698 points • 4 days ago • Article Link

如果你能把一块完整的桌面级 GPU 接到 MacBook Air 上会怎样？事实证明这是可行的。作者通过 Thunderbolt eGPU 扩展坞把 NVIDIA RTX 5090 连到 M4 MacBook Air，并在 macOS 上运行的 Linux 虚拟机中使用该 GPU 。这个项目涉及大量工程挑战——包括 PCI 直通、 DMA 映射的变通方案和 x86 仿真层——但最终让这套方案在游戏和 AI 推理上都能工作。

核心问题是 macOS 在 Apple Silicon 上没有对现代 NVIDIA 或 AMD GPU 的原生驱动。解决方法是让 ARM64 的 Linux 虚拟机接管并直通通过 Thunderbolt 连接的 GPU 。为此必须把 PCI 的基地址寄存器（BAR）映射到客户机虚拟机中。最初因为 QEMU 在使用 Apple 的 Hypervisor.framework 处理内存标志时的一个 bug，导致主机内核崩溃，修复方法是从设备内存映射中去掉 HV_MEMORY_EXEC 标志。至于直接内存访问（DMA），Apple 的 IOMMU 等价物 DART 施加了严格限制：总映射内存约为 1.5GB，且最多约 64k 个单独映射。为应对这些限制，作者实现了一个名为 apple-dma-pci 的虚拟 PCI 设备，拦截 DMA 映射调用，并用聚类方案把小的分配合并到更大的 256 KB 区域，从而显著减少映射数量。

其他性能优化还包括在 Apple Silicon 上启用硬件 Total Store Ordering（TSO）以通过 FEX-Emu 加速 x86 仿真、用 kprobes 修补 NVIDIA 驱动以应对 DART 的对齐限制，以及调整 QEMU 线程的调度优先级。即便如此，分层虚拟化带来的开销仍然不可忽视。基准显示，装在原生 PCIe 插槽中的相同 RTX 5090 的游戏 PC 比 MacBook Air 的 eGPU 方案快 2–4 倍。在 720p 等低分辨率下，由于仿真开销，M4 Air 的集成 GPU 反而比 eGPU 更快，但在 4K 下 eGPU 能把原本不可玩的帧率变为流畅画面。

在 AI 推理方面的表现尤其引人注目。运行 Qwen 3.6 时，eGPU 达到每秒 155 个 token，而 M4 Air 的集成 GPU 只有每秒 22 个 token，提升约 6.5 倍。更显著的是，对于 4K token 上下文的提示处理（prefill），耗时从 17 秒降到 150 毫秒，提速约 120 倍。 RTX 5090 在并发请求下的扩展性也更好，最多 4 路并发时几乎呈线性扩展。对于像 Gemma 4 这样的密集模型，性能差距更大，eGPU 的吞吐量大约是 Apple Silicon 集成显卡的 4 倍。

尽管技术上很出色，作者也承认这仍是个发烧友级别的方案。存在的稳定性问题包括 FEX-Emu 中 Steam 崩溃、 DMA 映射碎片化需要重置 GPU 、以及分发驱动所需的特殊 Apple 授权。该方案需要自行编译定制软件并加载内核扩展。作者正在与上游 QEMU 合作合并这些补丁，并希望未来在 Apple Silicon 上对 Linux 实现原生 Thunderbolt 支持能消除许多现有限制。目前这个项目证明了：一台 22W 的笔记本也能驱动一块 600W 的 GPU，但原生 PC 仍明显更快。 What if you could strap a full desktop GPU to your MacBook Air? Turns out, you can. The author set out to connect an NVIDIA RTX 5090 to an M4 MacBook Air using a Thunderbolt eGPU dock, running the GPU inside a Linux virtual machine on macOS. The project involved significant engineering challenges, including PCI passthrough, DMA mapping workarounds, and x86 emulation layers, but ultimately succeeded in making the setup functional for both gaming and AI inference.

The core technical challenge was that macOS lacks native drivers for modern NVIDIA or AMD GPUs on Apple Silicon. The solution involved running Linux in an ARM64 VM and passing through the Thunderbolt-connected GPU. This required mapping PCI Base Address Registers (BARs) into the guest VM, which initially caused host kernel panics due to a bug in how QEMU handled memory flags with Apple's Hypervisor.framework. The fix involved removing the HV_MEMORY_EXEC flag from device memory mappings. For Direct Memory Access (DMA), Apple's IOMMU equivalent called DART imposed strict limits of about 1.5GB total mapped memory and 64k individual mappings. To work around this, the author created a virtual PCI device called apple-dma-pci that intercepts DMA mapping calls and uses a clustering scheme to reduce the number of individual mappings by grouping small allocations into larger 256kB regions.

Additional performance optimizations included enabling hardware Total Store Ordering (TSO) mode on Apple Silicon to accelerate x86 emulation through FEX-Emu, patching the NVIDIA driver via kprobes to handle DART's alignment constraints, and adjusting QEMU's thread scheduling priorities. Despite these improvements, the layered virtualization approach still incurred significant overhead. Benchmarks showed that a gaming PC with the same RTX 5090 in a native PCIe slot was 2-4x faster than the MacBook Air eGPU setup. At lower resolutions like 720p, the M4 Air's integrated GPU actually outperformed the eGPU configuration due to emulation overhead, but at 4K the eGPU transformed unplayable framerates into smooth gameplay.

The most compelling results came from AI inference benchmarks. Running Qwen 3.6, the eGPU setup achieved 155 tokens per second compared to 22 tokens per second on the M4 Air's integrated GPU, a 6.5x improvement. More dramatically, prompt processing (prefill) for a 4K-token context dropped from 17 seconds to 150 milliseconds, a 120x speedup. The RTX 5090 also scaled much better with concurrent requests, nearly linear scaling up to 4 concurrent streams. For dense models like Gemma 4, the performance gap was even more pronounced, with the eGPU delivering roughly 4x the throughput of Apple Silicon's integrated graphics.

While the project demonstrates impressive technical achievement, the author acknowledges it remains firmly in the hobbyist category. Stability issues include Steam crashes in FEX-Emu, DMA mapping fragmentation requiring GPU resets, and the need for a special Apple entitlement to distribute the driver. The setup requires building custom software and loading kernel extensions. The author is working with upstream QEMU to integrate the patches and hopes future improvements like native Thunderbolt support in Linux on Apple Silicon could eliminate many of the current limitations. For now, it stands as a proof of concept that a 22W laptop can harness a 600W GPU, even if a native PC remains significantly faster.