VR2-0233
① SA Source
- Source: 開啟完整 SA 文章
- Section:
BlueField-4 - Line hint:
325
Context Before
Speaking of BlueField-4, it is important to discuss the new offering Jensen highlighted at CES earlier this year: ICMS, or Inference Context Memory Storage — a platform that we hear may be rebranded to “CMX” at GTC. ICMS, or CMX, introduces a third, entirely separate network dedicated solely to context memory. CMX is a purpose-built KV cache fabric. As long-context inference pushes context windows toward millions of tokens and agentic concurrency scales across users and services, the current memory hierarchy used to store KVcache begins to look insufficient.
KV cache grows linearly with sequence length and multiplicatively with workload parallelism, quickly expanding beyond what any single tier of memory was designed to hold. GPU HBM, while unmatched in bandwidth and latency, is not enough on its own to store KV especially for longer sequence length queries that are becoming popular between turns or tool calls. Host DRAM extends capacity but remains node-bound and limited in aggregate footprint and ultimately has limited capacity. Meanwhile, traditional shared storage—architected for durability rather than latency —has more access time and power overhead, making it unsuitable for participation in the decode loop.
Evidence
Wherever networking is used in this architecture, it is provisioned as a context memory network — isolated from general data movement and optimized for predictable decode latency
Context After
The challenge with this is that the volumes of SSDs going to ICMS / CMX are quite overblown by the industry. We worked through the math in the Memory model ↗ and Tokenomics model ↗.
BlueField-4 will be the silicon anchor of this third network. Positioned on the storage array, it terminates NVMe-oF and RDMA traffic at line rate and manages KV movement independently of host CPUs and GPUs. With 2×400G SerDes links providing 800Gb/s of bandwidth, integrated Grace CPU, and LPDDR, BlueField-4 would act as the controller for a distributed context memory fabric. In a preferred DGX-style configuration, a single BlueField-4 per tray may serve four Rubin processors, with the DPU dedicated purely to KV cache traffic and not shared with generic storage I/O.
② Atomic Claim
在此架構中，只要使用 networking，就會配置成 context memory network，與一般 data movement 隔離，並針對可預測的 decode latency 最佳化。
- Epistemic Mode:
ASSERTED - Mapping Status:
COMPLETE
③ Semantic Frame
{
"attribute": "LATENCY",
"context_nodes": [],
"entity": {
"id": "04_knowledge_base/Decode",
"label": "decode"
},
"frame_type": "ATTRIBUTE",
"qualifiers": {
"condition_text": null,
"numeric_mentions": [],
"temporal_mentions": []
},
"value": {
"numeric_mentions": [],
"value_text": "在此架構中，只要使用 networking，就會配置成 context memory network，與一般 data movement 隔離，並針對可預測的 decode latency 最佳化。"
}
}④ Canonical Entity Mapping
| Role | Surface Label | Canonical Target |
|---|---|---|
| entity | decode | Decode |
⑤ Human Review
請在 Properties 逐項確認：
- 原文 → Atomic Claim 是否忠實
- Atomic Claim → Semantic Frame 是否忠實
- Canonical Entity mapping 是否正確
- Epistemic mode 是否保留原文語氣
- 最後選擇
review_action
Review state
Markdown 內文不是正式 approval。只有 Apply bridge 寫入的 Decision Ledger event 才是正式決策。