Intel 144-core Sierra Forest CPU benchmark shows strong performance, but still falls behind AMD's Bergamo

Intel
(Image credit: YuuKi_AnS/Twitter)

Intel's upcoming Sierra Forest server chip has leaked onto Geekbench 6 with impressive multi-core performance (via BenchLeaks). Sierra Forest comes with up to 144  efficiency cores, and it's slated to arrive in the first half of 2024. The server CPU is expected to compete directly with AMD's Zen 4c Bergamo server CPUs and ARM-based server chips like those made by Ampere.

The Sierra Forest server tested on Geekbench 6 used a dual-socket configuration of two 144-core Sierra Forest CPUs. There is a 288-core version of Sierra Forest on the way, but Geekbench makes it clear that two CPUs were used.

Swipe to scroll horizontally
Row 0 - Cell 0 Sierra Forest 144-core x2Intel Xeon Platinum 8480+ x2AMD Epyc 9754
Single-Core Score8551,8971,597
Multi-Core Score7,7707,46716,455

With a multi-core score of 7,770, this dual-chip Sierra Forest machine is pretty fast. Most dual-socket systems equipped with Intel's top-end Xeon Platinum 8480+ score roughly between 6,500 and 7,500 on the Geekbench 6 database, and we've selected a result that's roughly in the middle. Of course, the single-core performance of Sierra Forest is much worse and isn't even half that of the 8480+.

That's all by design, though. The Sapphire Rapids-based 8480+ has 56 Golden Cove performance cores compared to Sierra Forest's 144 Glen E-cores, and P-cores are individually fast while E-cores are slow. However, E-cores are generally more efficient in terms of area and power, and Sierra Forest is expected to come in handy for servers where multi-core performance is all that matters.

Of course, the elephant in the room is AMD's Bergamo server CPUs, which take a pretty similar approach to what Sierra Forest does. Bergamo uses Zen 4c cores, which are architecturally identical to Zen 4 but have significantly reduced area and power consumption. This Geekbench 6 result, which is also about middle of the road, pretty much steamrolls both Sierra Forest and Sapphire Rapids in multi-core performance.

Though, the obvious caveat is that this Sierra Forest result comes from a pre-release system. For all we know, it could be massively underperforming on the multi-core score, and maybe also the single-core score too. It's also possible that Sierra Forest will come with unique instructions or other features that Geekbench can't really account for; it's how Intel got tapped for a Microsoft Azure instance that used AMD's MI300X GPUs, as Sapphire Rapids has AI-accelerated instructions.

Additionally, the benchmark also detailed Sierra Forest's cache configuration, but it's not clear if these specifications are accurate. We had expected Sierra Forest to have quad-core clusters that each had 4MB of L2 cache, which would mean 36 clusters on a 144-core chip, but Geekbench says there's just 18 clusters. However, there's also a reported 108MB of L3 cache, and since each cluster is supposed to come with 3MB of L3 cache, that would come out to 36 clusters.

We can probably assume Geekbench 6 just got the amount of L2 clusters wrong, which would mean a total of 144MB of L2 cache per CPU. That means Sierra Forest's combined L3 and L2 cache is 252MB, which is pretty big though still smaller than Bergamo's L3 cache on its own, which is 256MB.

Matthew Connatser

Matthew Connatser is a freelancing writer for Tom's Hardware US. He writes articles about CPUs, GPUs, SSDs, and computers in general.

  • zecoeco
    288 E-Cores as powerful as 112 P-Cores.
    That's very disappointing and it shows how a single E-Core has way less than 50% of the performance of a single P-Core.
    Reply
  • usertests
    Am I reading this right? 288 cores (144-core x2) of Sierra Forest are getting curbstomped by 128 cores of Bergamo?

    By the time the 288-core Sierra Forest appears, it might have to contend with 192 cores of Zen 5C. But it will already be slower than 128-core Bergamo unless dual-socket is used.
    Reply
  • Metal Messiah.
    usertests said:
    Am I reading this right? 288 cores (144-core x2) of Sierra Forest are getting curbstomped by 128 cores of Bergamo?

    Yeah, actually espcially in multi-core score.

    Bergamo seems to have a major lead with 256 threads versus Sierra Forest, and considering that E-Cores don't offer multi-threading and are lacking the same performance as the P-Core offerings, AMD seems to be competitive with its current and future Zen 'compute-density' offerings.

    It appears Intel will have higher cores but since the architecture is different, AMD seems to be more competitive, since the Bergamo/Zen 4C lineup shares the same ISA as the Genoa CPUs, albeit with more compute and efficiency.
    Reply
  • Order 66
    Wow, a pathetic 4% multicore performance increase going from 112 cores to 288. I understand that these are efficiency cores that are lower power and lower clock speed than the performance cores, but still, it is more than double the cores.
    Reply
  • endocine
    these things better have industry leading power efficiency to justify the low performance
    Reply
  • DavidC1
    You guys do realize that it hasn't been long since Intel announced Intel 3 was manufacturing-ready? These are extremely early Sierra Forest samples.

    The final variant will be very competitive with what's out there, yes even Turin Dense(successor to Bergamo).
    Reply
  • bit_user
    ... compared to Sierra Forest's 144 Glen E-cores, ...
    They're called "Sierra Glen".

    It's also possible that Sierra Forest will come with unique instructions or other features that Geekbench can't really account for
    It could have new accelerators (or versions of existing ones) that aren't being used, but software like Geekbench would never use those anyway.

    As for new instructions, we already know what it has. Intel telegraphs this stuff long in advance, so that software developers can be prepared in advance of each new hardware launch.
    1718933516535070950View: https://twitter.com/InstLatX64/status/1718933516535070950It's all pretty small tweaks, and mostly OS-level or aimed at tuning AI, I think.
    Reply
  • bit_user
    zecoeco said:
    That's very disappointing and it shows how a single E-Core has way less than 50% of the performance of a single P-Core.
    That's not what we saw when Alder Lake launched. Its E-cores were about 60% of a single P-core, and actually faster than a thread sharing a P-core with another (via hyperthreading).

    Since these E-cores are a whole generation newer, I'd expect this isn't the final word on their performance.
    Reply
  • rtoaht
    DavidC1 said:
    You guys do realize that it hasn't been long since Intel announced Intel 3 was manufacturing-ready? These are extremely early Sierra Forest samples.

    The final variant will be very competitive with what's out there, yes even Turin Dense(successor to Bergamo).
    Exactly.

    Everyone pretending it to be the final version with healthy Intel 3 process. Even Meteor Lake on Intel 4 isn't released yet. How can everyone expect full performance for Sierra Forrest on Intel 3? If anything, it just shows that very early Intel 3 Silicon is progressing well. This means both Sierra Forrest and Granite Rapids are on track. "5 nodes in 4-years" plan is on track.
    Reply
  • DavidC1
    bit_user said:
    Since these E-cores are a whole generation newer, I'd expect this isn't the final word on their performance.
    Sierra Glen is pretty much identical to Gracemont performance-wise. The client sister Crestmont gets 4-6% gain due to 6-wide rename/allocate and improved branch predictor. It's opposite for the P cores, where the client version only gets doubled L1i cache and server version gets improved branch prediction, lower FP latencies and better memory ILP.

    A leaker said Sierra Glen forgoes those improvements to get higher clocks.

    https://www.anandtech.com/show/20034/hot-chips-2023-intel-details-granite-rapids-and-sierra-forest-xeonshttps://www.anandtech.com/show/20046/intel-unveils-meteor-lake-architecture-intel-4-heralds-the-disaggregated-future-of-mobile-cpus/2
    Reply