# 专题日跟踪 · ASIC 托管 / 虚拟化 · 2026-09-03

> 日期：2026-09-03（Asia/Shanghai）。原文摘录保持源语言。不改写数字。本日检索无 2026-09-03 当天的 TPU / Gaudi / Neuron / 昇腾 / Groq HCI 发布。

## 摘要

- **今日无新的 ASIC 托管或 HCI 加速器 HCL 通稿落在 2026-09-03 当天。**
- 最强近期公开页仍是：**Intel Gaudi 官方虚拟化文档**（仅 PCI passthrough，明文无 SR-IOV / MIG，最小粒度单 HPU）+ **VMware VCF 9.0 Gaudi 3 Enhanced DirectPath I/O**（2025-08-07 博文，驱动页覆盖 ESXi 8.0U3 / 9.1）。
- Google 把 TPU 定义为自研 **ASIC**，消费面是 Cloud TPU VM / GKE slice；自定义 OS 文档（2026-08-26 更新）要求客机把 TPU 绑到 `vfio-pci`。TPU DRA 仍须申领节点上全部芯片。**无花名册 HCI 的 TPU HCL。**
- AWS Neuron DRA（2026-05-18）：Trainium 与 EFA 拓扑对齐；NeuronCore 是运行时切分，**不是** 第三方 hypervisor 的 PCIe SR-IOV。无本地可订板卡。
- 昇腾：Atlas 虚拟机指南支持 NPU 整卡直通与 vNPU 模板；FusionCube 1000H 规格可订 Atlas 300I Pro/300V/300V Pro。**Atlas 超节点不得写入 1000H / FusionOne。**
- SmartX（2025-04）超融合虚机直通昇腾推理卡 + MindIE；ZStack AIOS 手册列 Atlas 300I Pro、910B3/4（ARM 透传）、Hygon DCU、昆仑 P800、寒武纪等。深信服 AICP **厂商自称** 兼容昇腾/海光等，不是 aSV 级 HCL。
- Groq / Cerebras / SambaNova：机架一体机，花名册 HCI **全部无公开方案**。Cerebras CS-4 Nexus 报道日期 2026-08-19。
- 规划即时含义：HCI 全球盘仍 GPU 中心；中国区可产品化的是 Gaudi 直通对等、昇腾 300I 直通/vNPU、海光 vDCU。TPU/Trainium/LPU 答「非 HCI 形态」。

## 原文摘录

### 1. Gaudi：只有 PCI 直通（Intel 文档，现网 v1.24）

- URL：https://docs.habana.ai/en/latest/Virtualization/Configuring_VMs_on_Gaudi.html

> PCI passthrough is the only virtualization mechanism supported by Gaudi accelerators. There is no support for SR-IOV or MIG. The smallest granularity possible is with a single HPU.
>
> You can now enable Intel® Gaudi® PCI passthrough on a virtual OS using a Linux host server. This document describes how to allocate Intel Gaudi AI accelerator for KVM Guests on Ubuntu 22.04.5 LTS, Ubuntu 24.04.2 LTS and RHEL 9.4.

### 2. VMware：Gaudi 3 在 VCF 9.0 直通（2025-08-07）

- URL：https://blogs.vmware.com/cloud-foundation/2025/08/07/accelerating-ai-workloads-with-intel-gaudi-3-on-vmware-cloud-foundation-9-0/

> The latest release of VMware Cloud Foundation now supports Intel Gaudi 3 AI accelerators, enabling enterprises to run large-scale AI workloads within a fully virtualized environment.
>
> To streamline AI infrastructure management with Gaudi, Intel offers enhanced passthrough capabilities via its Enhanced DirectPath IO driver. When installed, such driver enables direct access to Gaudi 3 telemetry through the vCenter plugin, without compromising the isolation and flexibility of virtual machines. It also enforces device grouping, allowing multiple Gaudi 3 accelerators to be logically combined and presented as a single virtual instance.

驱动下载页（HL-338，ESXi 8.0U3 与 9.1）：https://www.intel.com/content/www/us/en/download/860900/passthrough-driver-for-intel-gaudi-ai-accelerators-for-vmware-esxi.html

> This software utilizes VMware* vSphere Enhanced DirectPath I/O in conjunction with VMware* vSphere Device Groups to passthrough grouped Intel® Gaudi® AI Accelerators to a virtual machine (VM), enabling hardware telemetry monitoring via VMware* vCenter.

### 3. TPU 是 ASIC；客机绑 vfio-pci（GKE / Cloud TPU 文档，2026-08-26 更新）

- URL：https://docs.cloud.google.com/kubernetes-engine/docs/concepts/tpus

> TPUs are Google's custom-developed, application-specific integrated circuits (ASICs) for accelerating ML workloads that use frameworks such as TensorFlow, PyTorch, and JAX.
>
> The number of requested `google.com/tpu` must be equal to the total number of available TPU chips on the TPU slice node. Any container in a GKE Pod that requests TPUs must consume all the TPU chips in the node.

