Company

Performance engineering for every model

Turbokern exists so that any AI model can run at its best on every NVIDIA GPU without a team of scarce performance experts.

TurbokernCompany
Stage
Pre-launch
Category
AI infrastructure
Focus
Inference on NVIDIA GPUs
Domain
turbokern.com
Pilot
2027

Why we started

The fastest engineers cannot be everywhere

Inference is becoming the largest recurring cost of AI. The work that cuts it, profiling, kernels, quantization and serving, is done by a small number of specialists, one model and one GPU at a time.

Each new GPU generation resets that work. We think the loop those engineers run can be automated, measured and repeated for every model, and that the data from each run makes the next one better.

MissionRun every AI model at its theoretical best on every NVIDIA GPU.
ApproachAn agent that measures everything and keeps only what it proved.
ScopeInference first; training, GPU FinOps and edge next.
EcosystemBuilt on CUDA, CUTLASS, TensorRT, Triton and Nsight.

Principles

How we work

01

Measured, then claimed

We publish numbers we measured, labelled with the hardware. Everything else is a target.

02

Correct before fast

A faster wrong answer is a bug.

03

Plain artifacts

Outputs are standard NVIDIA artifacts you can inspect and own.

04

Small and senior

A small team of performance, compiler and ML engineers.

Roadmap

What we plan to build, in order

Draft roadmap. Dates are targets.

  1. 2026 H2

    Profyle and harness

    Profiling and the real-GPU benchmark harness for LLM inference on H100.

  2. 2027 H1

    Pilot loop

    Kernl, Quantra and Servex with Simyx pruning, run with design partners.

  3. 2027 H2

    Blackwell and edge

    Re-optimization on Blackwell; Edgyn on Jetson Orin.

  4. 2028

    Platform

    General availability, on-prem, and training optimization research.

Careers

Open roles

Draft openings. Email careers@turbokern.com with a note about the hardest performance problem you have solved.

  • CUDA kernel engineerCUTLASS, attention and GEMM kernels on Hopper and Blackwell.Remote or SF
  • ML compiler engineerTensorRT-LLM engine building, graph rewrites and fusion.Remote
  • Inference serving engineerTriton, NIM, batching and tail-latency work under load.Remote
  • Research engineer, performance modelingLearned performance predictors and active learning for Simyx.Remote
  • Embedded GPU engineerJetson and DRIVE optimization under power and thermal limits.Remote

Our stack

What we build with

GPUCUDA, CUTLASSKernels
EnginesTensorRT-LLMInference
ProfilingNsight, CUPTITraces
DataRAPIDSTrajectory dataset
ServingTriton, NIMPackaging
EdgeJetPack, DeepStreamDevices
AgentNeMo-class LLM plannerStrategy
InfraKubernetes, SlurmHarness

Next step

Work on the hardest part of AI infrastructure

If you have tuned kernels or serving by hand and want to automate it, write to us.