Geekbench

[Geekbench @geekbench]

We've moved to Mastodon! You can find us at @[email protected].

[Ryan Shrout @ryanshrout]

Primate Labs released @geekbench 7 today. It brings new media and game physics workloads, heavier data sets, and a rebuilt multi-core methodology that only runs a test multi-threaded if the real application it models actually is. Our @Signal_65 team ran it alongside Geekbench 6.7.1 on four current laptops from Qualcomm, Intel, AMD, and Apple to see what changes in practice.

Key findings from our testing:

➡️ Single-core scores fall on every system, from about 9% on the Ryzen AI 9 465 to roughly 15% on the Snapdragon X2 Elite. Geekbench 7 is recalibrated against a 2,500 baseline, so GB6 and GB7 scores are not directly comparable

➡️ The multi-core redesign favors higher core counts. The Snapdragon X2 Elite gained almost 16%, from 19,764 to 22,871, while the Core Ultra X9 388H was the only system to lose ground

➡️ The Apple M5 remains the single-core leader, but the Snapdragon converts a 12% single-core deficit into a multi-core lead of nearly 30%

➡️ On-battery performance splits along architecture lines. The Qualcomm systems hold their scores unplugged while the Core Ultra X9 gives up close to 16% single-core and the Ryzen AI 9 sheds 23%

With one platform per CPU, these are early results rather than a verdict on the architectures. More testing to come.

Full analysis: https://t.co/4RiXhKHJfj

[Max Weinbach @mweinbach]

Geekbench 7 is out!

The long and short of it is the workloads are now far more accurate to real software and real usage, multi-core mode is better, and single-core scores are more accurate

I've been running my devices for a week or so, you can see some of my scores here!

[MacRumors.com @MacRumors]

Geekbench 7 Launches With Redesigned Multi-Core and GPU Benchmarks https://t.co/DvT0abfHL1

[AppleInsider @appleinsider]

Primate Labs released Geekbench 7 for macOS, iOS, Android, Windows, and Linux, redesigning its cross-platform benchmark to better reflect how current software uses CPUs and GPUs.

https://t.co/y08V1LWhTE

[Tom's Hardware @tomshardware]

Geekbench 7 introduces biggest overhaul yet — real-world CPU testing, new media workloads, AI benchmarks, and CUDA support https://t.co/KAIT5m3Ou1

[The Verge @verge]

Geekbench 7 will push your computer or phone even harder for better benchmarking https://t.co/q74MwToZLf

[Geekbench @geekbench]

Geekbench 7 is now available, with new media workloads, smarter multi-core tests, and a refreshed GPU benchmark. Find out more at https://t.co/x0Kk14WvqX

[Geekbench @geekbench]

Geekbench 6.7 is out! This release adds Intel BOT detection, fixes stability issues on Linux ARM systems, and improves hardware identification. Find out more at https://t.co/pwmdQnCwgh

[Geekbench @geekbench]

We spent a week investigating Intel's Binary Optimization Tool and discovered it doesn't just reorder code, it can vectorize workloads, boosting some Geekbench 6.3 scores by up to 30%. Find out more at https://t.co/1IhPznx1l0

[Geekbench @geekbench]

Intel's Binary Optimization Tool can boost Geekbench 6 scores by up to 8%, but those results aren't comparable with standard runs. We've added a warning to affected results on the Geekbench Browser to keep scores comparable. Find out more at https://t.co/OGSNMNtWNI

[Geekbench @geekbench]

@4k_isn Geekbench reports the target ABI, not the instruction set. Geekbench 6 uses ARM v9 instruction set extensions (e.g., SVE, SME) in several workloads.

[Geekbench @geekbench]

@pers0naluni0n @toniievych @7600chip @Cartidise I'm not sure why you think that -- Geekbench 6.0 and later use AVX-VNNI (in addition to AVX512-VNNI and AMX) to accelerate inference tasks.

[Geekbench @geekbench]

@toniievych @7600chip @Cartidise Most of the difference comes from ML workloads. Geekbench 6 uses AVX-VNNI (as well as AVX512-VNNI) to accelerate ML tasks. While Zen 4 supports AVX-VNNI, OpenVINO and oneDNN didn't use it until 2023 (the article you’re referring to was written in 2022).

[Geekbench @geekbench]

@toniievych @7600chip @Cartidise Zen 4 AVX-512 is "double pumped" so the advantage over AVX2 isn't huge. Disabling AVX-512 on Ice Lake or Tiger Lake causes overall scores to drop by 10% and individual workload scores to drop by 60%.

[Geekbench @geekbench]

@toniievych @Cartidise You can see which ARM and x86 instruction sets Geekbench 6 uses on pages 8 and 9 of the Geekbench 6 Internals document: https://t.co/S4gODS8Qdp

[Geekbench @geekbench]

@toniievych @Cartidise Geekbench tests uses instruction sets that aren't available on all CPUs (e.g., NEON, AVX2, AVX512). Developer use these instructions to improve application performance. If Geekbench didn't use them it wouldn't provide an accurate measure of performance.

[Geekbench @geekbench]

@toniievych @Cartidise Actually, Geekbench 6 scores are comparable between systems with and without SME support. You just need to be careful when using Geekbench 6.2 or earlier on systems with SME support.

[Geekbench @geekbench]

@Thortripathi @mweinbach

[Geekbench @geekbench]

If you’re a developer or software engineer: Do you use an LLM to help you with your work?

Opinions on their real-world utility vary (we’ve definitely got our own), but we’re curious to hear your take — whether that’s a nuanced answer or just “no.”

[Geekbench @geekbench]

Geekbench AI is barely a few weeks old, but v1.1 is already here, with lots of little improvements and bug fixes. You may see better scores on some devices, and the benchmark should take less time to run — give it a spin and let us know what you got!

https://t.co/PzcKDPMTXa