- URL：https://docs.cloud.google.com/tpu/docs/custom-os-image

> To allow the guest OS to access TPU hardware, you must bind TPU devices to the vfio-pci driver.
>
> `intel_iommu=on,sm_on`: Enables Intel Input-Output Memory Management Unit (IOMMU). Required for TPU7x and v5p architectures.

### 4. Neuron DRA：Trainium 在 EKS，不是 HCI SR-IOV（2026-05-18）

- URL：https://aws.amazon.com/blogs/containers/simplify-ai-infrastructure-for-aws-trainium-and-elastic-fabric-adapter-with-kubernetes-dynamic-resource-allocation/

> The Neuron DRA driver extends these benefits to AWS Trainium accelerator management by handling device allocation and per-workload configuration for Neuron devices.
>
> The EFA and Neuron DRA drivers are installed through Helm. These DRA drivers are recommended for new deployments on Amazon Elastic Kubernetes Service (Amazon EKS) clusters running Kubernetes version 1.34 or later … You can’t run DRA drivers on the same nodes as corresponding device plugins.

### 5. FusionCube 1000H 可订 Atlas 300I/300V（华为产品页；本抓取遇 403，规格以检索/既有核验为准）

- URL：https://e.huawei.com/cn/products/storage/hci/fusioncube1000h

> IO拓展 | 最多支持8个PCIe4.0 接口• 最多可支持8张全高半长PCIe卡• 最多可支持4张全高全长PCIe卡• 可支持Atlas 300I Pro/300V/300V Pro说明：根据不同配置，支持不同的卡和硬盘数量
>
> 处理器 | 支持2个处理器• 支持处理器型号（鲲鹏920 7260/5250/5220）

Atlas vNPU 直通（企业文档，与超节点硬件指南同册，**不等于** 1000H 可订超节点）：https://support.huawei.com/enterprise/zh/doc/EDOC1100568653/bf80825c

> 本章节以 Atlas 800 训练服务器（型号：9000） openEuler 20.03 LTS系统为例说明如何将vNPU（虚拟NPU）直通虚拟机

### 6. SmartX 昇腾直通（2025-04）

- URL：https://www.smartx.com/blog/2025/04/smartx-ai-ascend/

> 目前，SmartX AI 基础设施已全面兼容昇腾生态，支持用户基于虚拟机、虚拟化 Kubernetes 和裸金属 Kubernetes 环境部署使用昇腾 NPU 与 MindIE 平台
>
> 用户可以在 SmartX 超融合虚拟机环境中，通过昇腾推理硬件直通功能和 MindIE 镜像方式，实现昇腾 NPU 与 MindIE 的快速部署

（本抓取对该博文超时，摘录依据检索摘要；规划引用须再打开原文。）

### 7. Cerebras 仍是晶圆级一体机（2026-08-19）

- URL：https://www.nextplatform.com/compute/2026/08/19/cerebras-overclocks-wse-3-waferscale-engine-to-boost-inference-oomph-in-nexus-cs-4/5289400

报道口径：CS-4 “Nexus” 超频 WSE-3 Turbo，GA 目标 2026 Q3 晚些时候；背包式液冷机架。形态不是 HCI 节点。

## 对专题档案的影响

- asic.md 必须把 NVIDIA GPU 留在对照列，Gaudi / TPU / Neuron / 昇腾 / DCU 分列。
- FusionCube 只写 300I/300V 可选；超节点与 FusionOne 分料。
- 18 家里整行无 ASIC SKU 的写进对照表，不要用「异构加速器」补格子。

## 出处

- https://docs.habana.ai/en/latest/Virtualization/Configuring_VMs_on_Gaudi.html
- https://blogs.vmware.com/cloud-foundation/2025/08/07/accelerating-ai-workloads-with-intel-gaudi-3-on-vmware-cloud-foundation-9-0/
- https://www.intel.com/content/www/us/en/download/860900/passthrough-driver-for-intel-gaudi-ai-accelerators-for-vmware-esxi.html
- https://docs.cloud.google.com/kubernetes-engine/docs/concepts/tpus
- https://docs.cloud.google.com/tpu/docs/custom-os-image
- https://aws.amazon.com/blogs/containers/simplify-ai-infrastructure-for-aws-trainium-and-elastic-fabric-adapter-with-kubernetes-dynamic-resource-allocation/
- https://e.huawei.com/cn/products/storage/hci/fusioncube1000h
- https://support.huawei.com/enterprise/zh/doc/EDOC1100568653/bf80825c
- https://www.smartx.com/blog/2025/04/smartx-ai-ascend/
- https://docs.zstack.io/zh/docs/ai/aios/cloud-user-guide/topics/%E8%B5%84%E6%BA%90%E4%B8%AD%E5%BF%83/%E7%A1%AC%E4%BB%B6%E8%AE%BE%E6%96%BD/gpu%E8%AE%BE%E5%A4%87
- https://www.xfusion.com/cn/product/hyperconverged-solutions/hci
- https://www.nextplatform.com/compute/2026/08/19/cerebras-overclocks-wse-3-waferscale-engine-to-boost-inference-oomph-in-nexus-cs-4/5289400