zubairdroid

```rust use std::net::Ipv4Addr; use pnet::{packet::{arp::{ArpHardwareType, ArpOperation}, ethernet::EtherType}, util::MacAddr}; fn main() { // debug options std::env::set_var("RUST_BACKTRACE", "1"); // get all the interfaces and everything let interfac = pnet::datalink::interfaces()[3].clone(); println!("the interface name...

Anything at this point Just access to the the hardware will be enough its a shame really that taichi doesnt support it So much could be done if it did...

No need more granular control plus it would be better as a whole if we needed to depend less on outside packages given the nature of taichi

cuBLAS would be a better fit So whats the plan

@ailzhang Anything that allows fp16 taichi fields to use Nvidia tensor cores for multiply-accumulate while still in taichi scope (and preferably other data types)

any updates calling torch matmul from the python scope slows downthings considerably not to mention it limits the granularity taichi has currently its one of the main factors for people...