We publish practical research on workload optimization and orchestration across heterogeneous
GPUs. Find papers, reproducible benchmarks, grants, and media coverage

Peer-reviewed papers and preprints on efficient training, model offloading, inference latency, and GPU scheduling.
Highly-efficient billion-scale AI models training and inference using affordable GPUs
ZenFlow: stall-free offloading training via asynchronous updates
ZeRO-Offload and Sentinel for transformers
DyNN-Offload for Mixture-of-Experts (MoE)
TECO-Offload on disaggregated memory
ChunkFlow: communication-aware layerwise offloading for distributed diffusion transformer inference
Billion-scale graph neural network
Distributed AI training and inference based on parallelism and communication management
CoCoDiff: optimizing collective communications for distributed diffusion transformer inference under Ulysses sequence parallelism
BloomBee: internet-scale decentralized LLM inference with multi-dimensional communication optimization
Runtime Concurrency Control and Operation Scheduling
Tree structure-aware high performance inference engine
Hardware-aware AI optimization and performance profiling
CARDAN: multi-engine dataflow for MoE decoding on scratchpad-based tensor accelerators (AWS Trainium3)
Operator Profiler: hardware-attributed operator profiling for PyTorch
Energy-efficient training on GPU-FPGA accelerators
Processing-in-memory for energy-efficient DNN
Decentralized AI Computing Operating System for Accessible and Cost-Effective AI


Coverage of our research, open-source and product releases.