THEY'RE FAKE! 144 and 288 E-Core only Intel Sierra Forest CPUs coming next year with 172MB of cache

[filpo]

Summary

 

Quotes

Quote

Intel has given us a lot of details on its Sierra Forest Xeon CPUs which will be packing the Sierra Glen E-Core. This E-Core is an optimized version of the Crestmont core that will be incorporated within Meteor Lake & Arrow Lake client CPUs. The CPUs will come in 144 SP and 288 AP configs.

 

Now, we have the first appearance of the 144-core variant showing up within the Geekbench 6 benchmark database. This particular chip was running on the Beechnut City platform which is a reference evaluation platform featuring a 2S (Dual-Socket) design and supports 32 DDR5 DIMMs. The platform comes with the LGA 4677 socket, offering support for up to 350W CPU TPDs and up to 88 PCIe Gen5 lanes. Sierra Forest is expected to feature TDPs as low as 200W and will come in both 1S and 2S servers.

 

Quote

 

Quote

The particular Intel Sierra Forest Xeon CPU features 144 E-Cores, 144 threads, a base clock of 2.20 GHz, and lots of cache. There's 108 MB of L3, 64 MB of L2, 4 MB of L1 Data & 9 MB of L1 Instruction cache. Since this is a 2S platform we are looking at, two chips were running which is why 288 cores are mentioned instead of 144 but the cache count makes it clear that each chip was a 144 E-Core SKU and not the top 288 core SKU which should be expected a bit later. Additionally, the CPU was running alongside a total of 256 GB of DDR5 memory.

 

Quote

These higher core count E-Core CPUs are aiming at the cloud data centers where AMD has positioned its Zen 4C-powered Bergamo lineup. The Bergamo lineup shares the same ISA as the Genoa CPUs & offers some impressive compute and efficiency. Intel will have +12.5% higher cores vs Genoa's 128 cores but since the architecture is different, there aren't any available threads. This gives Bergamo a major lead with 256 threads versus Sierra Forest's 144-thread design. AMD EPYC CPUs also offer much higher cache than Sierra Forest with up to 384 MB pools on the top Bergamo SKU.

 

Based on the current Xeon roadmap, Intel's 5th Gen Emerald Rapids CPUs will be launching next week on the 14th of December followed by Sierra Forest in the first half of 2024 and these will soon be followed by the P-Core powered Granite Rapids family.

Performance

Quote

As for the performance, it looks like this particular Intel Xeon Sierra Forest CPU platform was an early ES design based on single and multi-core performance. It is rather low and while the E-Core architecture should not be compared to the performance-oriented P-Core designs, we still expect 144 E-Cores to do much better than the results posted here.

 

Quote

My thoughts

Was excited until I saw that it said E Cores, at least there's a lot of cache and could be useful in some applications

 

Sources

Intel 144 Core Sierra Forest Xeon CPU Leaks Out: Powered By Sierra Glen E-Core Architecture, 172 MB Cache (wccftech.com)

Edited December 7, 2023 by filpo

[YoungBlade]

I would think 288 E-cores would beat 128 Zen 4C cores pretty easily, and that would give Intel back the multi-core performance crown. Whether or not they beat Zen 5C, though, will remain to be seen.

 

If Intel can actually make these at scale, and they can get them out early next year, it could actually be a big deal in the server space.

 

My concern is that by the time Intel brings these to market in any real capacity, Zen 5C might already be out. Intel has had a bad habit recently of pushing back launches repeatedly.

[jagdtigger]

2 minutes ago, YoungBlade said:

I would think 288 E-cores would beat 128 Zen 4C cores pretty easily

It gets beaten by a 1950X. Or a more recent example inside my nas a 5950X, gory.......

[YoungBlade]

1 minute ago, jagdtigger said:

It gets beaten by a 1950X. Or a more recent example inside my nas a 5950X, gory.......

In Geekbench, which is not meant to emulate a workload that can take advantage of 288 cores. You'll notice that in Geekbench's rankings, lower core count CPUs regularly best higher core count ones. This is because the benchmark isn't taking full advantage of all of the cores.

Spoiler

So if your intention is to use one of these CPUs for everyday computing or gaming, then yes, it'll be bad. The question is how well it can work when all of its cores are actually being used in server or rendering workloads.

[jagdtigger]

2 minutes ago, YoungBlade said:

n Geekbench, which is not meant to emulate a workload that can take advantage of 288 cores. You'll notice that in Geekbench's rankings, lower core count CPUs regularly best higher core count ones. This is because the benchmark isn't taking full advantage of all of the cores.

