惯性聚合 高效追踪和阅读你感兴趣的博客、新闻、科技资讯
阅读原文 在惯性聚合中打开

推荐订阅源

Martin Fowler
Martin Fowler
WordPress大学
WordPress大学
月光博客
月光博客
OSCHINA 社区最新新闻
OSCHINA 社区最新新闻
大猫的无限游戏
大猫的无限游戏
让小产品的独立变现更简单 - ezindie.com
让小产品的独立变现更简单 - ezindie.com
博客园 - 聂微东
Apple Machine Learning Research
Apple Machine Learning Research
钛媒体:引领未来商业与生活新知
钛媒体:引领未来商业与生活新知
雷峰网
雷峰网
小众软件
小众软件
酷 壳 – CoolShell
酷 壳 – CoolShell
博客园 - 叶小钗
美团技术团队
宝玉的分享
宝玉的分享
Hugging Face - Blog
Hugging Face - Blog
阮一峰的网络日志
阮一峰的网络日志
A
About on SuperTechFans
Jina AI
Jina AI
D
Docker
Last Week in AI
Last Week in AI
MongoDB | Blog
MongoDB | Blog
Stack Overflow Blog
Stack Overflow Blog
Microsoft Azure Blog
Microsoft Azure Blog

Blog on KittenLabs

Marantz UD5007 power supply replacement GPD Pocket 4 Speaker DSP Extreme Pi Boot Optimization WiFi auth with OsmoHLR/SIM cards Windows 11 tweaks & usability improvements Using custom domains as a Fediverse redirect Manage RDP certificates on Windows using SSH Reviving a dead Gigabyte MJ11-EC1 mainboard Dead TROTEC PAC2000S air conditioner PlutoSDR standalone ADS-B FR24 feeder Micro8088 build log OpenWRT als WireGuard-Appliance GPS locked Raspberry Pi Es'hail-2 transceiver setup ThinkPad PowerSeries 820 PlutoSDR clock input WiFi over satellite TV coax
25GBit/s on macOS & iOS
By Manawyrm, tSYS and Ronja | Friday, May 17, 2024 · 2024-05-17 · via Blog on KittenLabs

Extreme network speeds on macOS/iOS devices using ConnectX network cards via USB4/Thunderbolt

macOS speed test

iOS speed test

Background info

Modern versions of macOS and iOS (starting with macOS Ventura/13) ship with NVIDIA/Mellanox ConnectX network drivers. AppleEthernetMLX5 is included by default and can be used with any NVIDIA/Mellanox ConnectX-4, 5 and 6 network card:

ConnectX-4 cards provide dual port 25 GBit/s connectivity (SFP28) and newer cards can go up to 200 GBit/s.
The maximum throughput is limited by the Thunderbolt/USB4 interface (to around 40 Gbit/s).

Photo of a Mac screen, showing Speedtest.net app with 20467 MBit/s download, 5670 MBit/s upload

We were even successful in connecting a ConnectX-4Lx NIC to an iPad Pro (M1):

iPad, connected via Thunderbolt to a ConnectX-4Lx network card, with SFP28 DAC, showing 10322/9359 MBit/s on the Speedtest.net app

Hardware setup

Thunderbolt/USB4 NVMe storage enclosures can be (ab)used as an external PCIe enclosure using a M.2 to PCIE x4 adapter. These enclosures are typically used to connect fast NVMe SSDs to a computer, but the used Thunderbolt 3 controller ICs (like Intels JHL7440) will just accept Thunderbolt and turn it into 4 PCIe lanes. These lanes can then be connected to any PCIe device.

Our tests were done using a TBU401 enclosure (~100$):
ConnectX-6 network card in a PCIe->M.2 adapter in a TBU401 Thunderbolt enclosure TBU401 internal PCB, JHL7440 IC, RTL9210B IC, ASM1480 IC, many other random storage and power ICs

Several vendors (on AliExpress, etc.) also offer little PCIe docks, like this “TH3P4 Lite GPU Dock”:
External GPU adapter with ConnectX-6 NIC

Full eGPU-enclosures (typically used for graphics cards) will also work well (but are large and very expensive).

Power supply

PCIe slots provide 2 voltage rails: 12V and 3.3V.
USB-C (without PD) provides 5V.

NVMe enclosures will convert the 5V supply down to 3.3V, but the 12V supply is still missing (and required for ConnectX cards).

Maximum available power on a USB-C port: 15W (5V/3A).

  • ConnectX-4 card: ~11W
  • Thunderbolt chipset: ~3W
  • Conversion losses (5V->12V step-up): ~2W

The total power of this setup will exceed the maximum available power on a USB-C port.

Powering such a setup without using an external power brick won’t be easily possible.
We used an external 12V/2.5A power supply for our tests.

GPUs on a Mac?

This same setup could also be used to connect an external GPU to a computer.
This will work well on Intel/x86 Windows and Linux machines.

There is no hope of GPU support on (current-gen) Apple Silicon, though. Apart from the fact that no drivers are available on macOS, even projects like Asahi Linux run into a hardware limitation with the way memory mapping works on M-series chips.

This is the same limitation also present on other Arm platforms like the Raspberry Pi. See PCIe problems on embedded systems.