Intel Xeon Gold & Xeon Platinum (Skylake-SP) Lineup Leaked

Status
Not open for further replies.
Shintai

Shintai

Supreme [H]ardness
Joined
Jul 1, 2016
Messages
5,678
Its all base clocks obviously.

6134 (without AVX512?) with 3.2Ghz base clock for 18 cores at 130W is quite interesting.

upload_2017-3-23_17-4-37.png


Thanks to sweepr for the finding.
 
drescherjm said:
I am very happy that Intel now has 200W+ parts. I have been hoping for them to do that for so long. With the larger socket cooling should not be a serious issue.
Click to expand...

This list is already inaccurate...there were Sandra benches that were recently posted of a Xeon Platinum 8180 with a TDP of 246W...:)

They are shaping up to be impressive chips...I just hope they leave the top SKUs (the high TDP models) unlocked...;)
 
lutjens said:
This list is already inaccurate...there were Sandra benches that were recently posted of a Xeon Platinum 8180 with a TDP of 246W...:)

They are shaping up to be impressive chips...I just hope they leave the top SKUs (the high TDP models) unlocked...;)
Click to expand...

Sure, and a 6950X at 4.4Ghz also only use 140W :p
http://ranker.sisoftware.net/show_r...d4ecdce4d4e4d6f082bf8fa9cca994a482f1ccfc&l=en
http://ranker.sisoftware.net/show_r...d4ecdce9d9eed7f183be8ea8cda895a583f0cdfd&l=en

Same for 4.3Ghz.
http://ranker.sisoftware.net/show_r...d4ecdcebd2eadafc8eb383a5c0a598a88efdc0f0&l=en

Same for 4.2Ghz.
http://ranker.sisoftware.net/show_r...d4ecdcebdce4d1f785b888aecbae93a385f6cbfb&l=en

Sandra is quite useless for the power numbers.
 
Shintai said:
Sure, and a 6950X at 4.4Ghz also only use 140W :p
http://ranker.sisoftware.net/show_r...d4ecdce4d4e4d6f082bf8fa9cca994a482f1ccfc&l=en
http://ranker.sisoftware.net/show_r...d4ecdce9d9eed7f183be8ea8cda895a583f0cdfd&l=en

Same for 4.3Ghz.
http://ranker.sisoftware.net/show_r...d4ecdcebd2eadafc8eb383a5c0a598a88efdc0f0&l=en

Same for 4.2Ghz.
http://ranker.sisoftware.net/show_r...d4ecdcebdce4d1f785b888aecbae93a385f6cbfb&l=en

Sandra is quite useless for the power numbers.
Click to expand...

Sandra is reporting the rated TDP specification of the processor being tested, not the TDP of the processor when it's overclocked. As the Xeon Platinum 8180 processor in question isn't overclocked, the report from Sandra is entirely valid.

http://ranker.sisoftware.net/show_r...d5e3dbebdce8ddfb89b484a2c7a29faf89fac7f7&l=en
 
lutjens said:
Sandra is reporting the rated TDP specification of the processor being tested, not the TDP of the processor when it's overclocked. As the Xeon Platinum 8180 processor in question isn't overclocked, the report from Sandra is entirely valid.

http://ranker.sisoftware.net/show_r...d5e3dbebdce8ddfb89b484a2c7a29faf89fac7f7&l=en
Click to expand...

As far as I known there is no TDP information read from the CPUs. But its rather hardcoded/BIOS information. Also why you can find plenty of Sandra results without power numbers.

Example:
https://www.custompcreview.com/wp-c...chmarks-leaked-sisoft-sandra-arithmetic-1.jpg

So the numbers from Sandra is meaningless.
 
Last edited:
Shintai said:
As far as I known there is no TDP information read from the CPUs. But its rather hardcoded/BIOS information. Also why you can find plenty of Sandra results without power numbers.

Example:
https://www.custompcreview.com/wp-c...chmarks-leaked-sisoft-sandra-arithmetic-1.jpg

So the numbers from Sandra is meaningless.
Click to expand...

And where do you think a BIOS gets it's info regarding a CPU's TDP? The rated TDP is read from the CPU either by the BIOS or Sandra.

I also suppose the fact that the CPU in the link I posted is clocked higher (3.8GHz turbo vs 3.7GHz) is also meaningless too right? So to is the fact that Intel already produces parts up to 260W TDP for the Socket 3647 platform (Xeon Phi).:rolleyes:
 
The 6136 and 6154 are listed as 18C with the same 3GHz base clock. I assume that is an error in the listing for the 6136 which makes more sense. I mean 105W for 18C at 3GHz won't happen soon...
 
Last edited:
32C may only be a custom product if any. The regular one tops out at 28.
 
Are they just disabling 6 to improve yields/meet TDP? It goes 28 -> 22 -> 16 doesn't it?
 
evilpaul said:
Are they just disabling 6 to improve yields/meet TDP? It goes 28 -> 22 -> 16 doesn't it?
Click to expand...

