Upgrade path eGPU

I have worked with Thunderbolt 3 eGPU for a while, it has its drawbacks namely throughput. Thunderbolt only has 4x PCIe lanes going to it.

If there could be a direct PCIe to the mainboard supporting 8x-16x PCIe lanes with an elegant expansion riser outside the laptop would be awesome. If there is a plan for AMD processors in the future this could be a path that could be included and possibly one of the first good options for the eGPU community since there has not yet been a AMD laptop that has Thunderbolt.

What my suggestion is.

Mainboard PCIe > PCIe Connector Adapter > Cable > PCIe Riser Bay (This could be a custom bay that includes powersupply)

Course PCIe is generally not hot-swappable so there would be some needed documentation explaining how this feature would work.


Fantastic idea. Thanks for sharing. I hope we can see this as a feature. If so, Framework will quickly become a Razer killer.


I’m unsure how we could integrate it as we can only fit a 1x slot on the expansion card slot in terms of size, not even thinking about all the cabling.

1 Like

I think that’d be another plus. If Framework gives eGPU support, they can advertise it as a workable everyday laptop, then you can bring it home, jack in, and game :smiley:

Heck, throw in some RGB component addons lol

In response to Josh_Cook statement. “I’m unsure how we could integrate it as we can only fit a 1x slot on the expansion card slot in terms of size, not even thinking about all the cabling.”
It is already achievable for a 4x slot. Some m.2 slots are 4x. It would be a custom slot to achieve 8x or 16x.


Congratulations on the project. I am not an engineer but I was wondering if it was possible to add, by moving some components here and there and by creating some new traces with a sharp knife :), a mezzanine connector on the lower portion of the motherboard (take a look at the attached image - box in red). The connector would be accessible on the bottom of the laptop. The latter would offer the possibility of connecting an external docking station. Only data would be conveyed through the connector while the docking station would provide the necessary power. The docking station should in fact allow the connection of a PSU and in particular of the 24-pin connector (if I remember correctly a pcie slot provides a maximum of 75W). The PSU would turn on automatically when the docking station is connected to the laptop (some pins of the mezzanine connector could act as a passive bridge and short the 4th and 5th pins of the 24-pin connector). The docking station would finally offer a PCI-e 16X slot: users, using a riser cable, can therefore connect their GPU. The docking station would also allow you to raise the laptop (favoring heat dissipation). It may also be possible to pair some low-profile fans (powered by the PSU’s 24-pin connector). The RPM could be adjusted through a potentiometer or through a PWM signal conveyed by the mezzanine connector. I hope I haven’t said too much bullshit :). Greetings from Italy!


alarico’s idea is pretty great. it would be like the old style thinkpad docking mechanism. So a future frame work motherboard upgrade could support that type of bottom connector option (that you would need to pair with a new bottom chassis upgrade) but after that you would now have your 16x PCIe connector. Pretty sweet.


For what it’s worth, the TB 4.0 ports already support 32gb of bandwidth which is the same as PCIe x16 3.0 as well as 100W of charging. The older docks were limited to 16gb or less, so the performance difference between native pci and TB->PCI should be pretty similar. No one has a TB4 GPU dock yet, so we’ll see.
While the framwork isn’t certified as TB4 yet, you can fire up that thunderbolt app and see for yourself that they are indeed TB4 devices with PCIE Tunneling enabled.
Your dream of a dock connector may already be installed :slight_smile:


Only around half of TB’s bandwidth can be used, as it only tunnels PCIe 3.0x4 iirc. The rest of the bandwidth is reserved for Display Port signalling from what I’ve read.

I believe it’s a 40GB max with 8GB used for video and 32GB available for PCI and all 4 ports have to support those data rates. I don’t have a TB4 device so there is a solid chance I’m wrong, but this article helped to clear some things up for me. Thunderbolt 4 Explained | Tripp Lite).

Something that may be of interest could potentially be using MXM, externally :slight_smile:

First, think of the enclosure being a bit smaller than an old external ZIP drive, when you consider the card, heat sink and fan.

The GPU would plug, right into the enclosure with an MXM connector, from there it would have cables connect to the motherboard and it’s eGPU port.

So anytime a GPU becomes obsolete, you buy a new card from Framework, and plug it into the previously mentioned enclosure.

1 Like

Yes exactly!

I think you got a bit confused with the units of measure. A TB 4.0 guarantees speeds of 32Gb/s (lowercase “b”) in both directions for PCIe link. That is 4GB/s (uppercase “B”) which is the performance of a PCIe 3.0 X4 slot. A PCIe 3.0 X16 slot instead allows a speed of 16GB/s in both directions (so 32GB/s in full-duplex).

I remain of the opinion that a native PCIe slot is the best solution for several reasons

Effective Bandwidth. The speeds mentioned above are “raw speeds”. TB technology uses 64b/66b encoding (66 bits are required to transfer 64 bits of actually useful data) with an overhead of approximately 3%. The PCIe protocol also has an overhead of about 1.5% (128b/130b encoding). I believe it is inefficient to encapsulate PCIe (which already has its own overhead) through communication that is double “inefficient”.

Latency. Take a look at the diagram at the bottom. A Tiger Lake-H processor offers 20 PCIe 4.0 lanes and 24 PCIe 3.0 lanes managed through the Platform Controller Hub (PCH). Current processors (UP3-class) used by the framework team offer 12 HSIO lanes (PCH) + 4 PCIe 4.0 (CPU). With a native PCIe slot the path would be CPU → GPU or worst case CPU → PCH → GPU. Using TB, CPU → Enclosure TB controller → GPU. In my opinion a PCH is more efficient than any external controller. Clearly, passing through PCH, it is possible to obtain the performance of a PCIe 3.0 X8 slot: DMI (Direct Media Interface) represents a bottleneck.

Costs and future expandability. When TB 5.0 is released you will need to update both the motherboard of your framework laptop and your eGPU enclosure. When PCIe 5.0 is released you will need to replace only the laptop motherboard instead.

I apologize for my bad English :slight_smile:
I hope I made myself clear!



ahhh good catch. Not the first time I’ve been wrong because of a little “b” :wink:

1 Like

I am very happy to see this suggestion gaining some traction.
I had forgotten the MXM connector the old Thinkpad Dock station had.
That would be nice to have a Dock station with that kind of design.
Turn off the computer. Dock it in. Turn it on and have the GPU, or additional PCIe riser cards in the dock station.