There's a surprising amount of argument over whether a small number of network connections means that GPU miners must be possible.
The number of network connections a particular client has is NOT an indicator of how many people are mining. I can spin up 100 servers with CPUMiner installed, and mine to one server, which would appear as one connection on the network.
Additionally, not everyone on a P2P network is necessarily connected to everyone else. If you are behind a firewall, then you can't directly connect to another peer behind a firewall. Sure, things like NAT punchthrough can help. Personally, I have 96 connections on an argon2 daemon running on an unfirewalled server, and 15 connections on one running inside a strict firewall.
The total network hash rate divided by the number of peers your client can reach isn't an estimation of the average mining speed--some nodes aren't mining at all, some are serving as private pools for tons of miners which don't appear as individual nodes on the network.
As for the feasibility of GPU mining, it certainly is possible. In fact, there's even an OpenCL implementation of Argon2d online. There are two versions of the Argon2 hashing algorithm: Argon2d and Argon2i. Argon2i is designed for password hashing (where side-channel timing attacks are a problem), while Argon2d is designed for uses like cryptocurrencies, focusing primarily on working well on CPUs compared to GPUs. Here's a quote from the specifications whitepaper:
Argon2 has two variants: Argon2d and Argon2i. Argon2d is faster and uses data-depending
memory access, which makes it suitable for cryptocurrencies and applications with no threats from side-channel
timing attacks. Argon2i uses data-independent memory access, which is preferred for password hashing and
password-based key derivation. Argon2i is slower as it makes more passes over the memory to protect from
tradeoff attacks.
There is an OpenCL implementation of Argon2d available in John the Ripper. Some benchmarking (
http://comments.gmane.org/gmane.comp.security.phc/3303) shows that extremely powerful GPUs (like the NVidia Titan) clock in below the speed of a 4770k. A Titan x is marginally faster than the 4770k. This doesn't, of course, mean that a faster implementation (on either platform, CPU or GPU) is impossible, but preliminary OpenCL code written by people who certainly know what they're doing shows that GPUs aren't great at crunching Argon2d.
It probably wouldn't be particularly hard for a GPU miner to be written given an OpenCL implementation of Argon2d already exists, but the benefits of such a GPU miner probably wouldn't be spectacular.
