
113
SDletmk said:
Does anyone know if ES processors have the same functionality as consumer in terms of optimization?
Click to expand...
if we are correct QS(Qualyfication Samples) are ES too but same than prod.unit(with same stepping)
there can be more than one prod.unit stepping, for each a matching QS.
with INTEL you can expect the same functionality of QS compared to prod.units.
 
S

SDletmk

New Member
Dec 30, 2023
9
0
1
bayleyw said:
It sounds like you are asking for consulting help. What is your workload? Saying "I need a GPU host for work but my boss won't let me use engineering samples unless they ship from the US" doesn't really help us help you (it sounds like you are trying to generate images, but that should run on the GPUs - why are you looking for a CPU that can run Stable Diffusion?)
In fact, if your boss were smart he or she probably shouldn't go into production with engineering samples regardless of where they ship from; all of the unsorted bugs and errata which are of inconsequential impact in single user bare metal workstation really add up in a multi-user virtualized environment.
Anyway, assuming my assumptions are right and you are hosting SD1.5 or 2.1 on a bunch of GPUs just get a 4028GR and fill it with used blower 2080 Ti. You'll burn more power over a bare bones mining rig, but unlike the mining right your inference server will be supported and reliable.
Click to expand...
Yes, the workload is generating images. I was requested to get the cheapest possible setup with multiple PCIE 4.0 or 5.0 lanes. As SD mainly runs on the GPU, all that was required of the CPU is to host PCIE lanes - which is why the Bronze 3408U was an option if it had worked with a PCIE 5.0 system. Although, it seems like PCIE 3.0 x16 might be good enough since all of the work is loaded onto the GPU, then loaded off, and the only bottlenecking that might occur even with PCIE 4.0 GPUs would be at those times. That 4028R might work if I can convince them it would do the job, so I'll look into it more.
Thank you.
 
B

bayleyw

Active Member
Jan 8, 2014
315
103
43
SDletmk said:
Yes, the workload is generating images. I was requested to get the cheapest possible setup with multiple PCIE 4.0 or 5.0 lanes. As SD mainly runs on the GPU, all that was required of the CPU is to host PCIE lanes - which is why the Bronze 3408U was an option if it had worked with a PCIE 5.0 system. Although, it seems like PCIE 3.0 x16 might be good enough since all of the work is loaded onto the GPU, then loaded off, and the only bottlenecking that might occur even with PCIE 4.0 GPUs would be at those times. That 4028R might work if I can convince them it would do the job, so I'll look into it more.
Thank you.
Click to expand...
Why do you need PCIe 4 lanes? The only thing that goes into the GPU is the prompt and the only thing that comes out is a 1024x1024 image, hardly high traffic.
 
M

myth0homelab

New Member
Nov 10, 2023
1
0
1
Does anyone know if
ASRock Rack SP2C741D16X-2T EEB Server
works with QYFQ CPUs?
Also, does anyone have a 2 socket motherboard that does work with QYFQ for sale?
 
S

scouzi

Member
Jan 8, 2024
39
7
8
RolloZ170 said:
with QS or stepping E/S ES(E0-E5,S0-S3) you can choose any motherboard.
but cheap SPR-SP D0 ES2 run only on Gigabyte (confirmed with latest BIOS R01) and ASUS W790.
Click to expand...
Can one run a Saphire Rapids Xeon E Socket on a 'W' Workstation MB? I'm eyeing a 8490H (KYFX D0) but the GIGABYTE MS33-AR0 PCIE slot design sucks for stacking long format Videocards such as ASUS W790? Downside I see from their KYFX D0 is that the base clock is 1.7 vs 1.9 GHz but I can live with that). Turbo boost is also a bit lower. I can live with a lower TDP and single core performance hit.
 
Last edited:
R

RolloZ170

Well-Known Member
Apr 24, 2016
5,475
1,659
113
scouzi said:
I'm eyeing a 8490H (KYFX D0) but the Gigabyte PCIE slot design sucks for stacking long format Videocards.
Click to expand...
stepping D0 SPR-SP (like QYFX,QYFQ,QYFP,QYFU,QYFV, and many more) are running on actual BIOS on ASUS W790 Ace and SAGE SE.
this is very wierd because stepping E0 and up are not working.
but SPR-SP have only 80 PCIe lanes, on the W790 SAGE SE you need Xeon W?34xx w. 112L to have all slots working,
with onyl 80lanes you will not have 2x M.2 from CPU. check the manual for that(Xeon W?24xx config)
 
S

scouzi

Member
Jan 8, 2024
39
7
8
Last edited:
R

RolloZ170

Well-Known Member
Apr 24, 2016
5,475
1,659
113
scouzi said:
Anyone know if we can run E sockets on ASUS Pro WS W790-ACE?
Click to expand...
EMR (Emerald Rapids) ??? needs a new BIOS to support EMR cpuid and microcode,
unless there is no EMR-WS this will not happen and even if i doubt that because
the reason of support of SPR-SP D0 stepping is absolutely unclear to me.
 
M

MillionMiles

Member
Aug 22, 2023
74
32
18
SPR-E0 looks good, their microcode is the same as the prod unit, does that mean they work on any c741? And get long-term microcode updates from intel? It seems to be expected that they will work similarly to the prod unit, like the ICX-D0

XCC SPR-D0 seems to have performance issue, especially in a dual socket configuration(GIGABYTE MS73), and it is unclear whether there has been any improvement in Windows 11

Dual D0 QYFS 112c CB-R23 76000
Dual E0 Q0KL 104c CB-R23 92000
 
Last edited:
R

RolloZ170

Well-Known Member
Apr 24, 2016
5,475
1,659
113
MillionMiles said:
SPR-E0 looks good, their microcode is the same as the prod unit, does that mean they work on any c741?
Click to expand...
they should work in theory....
E0 is one of the PRD cpuid and is currently shared over all E steppings.
but this can be ending if the E0 is removed from the MC container.
MillionMiles said:
XCC SPR-D0 seems to have performance issue, especially in a dual socket configuration(GIGABYTE MS73), and it is unclear whether there has been any improvement in Windows 11

Dual D0 QYFS 112c CB-R23 76000
Dual E0 Q0KL 104c CB-R23 92000
Click to expand...
have you made this yourself on same motherboard ?
can be just a misconfig. because single QYFS does 67000pts cbr23
 
Last edited:
R

RolloZ170

Well-Known Member
Apr 24, 2016
5,475
1,659
113
MillionMiles said:
I don't think this will change after the prod unit is released
Click to expand...
easy doable. the container supports
806F4 E0 can be removed, no issue at prod.units.
806F5 E2
806F6 E3
806F7 E4 final PRD(MCC)
806F8 E5 final PRD(XCC/MCC)
edit:
the container can be replaced by single microcodes, one for each cpuid.
and there is more than microcodes, each cpuid is suported by one or two(XCC/MCC) config ffs module.
 
Last edited:
You must log in or register to reply here.
Top Bottom