Improving Llama 2 Power Efficiency by Overclocking Arrow Lake Intel Graphics | SBCP#5

Today, I want to have a closer look at using the undervolting potential to improve the efficiency for Llama 2 generative AI workloads. For this purpose, I’ll rely on MLPerf Client – a relatively new benchmark I’ve been playing with.

In this video, I demonstrate that it’s possible to increase Llama 2 generative AI performance by almost 40% at nearly no efficiency cost by overclocking the graphics cores and memory subsystem. Similarly, you can improve power efficiency by about 4% by undervolting the graphics cores at default frequency.

SkatterBencher blog: https://skatterbencher.com/2025/05/15...
Join the Discord:   / discord  
Join the Patreon:   / skatterbencher  
Buy me a coffee: https://buymeacoffee.com/skatterbencher
Become a YouTube Member:    / @skatterbencher  

Hardware list:
Intel Core Ultra 9 285K https://amzn.to/4bB1OZA
ASUS ROG Z890 Apex https://amzn.to/4ig4BKf
EK-Quantum Velocity2
EK-Quantum Kinetic FLT 240 D5
EK-Quantum Surface P480M
EK-Furious Meltemi 120 (x4)
Corsair Vengeance 2x24GB DDR5-8000 CUDIMM https://amzn.to/4je6ebS
Enermax Revolution ATX 3.0 1200W https://amzn.to/3FhvnU4
AGI 512GB NVMe M.2 Gen3
Open Benchtable V2 Titanium https://openbenchtable.com/
BENCHLAB https://benchlab.io/

As an Amazon Associate I earn from qualifying purchases.

Any questions on our overclocking, ask in the comments or reach out via the contact page on my website.