Research2026-05-30
3D silicon stacking aims to extend Moore's Law
Researchers report a chip-stacking advance that could keep transistor density — and GPU throughput — climbing for years.
Source: sciencedaily.com
Research News
Papers and breakthroughs shaping what GPUs will do next. Back to all the latest.
Latest Research news
Research2026-04-30
FP4 training shown stable at scale
Researchers demonstrate stable four-bit training on Blackwell, cutting memory and energy for large runs without hurting quality.
Source: arxiv.org
Research2026-02-04
New MoE routing cuts inference cost
A routing scheme improves mixture-of-experts efficiency, lowering the cost of serving very large models.
Source: arxiv.org
Research2025-12-18
Diffusion LLMs promise faster generation
Diffusion-based language models show parallel decoding that could beat autoregressive speed.
Source: arxiv.org
Research2025-09-25
KV-cache compression extends context cheaply
New methods shrink the memory cost of long context, letting smaller GPUs handle bigger windows.
Source: arxiv.org
Research2025-07-03
1-bit LLMs become practical to serve
Ternary and 1-bit weight schemes cut memory enough to run large models on modest GPUs.
Source: arxiv.org