Look even if thats true we can still somewhat safely interpolate from the single core perf, and seeing how the epyc 9654 has  more than double of the performance i cant see any good reason to opt for these intel chips...

[GOTSpectrum]

I think price will be as interesting as performance... 

 

Will/can Intel match or beat AMDs pricing while using what are still relatively large chips? I know the target market isn't as price sensitive as us consumers but it will still be interesting to see 

[YoungBlade]

2 minutes ago, jagdtigger said:

Look even if thats true we can still somewhat safely interpolate from the single core perf, and seeing how the epyc 9654 has  more than double of the performance i cant see any good reason to opt for these intel chips...

Because SMT doesn't double CPU core performance and because 288 threads > 256 threads. So while a single core using a single thread is more than twice as fast, that same core having to juggle two threads will be slower than two cores each working on just one thread. In applications that are very good at using SMT, like Cinebench, the boost is only about 50%, meaning that the boost from a single core to all the cores on a 128 core EPYC CPU is only going to be about 192x, whereas all 288 of Intel's cores are working on just a single thread.

 

The EPYC 9654 is also based on Zen 4, not Zen 4C, which is what these CPUs are meant to compete against. Zen 4C has lower single core performance.

[8tg]

If these are accessible in some form on a consumer level I am buying one, ever since getting an N100 I wondered how just a shitload of ecores would handle some stuff. Behold, a shitload of ecores.

For reference the N100 gracemont cores are like haswell i5 cores, the 4 core configuration holds up nearly identical to an i5 4570.

144 of them would be 36 i5 4570’s, which is a dumb metric but I used to run a 64 player squad server on a 4570 with no issues. So a single chip could run 36 of them at extremely low power.

An that’s just using gracemonts performance as an example, this new sierra forest stuff should be faster.

 

I hope I can just buy one, that’s the hard part with any of this niche server stuff, is just getting your hands on it as a hobbyist.

[jagdtigger]

14 minutes ago, YoungBlade said:

So while a single core using a single thread is more than twice as fast, that same core having to juggle two threads will be slower than two cores each working on just one thread.

Doubt it, but we will see when independents get their hands on this. Ill put this debate on ice until then.....

[starsmine]

30 minutes ago, YoungBlade said:

Zen 4C has lower single core performance.

So long as the workload fits inside 4C cache, the IPC is exactly the same

[YoungBlade]

Just now, starsmine said:

So long as the workload fits inside 4C cache, the IPC is exactly the same

Yes, the IPC is the same, but they are clocked lower.

[Levent]

Why does that geekbench score look alarmingly low? lol

 

reference i3 12300 lol.

https://browser.geekbench.com/processors/intel-core-i3-12300

 

[porina]

28 minutes ago, YoungBlade said:

In applications that are very good at using SMT, like Cinebench, the boost is only about 50%

You're being generous. Cinebench R15 and R20 I've seen ~30% scaling with SMT/HT. R23 seemed to perform similarly to R20 but I didn't test it in depth, and haven't repeated this on 2024 yet.

 

I've only seen 50% uplift twice that I can remember. One was an old distributed computing project whose name I can't even remember any more. Something to do with simulating biodiversity or something like that. I ran that when P4 was current. That long ago. The other is the Ryzen Blender benchmark AMD made when they released Ryzen. However at that time Blender didn't support AVX which was added not long after, giving massive speedups all round making the older benchmark moot.

[igormp]

1 hour ago, YoungBlade said:

I would think 288 E-cores would beat 128 Zen 4C cores pretty easily, and that would give Intel back the multi-core performance crown. Whether or not they beat Zen 5C, though, will remain to be seen.

 

If Intel can actually make these at scale, and they can get them out early next year, it could actually be a big deal in the server space.

 

My concern is that by the time Intel brings these to market in any real capacity, Zen 5C might already be out. Intel has had a bad habit recently of pushing back launches repeatedly.

If AMD improves their routing, a 192 core bergamo should be totally doable.

 

Also, those cores are often used for webserver hosting, SMT does really well with those since EU occupancy per process is really low.

[porina]

8 minutes ago, YoungBlade said:

Yes, the IPC is the same, but they are clocked lower.

On this point, has there been any believable information on c core clocks at all? We know they will be lower, but I'm not aware of anyone saying how much lower.

