WiCi: Connecting GPUs over WiFi | Hacker News Reader

WiCi: Connecting GPUs over WiFi

wici.ai

18 pointsshubham020411 comments

For anyone, if not everyone, that finds site unintuitive, a technical report is also available[0]. Fwiw some other prior work, not cited in previous report, on remote GPU offloading is GVirtuS[1], which originated for use on HPCs[2] but later pivoted in low-power devices[3], and patents by Lenovo[4] and Nvidia[5] (abandoned; imo quite vague but worths showing that it's something that has existed as an idea).

That said, although title here focuses on one, WiCi seems to pursue a second related idea, that being SSD-backed LM inference. Code for this is open-sourced at https://github.com/wici-ai/inferssd.

[0]: https://arxiv.org/abs/2608.24204 [1]: https://github.com/gvirtus/GVirtuS [2]: https://link.springer.com/chapter/10.1007/978-3-642-15277-1_... [3]: https://onlinelibrary.wiley.com/doi/abs/10.1002/cpe.4286 [4]: https://patents.google.com/patent/US20250191108A1 [5]: https://patents.google.com/patent/US20140122558A1

I recently came across WiCi, a wireless GPU that connects to the host computer over WiFi. The host computer sees the wireless GPU as a native GPU connected to the computer and also shows up in the Windows Task Manager.

Demonstrations show it running DeepSeek and some graphics-intensive games. The GPU being shared over WiFi is an RTX 5090.

I wanted to dig deep into how WiFi as a protocol can be used for communicating with the GPU. In a normal situation, the GPU will communicate via PCIe with the CPU of the host computer, ensuring high-bandwidth data transfers. I found the following lines in a blog that WiCi has shared:

> So WiCi keeps the local-GPU contract while shrinking what actually crosses the link: cache stable facts, deduplicate repeated payloads, compress or stream large transfers, pipeline independent work, keep expensive state resident, and preserve true dependencies exactly.

What I feel is that WiCi has introduced an intermediate layer between the host and the GPU that facilitates communication. The commands intended for the GPU are first encoded in some intermediate language, passed through the network, and decoded back within the WiCi remote GPU (the external box holding the 5090) into commands that the CUDA driver understands.

WiCi does not seem to replace the physical PCIe protocol with IEEE 802.11 (WiFi) protocol over the physical medium. This is my assumption and I am not sure about it.

But if this product proves what it claims, then this would be a huge deal in my opinion. A shared GPU that different devices can use would be great for local LLMs.

Thanks. Site doesn't make a good job telling people what it's about. Though understandable to not want to say much, considering it's a commercialized tech using a proprietary protocol.

Feels like it could be potentially fairly slow

Nah. They are working on GPUs connected over RS-232

It's a wifi wrapper for an RTX GPU, they built a company around this?

I clicked the link: https://wici.ai/

I scrolled through the resulting cruft with all of the attention that its full-screen headings commanded. I did not watch the video.

I have no idea what the concept aims to improve, why it is useful, or how it is an advantage over any other concept.

Help?

Some of the things I have managed to find.

- It's GPU-over-WiFi. It shows up as a native, connected GPU on the client computer. I have no idea what drivers are required for this, or which OSs are supported

- It supports multiple users - i.e. multiple client devices can be connected to the same WiCi host device at the same time. I have no idea how it manages load across noisy neighbours.

- It can run DeepSeek V4 Flash. I don't know how well.

Maybe other things less relevant that I can't recall.

It seems like an interesting concept and product. I don't know how well it holds up, and it would be a very expensive gamble!

> I have no idea what the concept aims to improve, why it is useful, or how it is an advantage over any other concept.

The sell was "have a box full of compute, and a box full of GPUs, and mix-n-match at provision time." This stuff was directed at cloud providers, so you could let a customer move a slider to pick a core count and a gpu count and spin up a virtual machine without having to rewire any actual PCIe connections (n.b. this was before CXL existed).

Now they're trying to sell it on-prem as a box you throw on your network and provide GPU compute to your machine the way you provide storage with a NAS. Why this instead of just plugging in an eGPU? I dunno... wireless?

>[...] the way you provide storage with a NAS. Why this instead of just plugging in an eGPU? I dunno... wireless?

Why use a NAS rather plugging in an HDD enclosure? Same gains, same losses.

ignoring the real reasons that necessitate a big thick link to these cards for most real work, at that cost I see no benefit to this over a mini-pc+full size gpu hosting these services with a well known non-custom backend; a normal computer + a gpu.

yeah yeah yeah "it's done for you" but that's way less of a selling point post-agentic AI.

Reply on news.ycombinator.com