AI workloads & kernels
Custom GPU kernels, fused operations, and better memory use. Built for your model.
AI Performance / Numaro
Make AI
run faster.
Kernels, compilers, and specialized languages. We optimize the software around your workload and hardware.
Autonomous research. Applied to your stack.
Your performance target
- Latency
- Throughput
- Memory
- Energy
Across the stack
Three ways to go faster.
From one operation
to a whole new toolchain.
Compilers & runtimes
Scheduling, code generation, and workload mapping for GPUs and custom accelerators. Target stacks include MLIR/IREE, TVM, and LLVM, with hardware or simulator evaluation.
Specialized languages & toolchains
A language and compiler built around your domain. A deeper collaboration, from design to execution.
Our team guides the autoresearch engine and reviews the results.
How it worksFrom our optimization work
Measured speedups.
2–4.5×
Faster batched matrix multiplication
Selected small / medium matrices · batch ≥ 512
Relative speedHigher is faster ↗
Hatched area: variation across tested shapes.
NVIDIA L40S · Bit-exact in the reported tests
Benchmark conditions
Selected small and medium matrices, batch ≥ 512, versus torch.bmm / cuBLAS. Outputs matched the baseline bit for bit in the reported tests.
Measurements on one NVIDIA L40S in the report’s stated environment, using paired, interleaved timing on a shared GPU. Results depend on workload, hardware, and baseline; a pilot measures your stack and the end-to-end impact.
Read the GPU benchmarks ↗1.55×
Faster fused depthwise convolution
Fused 3×3 + bias + ReLU6 · FP16 geometric mean
Relative speedHigher is faster ↗
NVIDIA L40S · Bit-exact in the reported tests
Benchmark conditions
FP16 geometric mean for depthwise 3×3 convolution, bias, and ReLU6 versus cuDNN convolution followed by activation. Bit-exact in the reported tests.
Measurements on one NVIDIA L40S in the report’s stated environment, using paired, interleaved timing on a shared GPU. Results depend on workload, hardware, and baseline; a pilot measures your stack and the end-to-end impact.
Read the GPU benchmarks ↗2.93×
Faster compiled program execution
80 C-reference benchmarks · geometric mean
41 wins · 7 ties · 32 losses. 1.17× when excluding Numaro runs under 1 ms.
Relative speedHigher is faster ↗
x86-64 CPU · Program execution time
Benchmark conditions
Geometric-mean execution speedup across 80 C-reference benchmarks versus Clang 21 -O2, using Numaro’s experimental compiler. The programs express equivalent computations in different source languages and match the benchmark checksums. This measures execution time, not compilation time or Rust performance.
Recorded August 8, 2026 on an x86-64 CPU. Timings include process startup, with a warmup followed by the minimum of five runs pinned to one core. This is a historical benchmark snapshot, not a new timing run.
Against Clang 21 -O2: 41 wins, 7 ties within 2%, and 32 losses. The median runtime is 0.9733× the baseline. Large algorithmic gains and short-running cases have a substantial effect on the 2.93× geometric mean.
Restricting the comparison to the 66 cases where Numaro’s execution takes at least 1 ms gives a 1.17× geometric-mean speedup. Across all 80 cases, the geometric mean versus Clang -O3 -march=native is 2.48×.
The custom compiler is not a drop-in C or Rust compiler. These results do not establish faster compilation, a universal runtime advantage, or performance on a custom accelerator.
View the benchmark data ↗The people behind the work
Technical team.
Work directly with us,
from the first benchmark to handover.
Vadim Borisov, PhD
ML & autonomous research
Vadim guides workload selection, ML evaluation, and the research campaign. His work spans generative models and efficient inference.
Philipp Schuster, PhD
Programming languages & compilers
Philipp guides language and compiler design, optimization strategy, and technical review. His research connects programming language theory to native code generation.
Work with Numaro
Bring your bottleneck.
- 01
Scope it together
Your engineers and our team choose a workload, establish the baseline, and agree on performance and correctness targets.
- 02
Guide the research
Our team directs the campaign on your hardware or simulator, reviews candidates, and checks the results with you.
- 03
Technical handover
We deliver evaluated code and reproducible benchmarks, walk your engineers through the changes, and guide integration.
Targets follow baseline review. Simulated results are labeled.
contact@numaro.tech