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

RoundPipe

5 atoms · 跨 2 天 · 首见 2026-05-02 · 最近 2026-05-03

三色: 🟦 fact 5 · 🟥 take 0 · stance ▲1/▼0/◆0

来源: X 5

时态: fresh:4 · aging:1

标签: 好数字:5 · 好观点:1 · 好信源:1

2026-05-03 [老化中] · 性能指标
Achieves 1.5-2.2× speedups over SOTA baselines by dynamically dispatching stages in a round-robin manner for near-zero pipeline bubbles.
tags: 好数字·好信源 · → daily
value: qty=1.5-2.2× · date=baseline对比

2026-05-02 · 是唯一能在24 GB GPU上LoRA微调235B MoE模型的系统
RoundPipe is the only system capable of LoRA fine-tuning a 235B MoE model on 24 GB GPUs. Furthermore, its RTX 4090 throughput reaches at least 76% of existing A800 solutions across all models, effectively bridging the performance gap between consumer and data-center hardware.
tags: 好数字·好观点 · → daily
value: device=24 GB GPU · model_scale=235B MoE
2026-05-02 · 在8×RTX 4090服务器上相比最先进基线实现1.48-2.16倍加速
The results on an 8× RTX 4090 server show that RoundPipe achieves 1.48–2.16x speedups over state-of-the-art baselines when fine-tuning 1.7B to 32B models.
tags: 好数字 · → daily
value: qty=1.48-2.16x · device=8× RTX 4090 server
2026-05-02 · 在4090上实现高达2.16倍吞吐量和7.3倍更长序列
RoundPipe delivers, on 4090s, up to 2.16× higher throughput and 7.3× longer sequences; in turn, on NVIDIA A800s, it matches their throughput for small models, with up to 1.47× speedups and 5.6× longer sequences on large ones.
tags: 好数字 · → daily
value: qty=2.16x throughput, 7.3x sequence length · device=4090
2026-05-02 · 在RTX 4090上吞吐量达到现有A800方案的至少76%
its RTX 4090 throughput reaches at least 76% of existing A800 solutions across all models
tags: 好数字 · → daily
value: qty=≥76% of A800 throughput · device=RTX 4090 vs A800