About
An applied inference optimization lab
RiftStack works on the software layer between models and hardware. We build Emmy, a compiler that generates fast kernels, and a serving stack that runs them efficiently, so the same model costs less and responds faster on the GPUs teams already have.
The work runs in three tracks. Inference covers benchmarks and serving across consumer and datacenter GPUs, from the RTX 4090 to the B200. Virtualization covers how GPUs are run inside virtual machines, from VFIO passthrough to SR-IOV, and the hardware bugs found on the way. Compiler covers Emmy and the CUDA kernel work it is measured against. Results are published model by model and GPU by GPU, with the setup and the limits of each number stated, so a reader can tell what carries over to their own hardware.
RiftStack is operated by CloudRift, Inc., which runs a GPU cloud on the same hardware the research targets. The virtualization work comes out of that production environment, and the inference and compiler work goes back into it. Emmy is published under Apache-2.0 and installs from PyPI, and the engineering series on this site documents how it was built. We also run the Emmy Hackathon around it, and standout entries may lead to an interview with the team.
Team
Track record
A small team building low-level systems for performance-critical software, with backgrounds at Apple, Roblox, and Ubisoft.
50–60% over cuBLAS
FP32 SGEMM in batched mode on the RTX 5090. In non-batched mode cuBLAS still wins.
1,157 tok/s
Qwen3 Coder on a single consumer-class RTX 5090.
ML compiler from scratch
Emmy: tracing, fusion, scheduling, and CUDA codegen in Python.
Join us
Work with us
RiftStack is operated by CloudRift, and open roles are posted there.
See open roles at CloudRiftWe give priority consideration to people who have taken part in a RiftStack hackathon, because we have already seen how they work. Taking part does not guarantee an interview or an offer.