AI Briefing#ssd
The latest AI and developer news about #ssd, with the original source and a short summary.
The latest AI and developer news about #ssd, with the original source and a short summary.
A logging bug in Codex was confirmed to have written 37TB over 21 days, potentially accelerating SSD wear.
Running Qwen 35B-A3B always-on on a 16GB Mac M4 hit an SSD bottleneck before a RAM one.
Samsung achieved 97%+ accuracy and under 200ms latency across 7 languages with on-device subtitle spotting for TVs.
Kakao's in-house developed Wcache achieves hundreds of thousands of TPS and a cache hit rate of over 95% with its SSD-HDD dual-layer architecture.
A summary of 29 tips for SSD development and system optimization to minimize write amplification and leverage internal parallelism.
Small random writes cause internal fragmentation and degrade performance, so bulk writes aligned with the Clustered block size are recommended.
In sustained random write environments, 25% Over-provisioning space is effective for maintaining SSD performance, complemented by TRIM and internal parallel processing technologies.
This article details Write Amplification caused by the non-overwritable nature of NAND flash and the Garbage Collection mechanism of the FTL.
Analyzes the lifespan and latency differences among SLC, MLC, and TLC, emphasizing the importance of 'Pre-conditioning' through sustained random writes.
A 30-page guide based on Key-Value store development experience covers SSD internal structures and optimal access patterns.