[filpo]

35 minutes ago, Levent said:

Why does that geekbench score look alarmingly low? lol

 

reference i3 12300 lol.

https://browser.geekbench.com/processors/intel-core-i3-12300

 

Well they are efficiency cores, all of them. They have a nice amount of cache tho. Hopefully power consumption should be low/er (than usual)

[igormp]

42 minutes ago, porina said:

On this point, has there been any believable information on c core clocks at all? We know they will be lower, but I'm not aware of anyone saying how much lower.

https://www.phoronix.com/review/amd-epyc-9754-bergamo/9

Quote

The EPYC 9754 also has a lower base clock at 2.25GHz compared to 2.4GHz with the EPYC 9654 and a maximum boost clock of 3.1GHz compared to 3.7GHz with the EPYC 9654.

I guess we can assume the clock difference to be the same compared to consumer zen 4 vs 4c

[porina]

16 minutes ago, igormp said:

https://www.phoronix.com/review/amd-epyc-9754-bergamo/9

I guess we can assume the clock difference to be the same compared to consumer zen 4 vs 4c

Thanks. I haven't kept up to date on server CPUs so didn't know this was out already. That should be enough info for me to do some estimations on.

[RejZoR]

2 hours ago, YoungBlade said:

I would think 288 E-cores would beat 128 Zen 4C cores pretty easily, and that would give Intel back the multi-core performance crown. Whether or not they beat Zen 5C, though, will remain to be seen.

 

If Intel can actually make these at scale, and they can get them out early next year, it could actually be a big deal in the server space.

 

My concern is that by the time Intel brings these to market in any real capacity, Zen 5C might already be out. Intel has had a bad habit recently of pushing back launches repeatedly.

Be aware that Zen4c are SMT capable cores where Intel's efficiency cores aren't. So, depending on workloads, 128 cores behave like there is 256 cores available. Now we know SMT doesn't scale 100%, but it can have typical uplift from 20-50%, again depending on workloads.

[leadeater]

6 hours ago, YoungBlade said:

I would think 288 E-cores would beat 128 Zen 4C cores pretty easily

Not all that likely, E cores are only around Skylake performance which is not bad but it's nowhere near Zen4c.

 

Skylake is as bad as 30% the performance of to at best 60% while also being deficient in cache sizes even compared to Zen 4c.

 

Intel will struggle to do any better than "as good as" and that'll be limited to only the highest configuration part they have while Zen 4c will still maintain the luxury of more than good enough low thread performance.

[leadeater]

6 hours ago, YoungBlade said:

Because SMT doesn't double CPU core performance and because 288 threads > 256 threads. So while a single core using a single thread is more than twice as fast, that same core having to juggle two threads will be slower than two cores each working on just one thread. In applications that are very good at using SMT, like Cinebench, the boost is only about 50%, meaning that the boost from a single core to all the cores on a 128 core EPYC CPU is only going to be about 192x, whereas all 288 of Intel's cores are working on just a single thread.

 

The EPYC 9654 is also based on Zen 4, not Zen 4C, which is what these CPUs are meant to compete against. Zen 4C has lower single core performance.

These will all be used for VM hypervisors and every VM will have CPU cycle/time QoS limits on them and they won't be performance use case oriented so realistically the only metric the matters is the number of threads x average frequency, not cores since in that scenario threads is king not cores. IPC also matters too but less so here since again not really performance oriented.

 

This is why ESXi reports host capacity as total frequency (Mhz/Ghz) available, the aggregate of all core.

 

 

 

As you'll notice it's calculated on base frequency not boost and typically I see around 30% boost, this is also going to be true for Intel so lets sideline that.

 

The rated all core frequency for the AMD 9754 is 3.1Ghz so it's 128 x 3.1 = 396.8Ghz per CPU or 793.6Ghz for a standard dual socket VM host. All we care about as hypervisor managers is the occupancy of CPU cycles and thread utilization, we don't care about how much work is being done by the CPU cycles since we the hypervisor manager cannot allow sharing of actual cycles. A CPU thread can only do one thing at a time in any given cycle, only the customer cares how much work is being done.

 

So as long as the performance is sufficient for the workload then which is better is a simple calculation like above, but only if performance does not matter.

 

If we do the same for Intel E-Core Xeons then we get

• 144 x 2.2 = 316.8Ghz
• 288 x 2.2 = 633.6Ghz

 

The above assumes the 288 core will maintain the 2.2Ghz which odds are it will not. In any case since performance does matter for me I have zero interest in 288 E cores since we don't host thousands of web servers/containers. If that is you then 2x 288 E cores would be highly interesting.