以下部分引用 AI 总结,现阶段 AI 仍然有幻觉。请以内容的原文为准。
数据新鲜度群聊 07-14 ✓卖方 07-14 ✓Wrap 断档15天 ✕News 07-15 ✓X 断档12天 ✕页面生成 07-15
LENS
4 atoms · 跨 1 天 · 首见 2026-06-17 · 最近 2026-06-17
三色: 🟦 fact 4 · 🟥 take 0 · stance ▲0/▼0/◆1
来源: X 4
时态: fresh:4
标签: 好信源:4 · 好数字:3
事实 (fact) (4)
- 2026-06-17 · mean error on NPUs
The results across 248 cases on four commercial NPUs (Inferentia2, TPU v4, v5e, and v6e), three models, and four datasets show that LENS achieves a mean error of 2.15% and a median error of 1.69%, with 89.9% of cases within a 5% error margin.
tags: 好数字·好信源 · → daily
value: qty=2.15%
- 2026-06-17 · median error on NPUs
The results across 248 cases on four commercial NPUs (Inferentia2, TPU v4, v5e, and v6e), three models, and four datasets show that LENS achieves a mean error of 2.15% and a median error of 1.69%, with 89.9% of cases within a 5% error margin.
tags: 好数字·好信源 · → daily
value: qty=1.69%
- 2026-06-17 · percentage of cases within 5% error margin
The results across 248 cases on four commercial NPUs (Inferentia2, TPU v4, v5e, and v6e), three models, and four datasets show that LENS achieves a mean error of 2.15% and a median error of 1.69%, with 89.9% of cases within a 5% error margin.
tags: 好数字·好信源 · → daily
value: qty=89.9%
- 2026-06-17 · is a latency estimator
In this paper is presented LENS, a latency estimator that predicts NPU inference latency without information about the microarchitecture or compiler, and captures the nonlinear latency induced by bucketing.
tags: 好信源 · → daily
时间轴 (近 20)
- 2026-06-17 ·
fact · is a latency estimator · → daily
- 2026-06-17 ·
fact · mean error on NPUs · → daily
- 2026-06-17 ·
fact · median error on NPUs · → daily
- 2026-06-17 ·
fact · percentage of cases within 5% error margin · → daily
← 实体目录 · 系统日志