The latest articles on DEV Community by Carrie Guo (@carrie_netpreme). https://dev.to/carrie_netpreme https://media2.dev.to/dynamic/image/width=90,height=90,fit=cover,gravity=auto,format=auto/https:%2F%2Fdev-to-uploads.s3.us-east-2.amazonaws.com%2Fuploads%2Fuser%2Fprofile_image%2F4197010%2F79733c0d-68de-4b35-b29c-52537b124dc2.jpeg https://dev.to/carrie_netpreme en Carrie Guo Tue, 20 Oct 2026 23:27:29 +0000 https://dev.to/carrie_netpreme/xm-sparse-pushing-the-throughput-frontier-of-sparse-attention-with-netpreme-x-mem-1mk https://dev.to/carrie_netpreme/xm-sparse-pushing-the-throughput-frontier-of-sparse-attention-with-netpreme-x-mem-1mk <p>Sparse attention accesses fewer KV entries per decode step. The full KV cache still requires fast, high-capacity memory.</p> <p>Our new XM-Sparse testbed results with SGLang HiSparse and Netpreme X-Mem™ show how high-bandwidth access to offloaded KV and greater memory capacity can reduce time per output token (TPOT) and increase peak decode throughput compared with host DRAM.</p> <p>These results were measured in the X-Mem™ Pilot Virtual Machine. X-Mem™ silicon will be available later.</p> <p>Kudos to the Netpreme engineering team.</p> <p><a href="https://netpreme.com/blog/xm-sparse-sglang-hisparse-netpreme-xmem?utm_source=devto&amp;utm_medium=referral&amp;utm_campaign=xm_sparse_launch&amp;utm_content=community_post" rel="noopener noreferrer">Read the architecture and full benchmark results on Netpreme’s blog</a></p> ai llm performance machinelearning