⚙️ GearBench v0.1.2-alpha
-Scaling system has been removed
This release introduces a complete rewrite of the Read, Write, and Copy benchmark functions to achieve:
🎯 Accurate DRAM bandwidth measurement, bypassing CPU cache
⚡ AVX2-based parallelism inside each thread for full SIMD utilization
🧵 Automatic max thread usage to saturate the memory bus
🔁 Alternating memory buffer access to poison cache lines and ensure realistic results
🔍 Changes:
Brand new MeasureRead, MeasureWrite, and MeasureCopy logic
256B-aligned AVX2 streaming with cache-bypass logic
Time-based execution (TARGET_SECONDS = 5.0) instead of fixed loop count
More stable and fair results across CPUs with different core/thread counts
Copy operation refined to avoid write bottlenecks
📊 Results:
Greatly reduced CPU thread count distortion in measurements
Validated on DDR4 and DDR5 systems – even under cache poisoning
Latency metric remains untouched, still highly accurate using pointer chasing
This is still an alpha release, UI and results may change in future updates.