01
RTL Interconnect IP
CHI and XGMI protocol stacks designed at register-transfer level. Multi-GPU communication fabric engineered for scale, not retrofitted from legacy designs.
GPU interconnect IP for AI datacenter scale-up.
RTL-level design. FPGA-validated. AI-aware scheduling.
Luma Semiconductor builds GPU interconnect IP for AI datacenter scale-up. Our core work is RTL-level design and verification of multi-GPU communication fabric, validated through FPGA synthesis, not simulation alone.
We combine this with RTL-based traffic modeling and prediction to make interconnect scheduling AI-aware, targeting lower latency and better bandwidth utilization under real inference and training workloads.
CHI and XGMI protocol stacks designed at register-transfer level. Multi-GPU communication fabric engineered for scale, not retrofitted from legacy designs.
RTL-based traffic modeling and prediction to make interconnect scheduling AI-aware. Dynamic bandwidth allocation targeting lower latency and better utilization under real inference and training workloads.
FPGA synthesis and benchmarking on Xilinx Kintex UltraScale+. Every architecture decision backed by real silicon behavior, not simulation alone.
A unified stack, from protocol to fabric to AI scheduler,
designed to operate as licensable IP.
We are building toward licensable interconnect IP and reference implementations that datacenter and accelerator companies can integrate directly, without rebuilding the fabric from scratch.
Discuss licensing →Reach out to discuss partnerships, licensing, or integration.