Discussion Intel current and future Lakes & Rapids thread

[moinmoin]

More like fab at TSMC. It wouldn't fix all the problems but it'd be better at least.

You know all the back and forth with delays has me completely confused by now whether the design (be it silicon or packaging) or the node is mainly to blame. So you're sure it's the node?

Publicly available IFS nodes that Intel also uses internally can't come soon enough in any case.

 

[Geddagod]

86K points at 800W is in line with what desktop Raptor lake CPUs do - 40K points at 300W.

The only problem here is the manufacturing process.

Which Intel SPR does 86k points at 800W?
Plus RPL has E-cores, the 13900k is clocked to the sky (very inefficient), etc etc
I don't think we can compare like that

 

[Geddagod]

I am sure a better MFG process would help, but thats not the only problem. Look at Genoa vs Milan !

What?

 

[Geddagod]

It's Golden Cove. It's a big core that is not meant for Servers. More like desktop. Zen 4 is smaller than Golden Cove.

Core size, AFAIK, mostly just hurts yield. I guess since there is more physical silicon it might end up requiring more power, but this should also be counteracted by the fact that the architecture itself is wider. GLC, the design itself, is fine imo

 

[Kocicak]

Which Intel SPR does 86k points at 800W?
Plus RPL has E-cores, the 13900k is clocked to the sky (very inefficient), etc etc
I don't think we can compare like that

w9-3495 supposedly.

Ok, I might have said "broadly in line". 13900K is way in the inneficiency region, but the little E cores help with MT performance and it also does not have features that the server part has. All these and possibly others factors may cancel each other and what I wrote may be correct by chance.

 

[nicalandia]

Which Intel SPR does 86k points at 800W?

The W9 3495X, the OC Team Ran a all core OC of 3.5 Ghz and consumed 800 watts and yielded a CBR23 score of 86,000 points

 

[Edrick]

I really do love these forums. People jumping all over SPR (Xeon) because of a lower than expected Cinebench score (not by much) and a higher than expected power draw on said benchmark.

- SPR has up to 8 memory channels for memory intensive applications (not measured in Cinebench)
- SPR has up to 112 PCIe lanes for massive GPU connectivity and/or storage connectivity (not measured in Cinebench)
- SPR has a ton of AVX512 instructions for specialized purposes (not measured in Cinebench)
- SPR has new AMX extensions for AI/ML applications (not measured in Cinebench)
- SPR has a CXL technology, which looks very promising. Looking forward to how this plays out. (again, not measured in Cinebench)

SPR a big step forward for Intel, even though they are almost 2 years late to the party. I am not saying SPR is great by any means, but I do believe these CPUs will perform there intended role well, with Emerald and Granite Rapids being more fine tuned.

 

[nicalandia]

I really do love these forums. People jumping all over SPR (Xeon) because of a lower than expected Cinebench score (not by much) and a higher than expected power draw on said benchmark.

You would expect that at the very least the new Workstation HEDT Xeon would trash Old Rome based Threadrippers on regular MT workloads like CBR23, V-Ray, Blender and Geekbench5 But guess what? They are Dead Even in MT performance, so that is very bad news for Emerald Rapids which is not much of an update and really set to match Milan performance.

 

[jpiniero]

Hey, Dell is actually selling an SPR server! Two even. Both are only 2S XCC models but it does include the 56 core 8480+ as a (pricey) option. Cheaper one ships as soon as March 28th.

 

[nicalandia]

Hey, Dell is actually selling an SPR server! Two even. Both are only 2S XCC models but it does include the 56 core 8480+ as a (pricey) option. Cheaper one ships as soon as March 28th.

A single 9554 with 64 cores perform better thant two of those and cost Half

 

[jpiniero]

A single 9554 with 64 cores perform better thant two of those and cost Half

Probally. They want (at least on the site) an extra $13K per processor for the 8480+ over the 32 core.

 

[Markfw]

I really do love these forums. People jumping all over SPR (Xeon) because of a lower than expected Cinebench score (not by much) and a higher than expected power draw on said benchmark.

- SPR has up to 8 memory channels for memory intensive applications (not measured in Cinebench)
- SPR has up to 112 PCIe lanes for massive GPU connectivity and/or storage connectivity (not measured in Cinebench)
- SPR has a ton of AVX512 instructions for specialized purposes (not measured in Cinebench)
- SPR has new AMX extensions for AI/ML applications (not measured in Cinebench)
- SPR has a CXL technology, which looks very promising. Looking forward to how this plays out. (again, not measured in Cinebench)

SPR a big step forward for Intel, even though they are almost 2 years late to the party. I am not saying SPR is great by any means, but I do believe these CPUs will perform there intended role well, with Emerald and Granite Rapids being more fine tuned.

Their competition is Genoa, and its 96 cores will decimate SPR at a fraction of the power usage per core. And you can buy them today for $6788.

