# 计算专题日跟踪 2026-09-03

## 摘要

**今日无新的 hypervisor 版本或 CPU/GPU 上市通稿落在 2026-09-03 当天。** 最强近期公开页：NVIDIA MIG-backed vGPU 文档把 Blackwell 上「MIG 硬件切片 + 时间片」写成可售配置（RTX PRO 6000 单卡最多 12 VM），并写明跨 MIG profile **不能热迁**、多数训练卡上 Windows 客机不能用 MIG-backed。VMware Explore（2026-08-31）AI Factory 通稿把 GPU 池化与 AMD MI350/ROCm 零触摸部署绑进 VCF。华为侧仍以 HC 2025 路标为准：鲲鹏 950 / TaiShan 950 在 2026 Q1 窗口、昇腾 950DT / Atlas 950 在 2026 Q4；FusionCube 1000H 规格页同期仍列鲲鹏 920。RHV 已于 2026-08-31 EOL。

规划即时含义：计算差异化从「有 HA」变成「调度器认不认 LLC/NVLink 域」；GPU 必须分直通 / MIG / 时间片三 SKU；不要把 Atlas 超节点写进 FusionCube HCI 料单。

## 原文摘录

### 1. MIG-backed vGPU 与 Blackwell 密度（NVIDIA 文档，现网）

- URL：https://docs.nvidia.com/ai-enterprise/release-8/latest/infra-software/vgpu/features/mig-backed-vgpu.html

> A Multi-Instance GPU (MIG)-backed vGPU combines MIG hardware partitioning with vGPU virtualization. Instead of sharing a full physical GPU, each vGPU runs on a dedicated GPU instance - a hardware-isolated slice of a MIG-capable GPU.
> Universal MIG technology on Blackwell enables both compute and graphics workloads to be consolidated and securely isolated on the same physical GPU.
> On the NVIDIA RTX PRO 6000 Blackwell Server Edition, up to four MIG slices can be created on a single GPU. Within each MIG slice, one to three time-sliced vGPUs for Compute with 8 GB frame buffer each can be created … enabling up to 12 virtual machines to share a single physical GPU
> MIG-backed vGPU is not supported for Windows guest VMs on NVIDIA A30, A100, A800, H100, H200, B200, and B300 GPUs.
> Live migration between different MIG profiles (for example, migrating a VM from a 1g.5gb instance to a 2g.10gb instance) is not supported. The target host must have a matching MIG configuration.

### 2. VMware AI Factory（2026-08-31 通稿）

- URL：https://www.globenewswire.com/news-release/2026/08/31/3353363/0/en/broadcom-announces-vmware-ai-factory-enabling-faster-time-to-production-ai-and-greater-control-over-ai-tokenomics.html

> VCF pools and shares GPU resources across the organization so teams can run multiple models on shared hardware instead of dedicating infrastructure to each workload.
> Broadcom and AMD are collaborating to deliver a VMware AI Factory that pairs VCF with AMD Instinct GPUs and the open AMD ROCm software ecosystem. Zero-touch provisioning will orchestrate the end-to-end deployment of the entire stack, from vSphere and vSAN through Kubernetes and the AMD GPU operator, and the AMD DVX driver can attach GPUs to large VMs consumed by a VMware vSphere Kubernetes Service cluster.

### 3. 华为：鲲鹏 950 / 昇腾 950 / 超节点（HC 2025 演讲，2026 上市窗口）

- URL：https://www.huawei.com/cn/news/2025/9/hc-xu-keynote-speech

> 2026年Q1，我们将推出Kunpeng 950处理器，包括两个版本，分别是：96核/192线程和192核/384线程；支持通用计算超节点；安全方面新增四层隔离，成为鲲鹏首颗实现机密计算的数据中心处理器。
> Ascend 950PR 采用了华为自研的低成本HBM，HiBL 1.0 … 这颗芯片将在2026年一季度推出
> Ascend 950DT 将在2026年Q4推出。
> TaiShan 950超节点，基于Kunpeng 950打造，全球首个通用计算超节点，其最大支持16节点，32个处理器，最大内存48TB，同时支持内存、SSD、DPU池化。
> Atlas 950超节点的上市时间是：2026年四季度。

