MLA and FP8 MoE Serving: Multi-Head Latent Attention at Scale
Demystify Multi-Head Latent Attention (MLA) and low-precision FP8 Mixture-of-Experts serving. Slash KV-cache footprints by 93% on high-throughput GPUs.
Read Post →Featured Projects
View All Projects →Latest Blog Posts
Read All Posts →Disaggregated Inference: Separating Prefill and Decode Nodes at Scale
Scale LLM serving with disaggregated inference. Decouple compute-heavy prefill from memory-bound decode nodes to eliminate TTFT/TPOT interference.
Read Post →The Neural Rendering Matrix: Comparing NeRFs, Instant-NGP, 3D Gaussian Splatting, and NuRec
An exhaustive architectural, mathematical, and benchmark comparison across NeRFs, Instant-NGP, 3D Gaussian Splatting, and NVIDIA NuRec.
Read Post →