Machine learning systems sit at the heart of modern AI workloads. In these systems, performance often comes down to the quality of a small number of GPU kernels. Attention kernels, LLM prefill and decode kernels, low-precision block-scaled GEMMs, fused MoE layers, and other large fused kernels all directly shape end-to-end speed in both training and serving.
XiongjieDai/GPU-Benchmarks-on-LLM-Inference
Multiple NVIDIA GPUs or Apple Silicon for Large Language Model Inference?
GLaM: Efficient Scaling of Language Models with Mixture-of-Experts
Scaling language models with more data, compute and parameters has driven significant progress in natural language processing. For example, thanks to scaling, GPT-3 was able to achieve strong results on in-context learning tasks. However, training these large dense models requires significant amounts of computing resources. In this paper, we propose and develop a family of language models named GLaM (Generalist Language Model), which uses a sparsely activated mixture-of-experts architecture to scale the model capacity while also incurring substantially less training cost compared to dense variants. The largest GLaM has 1.2 trillion parameters, which is approximately 7x larger than GPT-3. It consumes only 1/3 of the energy used to train GPT-3 and requires half of the computation flops for inference, while still achieving better overall zero-shot and one-shot performance across 29 NLP tasks.
Can LLMs Be Computers? | Percepta
We build a computer inside a transformer — executing arbitrary C programs for millions of steps with exponentially faster inference via 2D attention heads.
AI GPUs probably live longer than three years
People who think current AI use is unsustainable often rely on the claim that inference GPUs only last “three years at the most” under load1. The idea here is that once the AI bubble money drains away, current infrastructure will rapidly become obsolete, and there won’t be enough money floating around to buy a whole slate of brand-new GPUs. Inference costs would thus rapidly become way too expensive for current AI products to make any financial sense.
Why Futhark?
A high-performance and high-level purely functional data-parallel array programming language that can execute on the GPU and CPU.
ST-MoE: Designing Stable and Transferable Sparse Expert Models
Scale has opened new frontiers in natural language processing -- but at a high cost. In response, Mixture-of-Experts (MoE) and Switch Transformers have been proposed as an energy efficient path to even larger and more capable language models. But advancing the state-of-the-art across a broad set of natural language tasks has been hindered by training instabilities and uncertain quality during fine-tuning. Our work focuses on these issues and acts as a design guide. We conclude by scaling a sparse model to 269B parameters, with a computational cost comparable to a 32B dense encoder-decoder Transformer (Stable and Transferable Mixture-of-Experts or ST-MoE-32B). For the first time, a sparse model achieves state-of-the-art performance in transfer learning, across a diverse set of tasks including reasoning (SuperGLUE, ARC Easy, ARC Challenge), summarization (XSum, CNN-DM), closed book question answering (WebQA, Natural Questions), and adversarially constructed tasks (Winogrande, ANLI R3).
Machine Learning Glossary | Google for Developers
This glossary defines a wide range of machine learning terms, including those specific to TensorFlow and large language models.
Stanford CS231N Deep Learning for Computer Vision | Spring 2025 | Lecture 1: Introduction
Async/await on the GPU
GPU code can now use Rust's async/await. We share the reasons why and what this unlocks for GPU programming.
Serpentine software
I don’t usually write hot takes. But StrongDM’s Software Factory post is worth it, so here goes. Their post is elegant, short, and well-written, so I’m going to just go through a bunch of pull quotes. Compounding correctness long-horizon agentic coding workflows [now] compound correctness rather than error This is a big, empirical claim. I’m already convinced by my own experiences that agents, suitably scaffolded, can form a Brownian ratchet, such that pouring on more compute helps, on balance.
Mixture-of-Experts Meets Instruction Tuning:A Winning Combination for Large Language Models
Sparse Mixture-of-Experts (MoE) is a neural architecture design that can be utilized to add learnable parameters to Large Language Models (LLMs) without increasing inference cost. Instruction tuning is a technique for training LLMs to follow instructions. We advocate combining these two approaches, as we find that MoE models benefit more from instruction tuning than dense models. In particular, we conduct empirical studies across three experimental setups: (i) Direct finetuning on individual downstream tasks devoid of instruction tuning; (ii) Instructiontuning followed by in-context few-shot or zero-shot generalization on downstream tasks; and (iii) Instruction tuning supplemented by further finetuning on individual downstream tasks. In the first scenario, MoE models overall underperform dense models of identical computational capacity. This narrative, however, dramatically changes with the introduction of instruction tuning (second and third scenario), used independently or in conjunction with task-specific finetuning. Our most powerful model, FLAN-MOE-32B, surpasses the performance of FLAN-PALM-62B on four benchmark tasks, while using only a third of the FLOPs. The advancements embodied byFLAN-MOE inspire a reevaluation of the design principles of large-scale, high-performance language models in the framework of task-agnostic learning.
On-device intelligence for every product
We're building a frontier lab for on-device AI. Small, specialized models for audio, vision, and text, faster than the cloud and free.
Evaluating and Sampling Glinty NDFs in Constant Time
Geometric features between the micro and macro scales produce an expressive family of visual effects grouped under the term 'glints'. Efficiently rendering these effects amounts to finding the highlights caused by the geometry under each pixel. To allow for fast rendering, we represent our faceted geometry as a 4D point process on an implicit multiscale grid, designed to efficiently find the facets most likely to cause a highlight. The facets' normals are generated to match a given micro-facet normal distribution such as Trowbridge-Reitz (GGX) or Beckmann, to which our model converges under increasing surface area. Our method is simple to implement, memory-and-precomputation-free, allows for importance sampling and covers a wide range of different appearances such as anisotropic as well as individually colored particles. We provide a base implementation as a standalone fragment shader.
Running local models is good now
Local agentic coding has gotten great over the past few months
Write-Only Code | Heavybit
AI may be accelerating software into a world of "write-only" code that forces a rethink of culture, processes, and tools. Heavybit’s Joseph Ruscio explains.
