Hardware1 hOctober 26, 2026, 16:38

Micron and Liqid boost LLM inference 2.8x with shared CXL memory pool

Micron built the Great Abaco cluster for PNNL, using a 160TB shared DDR5 pool over CXL 2.0 and Liqid fabric to extend the KV cache across GPUs. On Llama3-70B with 90% shared prefixes it delivered 2.8x the throughput and 2.9x faster time to first token versus DRAM of the same size.

Blocks & Files

Open source

Micron and Liqid boost LLM inference 2.8x with shared CXL memory pool

A single Abaco rack exposes up to 160TB of DDR5 as a shared addressable pool
Llama3-70B: 2.8x throughput and 2.9x faster time to first token
Cross-host KV block reuse gave 3x throughput with half the memory
CXL memory ran within 99% of host DRAM performance

#Micron#Liqid#Nvidia#CXL

Georgia Tech, Nvidia and Stanford's BOOST speeds LLM inference by 31%

Hardware

SK Hynix unveils MARS rack architecture with CXL memory pooling

Hardware

Micron and Liqid boost LLM inference 2.8x with shared CXL memory pool

Georgia Tech, Nvidia and Stanford's BOOST speeds LLM inference by 31%

SK Hynix unveils MARS rack architecture with CXL memory pooling

Samsung and AMD demo CXL memory pooling at OCP 2026: 38% lower TTFT

Prime Inference splits prefill and decode across separate GPU pools