Sakana AI’s new system uses LLMs and evolutionary optimisation techniques to automate this process, making high-performance CUDA kernel development more accessible. “The coolest autonomous coding ...

My first CUDA program is a simple Vector Addition kernel. CUDA (Compute Unified Device Architecture) is a parallel computing framework by NVIDIA that allows us to use GPUs for general-purpose ...

The unrolling process transforms the input image and kernel into a larger matrix. Each row of this matrix corresponds to a single "convolution window" – the set of pixels covered by the kernel at a ...

These CUDA kernels are 10-100 times faster than typical PyTorch versions Sakana AI is also releasing a dataset with over 30,000 CUDA kernels The company recently released The AI Scientist ...

2. Optimized CUDA kernel calls: To utilize FP8 matrix multiplications effectively, they refined CUDA kernel usage and even provided NVDA with suggestions for Tensor Core improvements. 3.

The company added that by writing instructions directly at the CUDA kernel level, it can achieve much higher performance for AI algorithms. Nvidia is an investor in Sakana. The startup noted that ...

作者｜冬梅、Tina3 月 1 日，DeepSeek 公司在发布了一系列开源项目后，又放了一个“大招”，首次公开了其模型推理系统 DeepSeek-V3 / R1 的技术细节和盈利数据。根据 DeepSeek 公布的信息，这个系统理论上一天能赚 56 ...

为了解决这一问题，近日字节跳动 ByteDance Research 团队开发并开源了 ByteQC —— 基于 GPU 加速的大规模量子化学计算工具集。该工具集使用强大的 GPU ...