Nvidia shrinks 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 costs and time-to-first-token by up to 8x for multi-turn AI applications.
TMCnet

Penguin Solutions Introduces Industry's First Production-Ready CXL-Based KV Cache Server

Accelerating memory-dependent AI processes, Penguin's MemoryAI KV cache server increases memory capacity by integrating 3 TB of DDR5 main memory and up to eight 1 TB CXL Add-in Cards (AICs). Penguin ...

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 — ...
Virtualization Review

What GPU You Really Need for AI Workloads

GPU memory (VRAM) is the critical limiting factor that determines which AI models you can run, not GPU performance. Total VRAM requirements are typically 1.2-1.5x the model size due to weights, KV ...
Semiconductor Engineering

Dynamic KV Cache Scheduling in Heterogeneous Memory Systems for LLM Inference (Rensselaer Polytechnic Institute, IBM)

A new technical paper titled “Accelerating LLM Inference via Dynamic KV Cache Placement in Heterogeneous Memory System” was published by researchers at Rensselaer Polytechnic Institute and IBM. “Large ...