Intel Ice Lake-SP Processor Spotted with 36 Cores and 3.6 GHz Base Clock

[AleksandarK]

Today, in the latest GeekBench 5 submission by ASUS, we have discovered something rather interesting. Intel's Ice Lake-SP processors were rumored to arrive with up to 28 cores and 56 threads at maximum, on a single chip. That was due to the 10 nm process used to make these chips, with suspicions that the yield of the node was not good enough to make any higher core count parts. Thanks to the GB5 listing, discovered by Leakbench on Twitter, the Intel Ice Lake-SP CPU engineering sample appeared with an amazing 36 cores with 72 threads. This is supposedly Intel's efforts to try and match the 64 cores and 128 threads of AMD's EPYC "Rome" CPUs, which are winning many server applications due to their performance.

The 36C/72T chip was paired with another similar chip in a 2P dual-socket configuration, which made the total core count rise to 72 cores and 144 threads, running inside of Asustek's Y4R-A1-ASUS-G1 server. The system was reporting a clock frequency of 3.6 GHz base speed, which means that the possible boost clocks could be higher. The CPU features a 1.25 MB level two (L2) cache per core (45 MB in total) and 54 MB of unified level three (L3) cache. That makes this CPU core quite an improvement over the past Cooper Lake generation. We are waiting for more information about these CPUs, and we are going to report on it in the coming time.


View at TechPowerUp Main Site

 

[phill]

I wonder what the TDP or power usage is for a few of these... Be interesting to find out to compare to AMD's Epyc....

 

[Caring1]

They might not even have a boost clock.

 

[Uroshi]

The evil gremlin in me is thinking 2x18 ... like ... we can't make those 28 and even less those 28x2 CPUs

 

[londiste]

The evil gremlin in me is thinking 2x18 ... like ... we can't make those 28 and even less those 28x2 CPUs

This is the likely scenario.
Whether they can make these or not is a bit irrelevant as smaller dies should be more cost effective in any case, especially on a cutting edge node.

 

[DeathtoGnomes]

The cost of using 2 of these to match up against 1 EPYC still needs to be addressed. Then there is the cost per unit, Intel just cant do it and profit enough.

 

[Vya Domus]

with an amazing 36 cores

What's amazing about that ?

 

[InVasMani]

Wonder if we'll start to see any 9c chips out of Intel/AMD anytime in the relatively near future they make a lot of sense for that performance and price gap between 8c and 12c/16c parts. I really think that could be a good Zen 3 ThreadRipper chip and could be similar to a 5600X for TR's higher TDP HDET segment. Chances are reasonable for TR AMD could do a single chip 9c part for Zen 3 that would perform great for gaming while offering all the excellent bandwidth advantages of quad channel. In fact with the BAR size it might made quad channel a lot more intriguing as well with the GPU having access to it. It would be pretty awesome for a APU as well especially true on that platform.

 

[birdie]

ASUSTeK COMPUTER INC. Y4R-A1-ASUS-G1 - Geekbench Browser

Benchmark results for an ASUSTeK COMPUTER INC. Y4R-A1-ASUS-G1 with an Intel $0000% processor.

browser.geekbench.com

Running at 2.2GHz, no boost.

This is either an ES or a bust.

 

[cosminmcm]

What's amazing about that ?

The fact that it is monolithic.

 

[Vya Domus]

The fact that it is monolithic.

So years of effort to make 10nm working got the them 8 more cores ? This better be a lower end SKU because that's terrible.

 

[MikeMurphy]

So years of effort to make 10nm working got the them 8 more cores ? This better be a lower end SKU because that's terrible.

A 28.5% increase in core-count on a monolithic die using a troubled 10nm node is an engineering feat.

It's not competitive against AMD's lower cost chiplet designs but it's purpose is really just to slow adoption of AMD products until Intel is in a position to launch something more competitive.

 

[Tom Yum]

The fact that it is monolithic.

How do you know it is monolithic? It could be a MCM made of two 18C/36T dies. The article doesn't really say one way or the other.

 

[efikkan]

It's not competitive against AMD's lower cost chiplet designs but it's purpose is really just to slow adoption of AMD products until Intel is in a position to launch something more competitive.

This is a typical mentality for consumers who don't understand servers.
Servers are built for a specific purpose in mind. Cinebench and Geekbench scores or average benchmark scores are completely irrelevant. The most important factor is the performance in the specific task the server will be running, along with various other constraints. So a 36 core Xeon could be superior in some workloads compared to a 64 core Epyc, and the completely opposite in others.
And lastly, the pricing of server parts is very fluid, so don't emphasize list price too much.

 

[yotano211]

ASUSTeK COMPUTER INC. Y4R-A1-ASUS-G1 - Geekbench Browser

Benchmark results for an ASUSTeK COMPUTER INC. Y4R-A1-ASUS-G1 with an Intel $0000% processor.

browser.geekbench.com

Running at 2.2GHz, no boost.

This is either an ES or a bust.

Of couse its a ES chip, they are not out yet.

 

[MikeMurphy]

This is a typical mentality for consumers who don't understand servers.
Servers are built for a specific purpose in mind. Cinebench and Geekbench scores or average benchmark scores are completely irrelevant. The most important factor is the performance in the specific task the server will be running, along with various other constraints. So a 36 core Xeon could be superior in some workloads compared to a 64 core Epyc, and the completely opposite in others.
And lastly, the pricing of server parts is very fluid, so don't emphasize list price too much.

My point was that on cost AMD's chiplet design is far less expensive to produce large core-count packages than the Intel monolithic monsters. Adjusting for price AMD is in a far better position.

I agree that Intel has some performance advantages but the AMD production cost advantage is significant.

 

[InVasMani]

One distinction difference is AMD has to pay TSMC to manufacture it's chips while Intel does it's own which is obviously in Intel's favor on the cost aspect of producing a chip itself. If Intel catches up to TSMC on that end it'll improve things a lot for them on the cost of business side. AMD does had a advantage in terms of not being a large and monolithic though for yields.

 

[Deleted member 24505]

One distinction difference is AMD has to pay TSMC to manufacture it's chips while Intel does it's own which is obviously in Intel's favor on the cost aspect of producing a chip itself. If Intel catches up to TSMC on that end it'll improve things a lot for them on the cost of business side. AMD does had a advantage in terms of not being a large and monolithic though for yields.

Many more advantages to Fabing your own chips than cost too.