3 ms·

The module on the framework basically is an e-gpu except directly on PCIE instead of thunderbolt. Why they couldn't make it quick swappable I don't know.

CuriousCosmic 3y ago
Hotswappable PCIE is really hard to do right, especially if you are using a non-standard connector. You have to guarantee that pins connect and disconnect in the right order or you risk damaging equipment. And on top of that, while a connector might be perfectly safe with cold swap (i.e. power off), you can damage the connector if the circuit is completed and power starts flowing but you only have partial contact. So it was probably preferable to just release a coldswap versions and try dabbling with a hotswap capable chassis in the future.
- adgjlsfhk1 3y ago
  it seems like the right approach would be to have a switch that disconnects the wires while the GPU is being connected to ensure everything is connected physically before anything is connected electricly.
  - rtkwe 3y ago
    It's two 74 pin connectors that's way too many to physically switch. Also signal integrity through that many switches sounds like an absolute nightmare.
- Dylan16807 3y ago
  Connecting the ground first has been solved in many ways. Are there other significant issues here for pin order?
DeIlliad 3y ago
Speaking from experience, there is a big performance difference connecting your eGPU over PCIE and connecting over Thunderbolt. Usually people just take the performance hit with Thunderbolt because of convenience and they are usually using overpowered desktop GPUs but there are eGPU PCIE connectors for people who want the most performance. Because this is already using a mobile chip with lower performance, Framework probably didn't want to tank performance any more by going with the lower bandwidth Thunderbolt option.
- izacus 3y ago
  The performance hit is because Thunderbolt has at most PCIe x4 bandwidth, while internal GPUs are connected to PCIe x16 lanes. That's a massive loss of bandwidth. But it also tells you a lot about why Framework didn't make x16 bus hot swappable, it's a significantly bigger engineering challenge.

CuriousCosmic 3y ago

Hotswappable PCIE is really hard to do right, especially if you are using a non-standard connector. You have to guarantee that pins connect and disconnect in the right order or you risk damaging equipment. And on top of that, while a connector might be perfectly safe with cold swap (i.e. power off), you can damage the connector if the circuit is completed and power starts flowing but you only have partial contact. So it was probably preferable to just release a coldswap versions and try dabbling with a hotswap capable chassis in the future.

adgjlsfhk1 3y ago
it seems like the right approach would be to have a switch that disconnects the wires while the GPU is being connected to ensure everything is connected physically before anything is connected electricly.
- rtkwe 3y ago
  It's two 74 pin connectors that's way too many to physically switch. Also signal integrity through that many switches sounds like an absolute nightmare.
Dylan16807 3y ago
Connecting the ground first has been solved in many ways. Are there other significant issues here for pin order?

DeIlliad 3y ago

Speaking from experience, there is a big performance difference connecting your eGPU over PCIE and connecting over Thunderbolt. Usually people just take the performance hit with Thunderbolt because of convenience and they are usually using overpowered desktop GPUs but there are eGPU PCIE connectors for people who want the most performance. Because this is already using a mobile chip with lower performance, Framework probably didn't want to tank performance any more by going with the lower bandwidth Thunderbolt option.

izacus 3y ago
The performance hit is because Thunderbolt has at most PCIe x4 bandwidth, while internal GPUs are connected to PCIe x16 lanes. That's a massive loss of bandwidth. But it also tells you a lot about why Framework didn't make x16 bus hot swappable, it's a significantly bigger engineering challenge.