Riptide

We make AI run faster on the hardware you already have.

Riptide builds the compute layer under AI: GPU kernels, a compiler and a runtime that get more out of every chip in your fleet.

Riptide symbol: a wave of square tiles cut by a blue tiled sword
What we do

The layer between your model and the silicon.

Most AI runs on general-purpose kernels that leave performance on the table. We write the code that runs on the chip, and the tools that put it there.

Kernels

Hand-tuned GPU kernels for the operations AI spends its time on, checked against high-precision references before they ship.

Compiler

Turns models into code shaped for each chip: tiling, memory layout and scheduling chosen for the hardware, not a lowest common denominator.

Runtime

Launches work with as little overhead as possible, so small models and decode steps stop waiting on the host.

Kernels

Where the time goes, we go.

Measured

Attention

FlashAttention-style forward and backward passes, full and causal, on the GPU's matrix cores. Plugs into PyTorch's standard attention call.

In development

Low-overhead launch and fused decode

Step replay that cuts per-launch cost, and fused decode for small models, where reading the weights takes a fraction of the time the launches do today.

In development

Fused memory-bound ops

Softmax, normalization and elementwise chains fused into single passes, so they run closer to the memory bandwidth the chip actually has.

Roadmap

Matrix multiply

The core of every model, next on the list.

Hardware

Hardware we work with.

Measured

AMD Instinct MI350X

gfx950
Built for

AMD Instinct MI300X

gfx942
Runs today

Apple Silicon

Metal
Runs today

CPUs

C++
Public work

riptide-amd

Open-source GPU kernels for AMD Instinct, Apache-2.0. Install it next to PyTorch for ROCm and opt in. No fork, no patches, no changes to your stack.

# coming soon
pip install riptide-amd

import riptide_amd
riptide_amd.enable()  # attention now runs on riptide
  • Pre-release: the package and benchmark data publish together. Coming soon.
  • Unsupported calls fall back to PyTorch unchanged, and the reason is counted.
  • Every result passes float64 correctness checks; 169 GPU tests over odd shapes, strides and masks.
2.31× Causal attention training step (fwd+bwd), 512 tokens, head size 64 95% CI 2.30–2.32×
2.03× Same step at head size 128 95% CI 2.01–2.04×
1.08–1.09× Forward, head size 64, 8k–16k tokens 95% CI 1.07–1.10×
9% Less energy per call, forward, head size 64, 16k tokens 3.27 J vs 3.59 J

Speed-up versus AMD's best kernel for each case: the faster of PyTorch's FLASH_ATTENTION (AOTriton) and EFFICIENT_ATTENTION backends, measured in the same session. Median of 30 runs, 95% CI, MI350X, bf16, 32 heads, ROCm 10.2, PyTorch 2.15, 2026-11-19.

In progress: head size 128 at long sequences, where AMD's kernels are strongest. At 16k tokens we are at 0.43–0.69× of AMD's best today, and closing that gap is the current work.
Private work

Beyond one vendor.

We're also building kernels and ports for other accelerators and chip makers. That work happens privately, across GPU and accelerator vendors, and we talk about it when it's ready to share.

Work with us

Bring us the workload.

Chip companies

Kernels and ports for your accelerator, measured honestly against the best available baseline.

AI teams

More throughput and less energy from the GPUs you already run, without changing your stack.

business@splitclub.com