There are at least 3 dies to begin with, lowest one with 10 cores.

Due to how software licensing works. Its often cheaper and better to have less fast cores than more slower cores. And some takes performance over cores at any cost (Financial transactions). SQL is a classic example of the licensing issue with between 3700 to 14000$ per core as list price. That's right, 14000 for enterprise SQL for a single core. If you got 28 cores that's 28x14000$.

Scaling of software isn't always what you think it is. More cores can actually give less performance besides the usual diminishing returns.

Example:
chris-hillman-beyond-mapreduce-scientific-data-processing-in-realtime-6-638.jpg
 
Last edited:
Last edited:
Considering the 6134 with 18 cores at 3.2Ghz is 130W. I think its save to assume that 3.6Ghz on 12 cores within 140W is quite doable. The question is how high the turbo will be with that headroom.
 
Shintai said:
Considering the 6134 with 18 cores at 3.2Ghz is 130W. I think its save to assume that 3.6Ghz on 12 cores within 140W is quite doable. The question is how high the turbo will be with that headroom.
Click to expand...

The 6130 with 16 cores is 2.1/3.4/3.7 GHz (base/all-core/single-core) on 120 W. 4GHz turbo for the 12C is a possibility.
 
I'd like to diverge this a bit.

Skylake-x is also coming out ASAP.....compare and contrast ?

I've got money in the bank for an upgrade , I WAS looking at 2687w v4's (dually), but this thread seems to say 'eff that.
If top bin skylake-x is 12 core, how does that compare to the skylake-sp 12 core ? same chip ??
That 18 core looks really sexy.

Any news on motherboards for these ?
 
Shintai said:
There are at least 3 dies to begin with, lowest one with 10 cores.

Due to how software licensing works. Its often cheaper and better to have less fast cores than more slower cores. And some takes performance over cores at any cost (Financial transactions). SQL is a classic example of the licensing issue with between 3700 to 14000$ per core as list price. That's right, 14000 for enterprise SQL for a single core. If you got 28 cores that's 28x14000$.

Scaling of software isn't always what you think it is. More cores can actually give less performance besides the usual diminishing returns.

Example:
chris-hillman-beyond-mapreduce-scientific-data-processing-in-realtime-6-638.jpg
Click to expand...

Our SQL setup is highly multi threaded so having 32C/64T makes sense for where I work.

Your statement on SQL enterprise licensing I think is a bit off. If I remember correctly the license cost is per 4 cores not one. You have to dig a bit to find that but considering the side of our systems I was VERY concerned with this value.

Oh and it is NOT a value proposition to run MSSQL Enterprise in any kind of VM infrastructure. You may have to license all cores on the largest host in a cluster if not on ALL HOSTS in a cluster. And THAT is a RIPOFF.
 
Well its halfway between you and me. I got the public one wrong.
https://www.microsoft.com/en-us/sql-server/sql-server-2016-pricing

You can only buy in stacks of 2 cores at a time and that's what the price is.

https://www.cdw.com/shop/products/M...16-Standard-Core-license-2-cores/4149038.aspx
https://www.cdw.com/shop/products/M...ise-Core-license-2-cores/4149037.aspx?pfm=srh

So for your 32 cores and no leasing just plain upfront cost without discount. Its roughly 60K$ to get it on 32 cores for the cheap standard edition. And ~220K$ for enterprise.

But fully agree, SQL is where you bring lube, lots of lube. Oracle for example still wants 20% of the entire cost in mandatory support too. Everyone knows its big business.
 
Last edited:
Interesting data.

On the Linux kernel compilation test, it took the two-socket Epyc machine 15.7 seconds to do the compile, compared to 22.5 seconds on the Xeon E5 machine using the top-bin part. That is 30 percent better performance for the AMD system, and it also implies that on this kind of work, the Broadwell core does about 11.7 percent more work than a Zen core at their respective clock speeds.
Click to expand...

Thus, SKL-X core will be about 20--25% faster clock-for-clock on non-AVX workloads. Therefore the Xeon 8110 [28 SXL cores @2.5GHz] will be about 30% faster than Naples [32 Zen cores @1.9GHz] with similar wattages 205W vs >200W.
 
juanrga said:
Interesting data.



Thus, SKL-X core will be about 20--25% faster clock-for-clock on non-AVX workloads. Therefore the Xeon 8110 [28 SXL cores @2.5GHz] will be about 30% faster than Naples [32 Zen cores @1.9GHz] with similar wattages 205W vs >200W.
Click to expand...

Naples is nowhere near SKL-SP, it cant even remote touch Broadwell-EP. But we already knew that, because outside the slideware BS and rigged demos they dont expect to make any revenue increase as such.
 
Status
Not open for further replies.
Back
Top