New KV cache compaction technique cuts LLM memory 50x without accuracy loss

MIT researchers developed Attention Matching, a KV cache compaction technique that compresses LLM memory by 50x in seconds — without the hours of GPU training that prior methods required.

Nvidia says it can shrink LLM memory 20x without changing model weights

Nvidia's KV Cache Transform Coding (KVTC) compresses LLM key-value cache by 20x without model changes, cutting GPU memory ...
Embedded

Understanding cache placement

In a computer, the entire memory can be separated into different levels based on access time and capacity. Figure 1 shows different levels in the memory hierarchy. Smaller and faster memories are kept ...
TechRepublic

Cache-Only Memory Architecture

A Cache-Only Memory Architecture design (COMA) may be a sort of Cache-Coherent Non-Uniform Memory Access (CC- NUMA) design. not like in a very typical CC-NUMA design, in a COMA, each shared-memory ...