SPEC2017 Single-Threaded Results

SPEC2017 is a series of standardized tests used to probe the overall performance between different systems, different architectures, different microarchitectures, and setups. The code has to be compiled, and then the results can be submitted to an online database for comparison. It covers a range of integer and floating point workloads, and can be very optimized for each CPU, so it is important to check how the benchmarks are being compiled and run.

We run the tests in a harness built through Windows Subsystem for Linux, developed by Andrei Frumusanu. WSL has some odd quirks, with one test not running due to a WSL fixed stack size, but for like-for-like testing it is good enough. Because our scores aren’t official submissions, as per SPEC guidelines we have to declare them as internal estimates on our part.

For compilers, we use LLVM both for C/C++ and Fortan tests, and for Fortran we’re using the Flang compiler. The rationale of using LLVM over GCC is better cross-platform comparisons to platforms that have only have LLVM support and future articles where we’ll investigate this aspect more. We’re not considering closed-source compilers such as MSVC or ICC.

clang version 10.0.0
clang version 7.0.1 (ssh://git@github.com/flang-compiler/flang-driver.git
 24bd54da5c41af04838bbe7b68f830840d47fc03)

-Ofast -fomit-frame-pointer
-march=x86-64
-mtune=core-avx2
-mfma -mavx -mavx2

Our compiler flags are straightforward, with basic –Ofast and relevant ISA switches to allow for AVX2 instructions.

To note, the requirements for the SPEC licence state that any benchmark results from SPEC have to be labeled ‘estimated’ until they are verified on the SPEC website as a meaningful representation of the expected performance. This is most often done by the big companies and OEMs to showcase performance to customers, however is quite over the top for what we do as reviewers.

SPECint2017 Rate-1 Estimated Scores

Opening things up with SPECint2017 single-threaded performance, it's clear that Intel has improved ST performance for Raptor Lake on generation-upon-generation basis. Because the Raptor Cove P-cores used here don't deliver significant IPC gains, these performance gains are primarily being driven by the chip's higher frequency. In particular, Intel has made notable progress in improving their v/f curve, which allows Intel to squeeze out more raw frequency.

And this is something Intel's own data backs up, with one of Intel's performance breakdown slides showing that the bulk of the gains are due to frequency, while improved memory speeds and the larger caches only making small contributions.

The ST performance itself in SPECint2017 is marginally better going from Alder Lake to Raptor Lake, but these differences can certainly be explained by the improvements as highlighted above. What's interesting is the performance gap between the Core i9-13900K and the Ryzen 9 7950X isn't as far apart as it was with Alder Lake vs. Ryzen 9 5950X. In 500.perlbench_r, the Raptor Lake chip actually outperforms the Zen 4 variant by just under 4%, while Ryzen 9 7950X is a smidgen over 10% better in the 505.mcf_r test. 

SPECfp2017 Rate-1 Estimated Scores

Looking at the second set of SPEC2017 results (fp), the Ryzen 9 7950X is ahead of the Core i9-13900K by 16% in the 503.bwaves_r test, while the Raptor Lake chip is just under 10% better off in the 508.namd_r test. The key points to digest here is that Intel has done well to bridge the gap in single-threaded performance to Ryzen 7000 in most of the tests, and overall, it's a consistent trade-off between which test favors which mixture of architecture, frequency, and most importantly of all, IPC performance.

While we highlighted in our AMD Ryzen 9 7950X processor review, which at the time of publishing was the clear leader in single-core performance, it seems as though Intel's Raptor Lake is biting at the heels of the new Zen 4-core. In some instances, it's actually ahead, but stiff competition from elsewhere is always good as competition creates innovation.

With Raptor Lake being more of a transitional and enhanced core design that Intel's worked with before (Alder Lake), it remains to be seen what the future of 2023 holds for Intel's advancement in IPC and single-threaded performance. Right now, however SPEC paints a picture where it's pretty much neck and neck between Raptor Cove and Zen 4.

Core-to-Core Latency SPEC2017 Multi-Threaded Results
POST A COMMENT

169 Comments

View All Comments

  • OreoCookie - Tuesday, October 25, 2022 - link

    Yes, TDP has a meaning, and technically, neither company is using it correctly. Back in the good-ol’ days when TDP was really max power under load, it easily allowed you to spec a cooler. Clock boosts were meant to be temporary, transient states so that *on average*, you’d still lie within the thermal budget of the cooler. Obviously, we are well past that.

    So yes, AMD is playing it a bit loose (+31 %). But Intel is playing it ridiculous: the i9’s max power (as tested here) is 2.7x (!) their “TDP”.
    Reply
  • shaolin95 - Thursday, October 20, 2022 - link

    AMD does the same thing. dont be a fanboy Reply
  • yh125d - Thursday, October 20, 2022 - link

    If you're equating AMD going ~50w over TDP to intel going 210w over TDP, you're being the fanboy. Reply
  • Yojimbo - Friday, October 21, 2022 - link

    AMD's turbo clocking is more than 50W. Reply
  • Yojimbo - Friday, October 21, 2022 - link

    i checked and it's 60 W. That doesn't make AMD "less dishonest”. Neither company are being dishonest. It means AMD does not intend their desktop products to be used in lower power products. If you want to design a product around a Ryzen 7950X you need a 170 W cooling solution. Whereas you can put an i9 13900K in a product that can only dissipate 125 W. That's the difference between the two processors in terms if the TDPs. That's what TDP means. Reply
  • Truebilly - Friday, October 21, 2022 - link

    I'd like to see someone run that 13900k with 120mm rad Reply
  • Wrs - Friday, October 21, 2022 - link

    I mean, it works. The processor automatically steps down the v/f curve and doesn't hiccup with a puny cooler good for 140'ish W. I tested a 12900k with a low-profile AXP-200 from my Skylake days. Performance wasn't bad, over 4GHz all 16 cores. I left all the OC settings on, or else stock E-cores would be 3.9GHz. Reply
  • nandnandnand - Thursday, October 20, 2022 - link

    Go look at some efficiency curves for the 7950X and 13900K, for example at 19:00 in Hardware Unboxed's review: https://www.youtube.com/watch?v=P40gp_DJk5E Reply
  • Yojimbo - Friday, October 21, 2022 - link

    none of the companies "do” anything here. The "doing" is by the people who, though they are ignorant, write seething rants in comment sections damning the companies. Reply
  • bji - Friday, October 21, 2022 - link

    This issue would be a lot less contentious if technical sites like Anandtech actually used their expertise to curate information presented. They just shouldn't even show TDP as it's simply not relevant to the end users who are reading the articles. They should have some standard benchmark they run to determine peak and maximum sustained power draws and show ONLY those values in any charts. Reply

Log in

Don't have an account? Sign up now