That is a serious problem, and what everyone here is talking about. And Genoa has avx-512. And 128 PCIE 5.0 lanes. Genoa has 12 channels of memory. SPR is so short in performance and features, its virtually DOA.

 

[Edrick]

Their competition is Genoa, and its 96 cores will decimate SPR at a fraction of the power usage per core. And you can buy them today for $6788.

That is a serious problem, and what everyone here is talking about. And Genoa has avx-512. And 128 PCIE 5.0 lanes. Genoa has 12 channels of memory. SPR is so short in performance and features, its virtually DOA.

But its not really. We all know SPR is 2 years late due the 10nm disaster at Intel. It was designed (at the time) to compete with Rome/Milan. Intel never claimed that SPR would compete with Genoa, they (and we) know that's a lost battle. Everyone knew this before SPR launched and yet people still seem surprised. Hell, even Granite Rapids next year (maybe) will have trouble matching Genoa. Diamond Rapids is Intel's first true hope to catch up to AMD.

I was not arguing that Intel was finally at AMD levels, it has not. I was mostly arguing that Cinebench was not the best benchmark to rate CPUs in this niche market (Genoa included). I am holding off my final judgement until I see test cases based on AVX512/AMX extensions.

 

[nicalandia]

I was not arguing that Intel was finally at AMD levels, it has not. I was mostly arguing that Cinebench was not the best benchmark to rate CPUs in this niche market (Genoa included). I am holding off my final judgement until I see test cases based on AVX512/AMX extensions.

The Xeon w9-3495X is also either a match or slower than the Threadripper Pro 3995WX on Geekbench5, Blender and V-Ray

 

[Edrick]

The Xeon w9-3495X is also either a match or slower than the Threadripper Pro 3995WX on Geekbench5, Blender and V-Ray

I am going to pretend you didn't say Geekbench, but do you have a source for the Blender and V-Ray comparison? I honestly have not seen that yet.

 

[Markfw]

But its not really. We all know SPR is 2 years late due the 10nm disaster at Intel. It was designed (at the time) to compete with Rome/Milan. Intel never claimed that SBR would compete with Genoa, they (and we) know that's a lost battle. Everyone knew this before SBR launched and yet people still seem surprised. Hell, even Granite Rapids next year (maybe) will have trouble matching Genoa. Diamond Rapids is Intel's first true hope to catch up to AMD.

I was not arguing that Intel was finally at AMD levels, it has not. I was mostly arguing that Cinebench was not the best benchmark to rate CPUs in this niche market (Genoa included). I am holding off my final judgement until I see test cases based on AVX512/AMX extensions.

OK, well, I guess you are saying that a site like Phoronix has the types of benchmarks for these server CPUs, but I can't find one for the 3495x. All I can find out there is cinebench. It does give a good indicator of a lot of areas, and a good idea of performance and efficiency. And while many would argue with me, taking a 12700k and disabling the e-cores (version before they disabled avx-512), gives a good idea of what the 3495x would do, if you multiply power and compute power times 7 (=56 cores). And it would say exactly what we are seeing, horrible efficiency and huge power usage. I personally had one, so I know of what I speak.

 

[Exist50]

I personally had one, so I know of what I speak.

Lmao, are we going to pretend you didn't make ridiculous false claims about it? Yes, SPR is uncompetitive, but waving around Cinebench and "trust me bro" isn't exactly a compelling way to demonstrate that.

 

[igor_kavinski]

I personally had one, so I know of what I speak.

Went to family member?

 

[Markfw]

Went to family member?

good friends son. It suited him, as he needed a general usage PC, hardly any gaming, no heavy use and not powered on all the time.

 

[nicalandia]

I am going to pretend you didn't say Geekbench, but do you have a source for the Blender and V-Ray comparison? I honestly have not seen that yet.

Even if you dont like it Geekbench5 is actually a pretty good test and needs to be taken into account, here the Threadripper is very competitive

HP HP Z8 Fury G5 Workstation vs ASUS System Product Name - Geekbench

browser.geekbench.com

Also, PugetSystem did not include the Threadripper Pro 3995WX on their Xeon W9 benchmarks, luckely we have numbers from previous Reviews made by them, those Threadripper are air cooled, stock Threadripper on a waterloop will be even more impressive.

To the MT numbers.

In Blender the 3995WX gets 1093 points, the W9 3495X gets 1078, so the 3995WX has 1.4% advantage

In V-Ray the 3995WX gets 48,300 points and the 3495X gets 50,800 so thats a 5% advantage that goes to the 3495X.


