ML Research Scientist

Cerebras Systems builds the world's largest AI chip, 56 times larger than GPUs. Our novel wafer-scale architecture provides the AI compute power of dozens of GPUs on a single chip, with the programming simplicity of a single device. This approach allows Cerebras to deliver industry-leading training and inference speeds and empowers machine learning users to effortlessly run large-scale ML applications, without the hassle of managing hundreds of GPUs or TPUs.  

Cerebras' current customers include global corporations across multiple industries, national labs, and top-tier healthcare systems. In January, we announced a multi-year, multi-million-dollar partnership with Mayo Clinic, underscoring our commitment to transforming AI applications across various fields. In August, we launched Cerebras Inference, the fastest Generative AI inference solution in the world, over 10 times faster than GPU-based hyperscale cloud inference services.

About The Role 

The Core ML team at Cerebras is looking for both senior and junior research scientists dedicated to the development of novel, state-of-the-art ML algorithms. Our team focuses on research areas that emphasize our hardware's unique capabilities: 

• Accelerated unstructured sparse matrix multiplications enable weight sparsity (LINK) and activation sparsity.
• Unprecedented on-chip memory bandwidth enables >10x faster inference than GPU-based hyperscale cloud services (LINK).
• Weight streaming execution mode enables training 1T+ parameter models on 1 WSE, simple data parallel scaling instead of 3D+ parallelism required on GPUs.
• Flexible support to use low-precision numerics to improve performance of many aspects of training (LINK).

We also invest in advancing our fundamental understanding of ML training dynamics and distilling these insights into recipes which systematically improve existing customer ML workloads.  

Here is a sampling of our recent publications and releases:  

• Straight to Zero: Why Linearly Decaying the Learning Rate to Zero Works Best for LLMs, ICLR 2025 [paper](https://www.arxiv.org/abs/2502.15938) 
• The practitioner's guide to the maximal update parameterization, 2024, https://cerebras.ai/blog/the-practitioners-guide-to-the-maximal-update-parameterization 
• Sparse maximal update parameterization: A holistic approach to sparse training dynamics, NeurIPS 2024 [paper](https://arxiv.org/abs/2405.15743) 
• Normalization Layer Per-Example Gradients are Sufficient to Predict Gradient Noise Scale in Transformers, NeurIPS 2024 [paper](https://arxiv.org/abs/2411.00999) 
• Sparse-IFT: Sparse Iso-FLOP Transformations for Maximizing Training Efficiency, ICML 2024 [paper](https://arxiv.org/abs/2303.11525) 
• SPDF: Sparse Pre-training and Dense Fine-tuning for Large Language Models, UAI 2023 [paper](https://arxiv.org/abs/2303.10464) 
• BTLM-3B-8K: 7B Parameter Performance in a 3B Parameter Model https://huggingface.co/cerebras/btlm-3b-8k-base (most popular 3B model on Hugging Face), 2023 [paper](https://arxiv.org/abs/2309.11568) 
• SlimPajama: A 627B token cleaned and deduplicated version of RedPajama, 2023, https://www.cerebras.net/blog/slimpajama-a-627b-token-cleaned-and-deduplicated-version-of-redpajama