### 4. FusionCube 1000H 现网仍是鲲鹏 920（华为产品页）

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

> 处理器 | 支持2个处理器• 支持处理器型号（鲲鹏920 7260/5250/5220）
> 虚拟机管理：创建虚拟机，虚拟机模板管理，迁移虚拟机

### 5. KubeVirt SR-IOV：直通虚机不能热迁（Network Operator 26.7，文档 2026-09-01）

- URL：https://docs.nvidia.com/networking/display/kubernetes2670/platform-support.html

> NVIDIA Network Operator supports attaching SR-IOV Virtual Functions (VFs) to KubeVirt and Red Hat OpenShift Virtualization virtual machines via VFIO PCI passthrough.
> Live migration is not supported for VMs that use SR-IOV VFIO passthrough.

### 6. XCP-ng：NVIDIA vGPU 仍无公开方案

- URL：https://docs.xcp-ng.org/compute/

> Due to proprietary XenServer code, XCP-ng doesn't have (yet) support for NVIDIA vGPUs.

### 7. 拓扑感知调度（NVIDIA 技术博文）

- URL：https://developer.nvidia.com/blog/running-ai-workloads-on-rack-scale-supercomputers-from-hardware-to-topology-aware-scheduling/

> Topology-aware scheduling: When a network topology is defined for a node pool, NVIDIA Run:ai applies topology-aware scheduling preferences to keep all pods in a distributed workload as close as possible, only expanding outward when necessary.
> ComputeDomain-backed workload placement: When a job is submitted, it automatically creates and attaches a ComputeDomain. Pods are placed on nodes that share NVLink connectivity and a correctly scoped IMEX domain.

### 8. Linux 7.2 Cache Aware Scheduling（报道）

- URL：https://www.kad8.com/software/linux-7-2-brings-cache-aware-scheduling-for-modern-cpus/

> Cache Aware Scheduling (CAS), a Linux scheduler enhancement primarily developed by Intel engineers over the past year … officially entered the TIP sched/core branch in May 2026 and is expected to merge during the Linux 7.2 merge window.
> Contemporary architectures such as AMD EPYC Turin and Intel Xeon 6 Granite Rapids are built around multiple chiplets, tiles, and independent LLC (Last Level Cache) domains.

## 出处

- https://docs.nvidia.com/ai-enterprise/release-8/latest/infra-software/vgpu/features/mig-backed-vgpu.html
- https://docs.nvidia.com/ai-enterprise/release-8/latest/infra-software/vgpu/reference/blackwell.html
- https://www.globenewswire.com/news-release/2026/08/31/3353363/0/en/broadcom-announces-vmware-ai-factory-enabling-faster-time-to-production-ai-and-greater-control-over-ai-tokenomics.html
- https://www.huawei.com/cn/news/2025/9/hc-xu-keynote-speech
- https://e.huawei.com/cn/products/storage/hci/fusioncube1000h
- https://www.guancha.cn/economy/2026_03_20_810815.shtml
- https://docs.nvidia.com/networking/display/kubernetes2670/platform-support.html
- https://docs.xcp-ng.org/compute/
- https://developer.nvidia.com/blog/running-ai-workloads-on-rack-scale-supercomputers-from-hardware-to-topology-aware-scheduling/
- https://www.kad8.com/software/linux-7-2-brings-cache-aware-scheduling-for-modern-cpus/
- https://aws.amazon.com/about-aws/whats-new/2026/02/amazon-ec2-nested-virtualization-on-virtual/
- https://documentation.suse.com/cloudnative/virtualization/v1.8/en/hosts/vgpu-support.html