External GPUs (eGPU) have been tried time and again, and while they have found their niche among laptop users, they're not commonplace. Now that the AI craze is in full swing, more people are looking into owning a graphics card of some sort for the single purpose of running LLMs locally. The WiCi One is designed with that specific scenario in mind, but adds wireless connectivity, multi-device support, and onboard storage.
That list of words usually doesn't belong in the same sentence as GPUs, so it merits close inspection. First and foremost, the WiCi One leverages the WiFi 7 protocol to transmit data to and from the card. The firm says that client machines keep a GPU driver that makes the operating system think it's a graphics card just like any other, while in reality they'll be talking to the WiCi One over the network.
However, as WiCi explains in great detail, just blindly doing that would make for a terrible eGPU, as the standard communication between host and card is extremely chatty. That's not a problem with the sub-microsecond latency of PCIe 5.0, but would quickly turn even basic status querying operations into multi-second affairs. Large computations with lots of data shuffled back and forth could hypothetically take hours.
The company claims its driver software performs a great deal of data caching, deduplication, and compression, and is also very judicious about what type of hardware requests can be handled locally versus those that really do need to go over the airwaves.
Source link