For the "Mid-Range" HEDT Pay attention to the 32 core 3975WX, its also a match for the 28 core 3465X

AMD Threadripper PRO 5000 WX-Series Content Creation Review

AMD's new Threadripper PRO 5000 WX series of CPUs are here, providing greater performance over the previous generation while maintaining the large memory capacity and high PCIe lane count that Threadripper Pro is known for. But just how much faster are these new processors in content creation...

www.pugetsystems.com

I wish PugetSystem would have included the 3000 series Threadripper Pro on their tests, many are on that Generation of HEDT CPUs and Would like to know if its okay to upgrade now or wait for Zen4 based Threadripper.

 

[DrMrLordX]

What is the drama all about? They just built a CPU from the cores they have, on the process they have with expected power to performance ratio, with added features compared to desktop models as memory support and more PCIe lanes, and are getting expected performance.

That is not the expected performance.

More like fab at TSMC. It wouldn't fix all the problems but it'd be better at least.

They could do that, but once they go down that road, their role as an IDM is finished, and the relevance of Intel may wane considerably.

I really do love these forums. People jumping all over SPR (Xeon) because of a lower than expected Cinebench score (not by much) and a higher than expected power draw on said benchmark.

- SPR has up to 8 memory channels for memory intensive applications (not measured in Cinebench)
- SPR has up to 112 PCIe lanes for massive GPU connectivity and/or storage connectivity (not measured in Cinebench)
- SPR has a ton of AVX512 instructions for specialized purposes (not measured in Cinebench)
- SPR has new AMX extensions for AI/ML applications (not measured in Cinebench)
- SPR has a CXL technology, which looks very promising. Looking forward to how this plays out. (again, not measured in Cinebench)

None of that explains the ridiculous power draw or low performance. Which is especially embarassing for Intel since Cinebench uses Embree.

 

[eek2121]

86K points at 800W is in line with what desktop Raptor lake CPUs do - 40K points at 300W.

The only problem here is the manufacturing process.

Correct.

Yes. The question is, why Intel decided to use this inefficient process to make these CPUs, are there technical reasons (this design is for example impossible to make on any TSMC process), financial (it would be too expensive to use any other process than their own), strategic (perhaps Intel does not even want or need to sell a lot of these CPUs, because they know the next gen will be much more competitive), purely image building exercise so that they can present their workstation lineup to somebody, contractual (they are oblidged by some contract to make and deliver such CPUs), or what else there could be?

Intel using TSMC for their core (little c, not the product) IP would be a disaster. Investors would be clamoring for them to ditch IFS. Press headlines would be similarly dooming.

On the desktop they just happened to roll with a configuration that is competitive with AMD. They still have a bit of time to play the Intel name in the server market. Laptop will be a coin toss.

That is not the expected performance.



They could do that, but once they go down that road, their role as an IDM is finished, and the relevance of Intel may wane considerably.



None of that explains the ridiculous power draw or low performance. Which is especially embarassing for Intel since Cinebench uses Embree.

I kind of expected the performance. Power usage is higher than I expected.

 

[semiman]

I really do love these forums. People jumping all over SPR (Xeon) because of a lower than expected Cinebench score (not by much) and a higher than expected power draw on said benchmark.

- SPR has up to 8 memory channels for memory intensive applications (not measured in Cinebench)
- SPR has up to 112 PCIe lanes for massive GPU connectivity and/or storage connectivity (not measured in Cinebench)
- SPR has a ton of AVX512 instructions for specialized purposes (not measured in Cinebench)
- SPR has new AMX extensions for AI/ML applications (not measured in Cinebench)
- SPR has a CXL technology, which looks very promising. Looking forward to how this plays out. (again, not measured in Cinebench)

SPR a big step forward for Intel, even though they are almost 2 years late to the party. I am not saying SPR is great by any means, but I do believe these CPUs will perform there intended role well, with Emerald and Granite Rapids being more fine tuned.

Even if we put all these things into account, 3495X is still a huge disappointment. Their workstation counterpart, 5995WX also has 8 channels MC, and 128 PCIe lanes. 3495X was supposed to win 5995WX in legacy programs and roflstomp 5995WX in simulation and AI applications. But they're even losing ST performance!
SPR is obviously suffering from problems which don't exist in Alder lake and its descendants. If we look into Raptor lake vs Zen 4 comparison, then high core count + big cores should be much better than this.

 

[Accord99]

SPR is obviously suffering from problems which don't exist in Alder lake and its descendants. If we look into Raptor lake vs Zen 4 comparison, then high core count + big cores should be much better than this.

Stock voltage is very high, Raptor Lake's P-cores doesn't need 1.2v to hit 5 GHz, and can do 4 GHz at 0.95v or less. It'll be interesting to see what undervolting does, SkatterBencher's 24-core 2495X uses 1.15v at 4.4 GHz:

https://twitter.com/x/status/1632990592908263426

 

[Exist50]

Intel using TSMC for their core (little c, not the product) IP would be a disaster. Investors would be clamoring for them to ditch IFS. Press headlines would be similarly dooming.

Seems like they're already set to do that with Arrow Lake and Lunar Lake.