News Intel expects behemoth 144-core Sierra Forest chips to boost per-rack performance by 2.7X at lower power

[Admin]

Intel gives a brief idea what to expect from its Xeon 'Sierra Forest' processor.

Intel expects behemoth 144-core Sierra Forest chips to boost per-rack performance by 2.7X at lower power : Read more

 

[bit_user]

Headline: Intel expects behemoth 144-core Sierra Forest chips to boost per-rack performance by 2.7X at lower power

Wrong.

Try reading your own source, maybe??

"Intel Xeon next-gen processors, code-named Sierra Forest, launching later this year will offer up to 288 cores and help operators recognize a 2.7x performance per rack[1] improvement – an industry leading performance per rack for 5G Core workloads[2]."

Notes:

1. Based on estimated architectural projections as of Feb. 14, 2024, vs. prior-generation platforms as of 2021. Your results may vary.
2. Based on Intel analysis and publicly available data as of Feb. 20, 2024.

Source: https://www.intel.com/content/www/u...c-2024-next-gen-xeon-improvement-5g-core.html
​

This clearly suggests they're talking about the 288-core version delivering the 2.7 speedup! "Prior-generation as of 2021" means they're probably referring to Ice Lake.

That puts 106 cores at just below one 40-core (80-thread) Ice Lake, which is very plausible, since each E-core is said to have similar IPC to a single-threaded Skylake, while Ice Lake delivered real IPC gains relative to Skylake and could run at slightly higher frequencies than Sierra Forest. Not to mention hyper-threading.

In contrast, the notion that their 144-core CPU is 2.7x as fast as an 80-thread Ice Lake-SP, just isn't credible.

@JarredWaltonGPU can you check me on this?

 

[PaulAlcorn]

Wrong.

Try reading your own source, maybe??

​

"Intel Xeon next-gen processors, code-named Sierra Forest, launching later this year will offer up to 288 cores and help operators recognize a 2.7x performance per rack[1] improvement – an industry leading performance per rack for 5G Core workloads[2]."​

​

Notes:​

1. Based on estimated architectural projections as of Feb. 14, 2024, vs. prior-generation platforms as of 2021. Your results may vary.
2. Based on Intel analysis and publicly available data as of Feb. 20, 2024.

​

Source: https://www.intel.com/content/www/u...c-2024-next-gen-xeon-improvement-5g-core.html​

​

This clearly suggests they're talking about the 288-core version delivering the 2.7 speedup! "Prior-generation as of 2021" means they're probably referring to Ice Lake.

That puts 106 cores at just below one 40-core (80-thread) Ice Lake, which is very plausible, since each E-core is said to have similar IPC to a single-threaded Skylake, while Ice Lake delivered real IPC gains relative to Skylake and could run at slightly higher frequencies than Sierra Forest. Not to mention hyper-threading.

In contrast, the notion that their 144-core CPU is 2.7x as fast as an 80-thread Ice Lake-SP, just isn't credible.

@JarredWaltonGPU can you check me on this?

Thanks for the feedback, as always! The source doesn't say they used a 288-core, as noted in the text, and they didn't. Here is the response from Intel on the test setup:

"Here’s the claim: (2.7x improvement) This claim is comparing SRF 2S 144C vs widely deployed Telco config of 6252N 2S. So 144-core."

I added some text in the article to represent the additional info. We'll see if INtel will also clear up the workload info, too.

 

[TerryLaze]

In contrast, the notion that their 144-core CPU is 2.7x as fast as an 80-thread Ice Lake-SP, just isn't credible.

It's marketing but also true, the key word here is low(er) power, and we know that at very low power the e-cores will be a lot more efficient than the p-cores.
Especially since they are not saying "as fast" as you are saying, they say performance per rack at low power.

At least that's what I think.

 

[bit_user]

Thanks for the feedback, as always!

Thanks for checking with Intel. This is very helpful, due to the way they worded that sentence.

The source doesn't say they used a 288-core, as noted in the text, and they didn't. Here is the response from Intel on the test setup:

"Here’s the claim: (2.7x improvement) This claim is comparing SRF 2S 144C vs widely deployed Telco config of 6252N 2S. So 144-core."

Okay, so we should consider the match-up of 144-core Sierra Forrest vs. 24-core Cascade Lake. That would be six times as many cores and three times as many threads. So, a 2.7x speedup indeed seems plausible (since IPC is roughly similar and scaling is usually sub-linear).

https://www.intel.com/content/www/u...sor-35-75m-cache-2-30-ghz/specifications.html
​

I added some text in the article to represent the additional info. We'll see if INtel will also clear up the workload info, too.

Thanks! Two observations about that...

First, this part:

"... will offer up to 2.7X higher performance per rack ..."
​

Seems like it should be "... will offer up to 2.7X as much performance per rack ...".

Second, the sentence:

"Intel says that its choice of comparing two 24-core processors against two 24-core models is to highlight a widely deployed telco config."
​

Did you perhaps mean to say "... two 24-core processors against two 144-core models ..." ?

While I'm nit-picking:

"a 2.7 times performance increase over a three-year-old platform is an impressive improvement"
​

Cascade Lake began launching in Q2 of 2019. It's essentially the same microarchitecture as the Skylake-SP series, that launched on virtually the same process node, back in Q3 2017. So, I guess the real question is whether Sierra Forest is about 5 or 7 years newer than what they're comparing it to?
: D