Xe-Forge: Multi-Stage LLM-Powered Kernel Optimization for Intel GPU
- Uncategorized
- manufacturer agreement
- June 20, 2026
Abstract:Porting deep studying algorithms to new hardware accelerators requires developers to repeatedly apply the identical low-degree optimizations — quantization, memory entry coalescing, tile dimension tuning, and architecture-particular workarounds — to every Triton kernel in their code-base. This handbook, repetitive effort is a serious bottleneck: each kernel calls for the identical cycle of trial-and-error profiling in opposition to hardware constraints that fluctuate across devices, yet the underlying optimization patterns remain largely constant. We present Xe-Forge, a multi-stage LLM-powered pipeline that automates this course of for Intel GPU. Given a functionally correct Triton kernel, the system applies up to nine optimization stages — from algorithmic restructuring and operator fusion via block pointer modernization, GPU-specific tuning, and open-ended discovery — every pushed by a chain-of-Verification-and-Refinement (Cover) agent that generates candidates, validates them on actual hardware, and iterates on failures. A curated information base encodes Intel GPU constraints (energy-of-two warp counts, GRF modes, SLM sizing) which can be absent from LLM training information, maintaining the model within architecturally legitimate bounds. We evaluate Xe-Forge on 97 Level-2 KernelBench kernels and Flash Attention on the Intel Arc Pro B70, achieving a 1.17x geometric imply speedup over PyTorch keen with 67% of kernels enhancing, nine kernels exceeding 5x (as much as 82x), and 2–13.3x speedups on Flash Attention across all examined configurations without regression — demonstrating that structured area data with hardware-in-the-loop verification can systematically remove the repetitive porting effort that presently gates algorithm deployment on new accelerators.
Setgraph is free; Strong requires premium for full options. Try both and see which interface you desire. Hevy or JEFIT. Both offer muscle group analytics, intensive train libraries, and options specific to bodybuilding coaching. Hevy has a cleaner interface; JEFIT has more options. Setgraph (iOS and Android) or FitNotes (Android only). Both are completely free with no premium upsells. Setgraph or FitNotes. Both deal with core functionality without bloat. Hevy. The social feed and community facets are properly-applied if that motivates you. Don’t pay for premium options you do not need. Setgraph and FitNotes are utterly free. In the event you must have features from a paid app, Strong at $4.99/month is more inexpensive than JEFIT or Hevy. Boostcamp includes common applications and guides you thru them session by session. Calibr offers AI-powered teaching, however at $39.99/month, it’s costly. Setgraph’s AI workout generator offers program design with out the month-to-month cost. The worst selection is evaluation paralysis-spending so much time researching apps that you don’t actually prepare.
Abstract:We present AlphaLab, an autonomous research harness that leverages frontier LLM agentic capabilities to automate the total experimental cycle in quantitative, computation-intensive domains. Given only a dataset and a pure-language objective, AlphaLab proceeds by three phases without human intervention: (1) it adapts to the domain and explores the data, writing evaluation code and producing a research report; (2) it constructs and adversarially validates its personal evaluation framework; and (3) it runs large-scale GPU experiments via a Strategist/Worker loop, accumulating domain knowledge in a persistent playbook that capabilities as a form of online prompt optimization. All domain-specific conduct is factored into adapters generated by the model itself, so the same pipeline handles qualitatively totally different duties with out modification. We evaluate AlphaLab with two frontier LLMs (GPT-5.2 and Claude Opus 4.6) on three domains: CUDA kernel optimization, the place it writes GPU kernels that run 4.4x quicker than this http URL on common (as much as 91x); LLM pretraining, the place the complete system achieves 22% decrease validation loss than a single-shot baseline utilizing the identical model; and site visitors forecasting, the place it beats commonplace baselines by 23-25% after researching and implementing printed model households from the literature. The two models uncover qualitatively completely different options in every domain (neither dominates uniformly), suggesting that multi-mannequin campaigns provide complementary search protection. We additionally report results on monetary time sequence forecasting within the appendix, and release all code at this https URL.
The author was Columba’s ecclesiastical successor and his kinsman, and in his youth knew some who had been contemporaries of the saint. The earliest existing manuscript of the life is sort of as old as the time of Adamnan. Carlyle had read the book typically and admired it. You can see,’ he said, ‘ that the man who wrote it will tell no lie ; what he meant you can not always discover out, however it is clear that he instructed things as they appeared Agreement to Protect Property During a Split-Up him.’ The article of the life is just not to present dates or descriptions, but to exhibit the saintly character of Columba. In the account, nonetheless, of his prophetic revelations, of his miracles, and of his angelic visions, the three sections of the biography, his manner of life, his disposition, and his tastes, are easily realized. Most of what are described as wonders are simple occasions which take their miraculous colour from the observer’s perception within the constant interposition of providence in each day life.