English(EN) WideEP dispatch and combine need to happen twice on every layer and for every forward pass. That means that, for each forward pass, dispatch and combine will ne

Kimi K3 的海量规模尽管经过优化，仍需要密集的网络连接 · 跟踪 8 个来源

作者 PulseAugur 编辑部 · [8 个来源] · 2026-07-18 21:35

SemiAnalysis 报道称，拥有 2.8 万亿参数的 Kimi K3 模型，尽管经过了 Kimi Delta Linear Attention (KDA) 等优化，仍需要显著的网络带宽。该模型的架构需要 WideEP 服务优化，将 896 个专家分布在许多 GPU 上，导致密集的网络使用。这种复杂性意味着，虽然 KDA 减少了 KV 缓存传输，但 Kimi K3 的整体网络需求仍然很大，可能会影响 AI 网络交换机市场。 AI

影响 Kimi K3 复杂的架构和高带宽需求凸显了高效扩展大型 AI 模型所面临的持续挑战和创新。

排序理由该集群包含来自 SemiAnalysis 的多条推文，讨论了 Kimi K3 模型架构和服务的技术方面及其影响，而不是主要公告或发布。

在 X — SemiAnalysis 阅读 →

AI 生成摘要 · Google Gemini · 来自 8 个来源。我们如何撰写摘要 →

Kimi K3 的海量规模尽管经过优化，仍需要密集的网络连接 · 跟踪 8 个来源

报道来源 [8]

X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

更高效的注意力机制将把上下文长度从100万推至500万以上，并减少上下文损耗。Jevons悖论意味着注意力机制的效率提升将...

More efficient attention will further push context lengths from 1M to 5M+, with less context rot. Jevons’ Paradox means that making attention more efficient will lead to wider AI adoption, which will require more networking. 8/8🧵
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

此外，通过增量KV缓存传输等优化，预填充只需传输非解码实例缓存的部分，因此KV传输

Furthermore, with optimizations like incremental KV-cache transfers, prefill only needs to transfer the portions cached by the non-decode instance, so KV transfer does not take up much networking bandwidth relative to WideEP, even before KDA. 7/8🧵
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

KV缓存预填充和解码之间的传输每轮只发生一次，而WideEP在每次输出前向传递中发生120多次，潜在输出为500多

KV-cache transfer between prefill and decode happens only once per turn, while WideEP happens 120+ times per output forward pass, with potentially 500+ output tokens per turn. Thus, the relative KV-cache transfer savings from KDA linear attention are dwarfed by the increase in
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

WideEP 在每一层和每一次前向传播中都需要进行两次分派和合并。这意味着，对于每一次前向传播，分派和合并将需要

WideEP dispatch and combine need to happen twice on every layer and for every forward pass. That means that, for each forward pass, dispatch and combine will need to run 120+ times. 5/8🧵
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

WideEP 是 Kimi K3 需要使用的、非常消耗网络资源的推理优化技术。WideEP 通过将 896 个专家模型分布到多块 GPU 上来实现大规模扩展

WideEP is a very network-intensive serving optimization that Kimi K3 will need to use. WideEP works by spreading 896 experts across many GPUs within a scale-up domain, then dispatching tokens to the correct GPU and combining the results. 4/8🧵 https://t.co/ZciwDfWfZR
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

因为Kimi K3拥有2.8万亿参数，即使在MXFP4下，每次前向传播也需要1.5TB的HBM带宽。这意味着，即使使用spec decode，服务

Because Kimi K3 has 2.8 trillion parameters, even at MXFP4, each forward pass will require 1.5 TB of HBM bandwidth. This means that, even with spec decode, serving it profitably at a reasonable level of interactivity requires aggregating many chips together over a high-bandwidth
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

虽然Kimi K3确实在其四分之三的层中使用Kimi Delta Linear Attention (KDA)，并且KDA将KV缓存传输带宽减少了多达10倍，但

While it is true that Kimi K3 uses Kimi Delta Linear Attention (KDA) in 3 out of every 4 layers and that KDA reduces KV-cache transfer bandwidth by up to 10x compared with comparable full global-attention models, the important missing piece is that Kimi K3 requires WideEP to http…
X — SemiAnalysis TIER_1 English(EN) · SemiAnalysis_ · 2026-07-18 21:35

与2025年1月的DeepSeek类似，Panicans可能认为AI网络交换机TAM将大幅萎缩，因为Kimi K3使用了KDA Attention，这减少了

Similar to DeepSeek in January 2025, Panicans may think that the AI networking switch TAM will massively shrink because Kimi K3 uses KDA Attention, which reduces KV-transfer networking bandwidth by up to 10x. But the opposite is true, as we explain below. 👇️ 1/8🧵 https://t.co/FNr…

报道来源 [8]

相关实体

相关话题