ReCIM-HD is an on-device learning (ODL) accelerator that integrates gradient-free hyperdimensional computing (HDC) with resistive RAM (ReRAM)-based analog compute-in-memory (CIM). ReCIM-HD introduces three key techniques: (1) ultra-low-power HDC-based edge ODL without costly back-propagation operations; (2) non-volatile ReRAM-based storage for class hypervectors and CIM similarity search, enabling energy-efficient inference and power-gating during idle periods without data loss; and (3) a hybrid digital-analog CIM inference flow that protects a small accuracy-critical subset digitally to preserve accuracy. Additionally, it employs a Kronecker HD encoder (KE) to reduce encoding overhead and a customized instruction set architecture (ISA) to support flexible programmability. Fabricated in 40 nm, ReCIM-HD achieves 44.17 TOPS/W (hybrid mode) and 75.8 TOPS/W (ReRAM CIM mode), delivering up to 20.1x higher energy efficiency than the state-of-the-art accelerator.
On-policy distillation (OPD) trains a student on its own rollouts under a frozen teacher's per-token supervision. Does this dense signal reduce the training FLOPs needed to reach a target accuracy compared with reinforcement learning? We systematically compare vanilla reverse-KL OPD with Group Relative Policy Optimization (GRPO) across task difficulty, student and teacher size, domain, and optimization choices, measuring held-out accuracy against training FLOPs. With an external post-trained teacher, OPD generally loses to GRPO in accuracy per training FLOP at matched settings, even when the teacher is further trained with GRPO, and often suffers response-length inflation. Yet a different teacher changes the picture: when the teacher is the student's own GRPO checkpoint, OPD recovers that checkpoint's accuracy on math with $50\times$ fewer distillation FLOPs than GRPO needs to reach the same accuracy. Following the same teacher setting, we consistently observe this high efficiency with a larger student, on code, and in a second model family. On a single domain, however, the GRPO checkpoint is already the desired model, so distilling it adds to total compute. Across domains, separate GRPO checkpoints hold different capabilities, and no single checkpoint serves every domain. Multi-teacher OPD (DeepSeek-AI, 2026) consolidates per-domain GRPO experts into a single student that matches each expert's accuracy, with the distillation stage costing only 5.2% (on 2 domains) and 8.0% (on 4 domains) of the specialist training FLOPs. Counting specialist training and distillation together, this pipeline reaches joint GRPO's performance on the 2-domain mix with 42.0% fewer total FLOPs.
John Martabano Landy, Louis C. Tiao, Yao Xuan, Luc Kien Hang, Rupinder P. Khandpur, Arvind M. Vepa, Srinidhi Raghavendran, Anirudh Agrawal 等 13 位作者
原文摘要
Large-scale ranking and recommendation models of Web platforms benefit from increased compute much as large language models do, and their scaling behavior is broadly power-law in form. Yet ranking models have heterogeneous architectures which makes their scaling law less straightforward than that of LLMs. We address how to discover the scaling law of an arbitrary ranking model efficiently. We introduce a scaling layer, which groups primitive architecture parameters into semantically coherent groups and defines how they co-vary, inducing a mapping from a low-dimensional space of hyper-hyperparameters into the full architecture space. On top of it, Axia fits scaling-informed surrogates, selects configurations by a posterior-based acquisition criterion, and returns a Pareto frontier from which per-model operating points are validated -- replacing exhaustive sweeps with a small number of informative runs. A fleet-level evaluation protocol, computed on a fixed representative model set under matched trial budgets, makes refinements to either component measurable. Our contribution is an automated system that learns model-specific complexity-quality relationships and gives uncertainty-aware guidance for allocating compute across a heterogeneous model ecosystem. Across 34 model iterations in two development cycles on Meta's ads ranking and retrieval system, the workflow yielded a mean 0.66% offline normalized-entropy improvement and 0.36% cumulative online lift.