Wired communication at VRAM speeds on a single low (production) cost chip?! This is going to be game changing for AI and super-computing. Wow.
At these speeds clustering and distributed computing will be a thing of the past in supercomputing.
The supercomputer of next year may functionally be an enormous symmetric multiprocessor bottle-necked only by physical space and power.
The same code could be run on a laptop or supercomputer without having to optimize it. This will make development of large scale applications way more accessible.
>A number of products based on the Odin 32 are expected to be announced by late next year.
Whenever something is too good to be true I am always skeptical. But since they are actually talking about shipment rather than R&D may be it really is that revolutionary.
I remember Intel has a concept where CPU, Memory, and Storage are all in their own separate rack and linked up via similar photonics connection and can scale up and down as needed. No more server in the traditional sense.
Photons are still a lot faster than electrons. Latency at the endpoints is still a thing, but this allows a lot of 800 Gbps channels along distances copper can't match.
I don't want to think about the NUMA model for a beast that takes this to its upper limits.
What is the class of problem that can be solved by computers of a certain size (compute density metric)? If a computer is basically instantaneous, but has to communicate over vast distances, it still couldn't solve large problems.
Couldn't we have already made 100k core single image OSs running over RDMA?
I'm sure that, as we extend the capabilities of our machines, we'll find new problems or approaches to known problems that were not practical up until then.
Like I mentioned, this can increase bandwidth and reduce latency between system components, making warehouse-sized computers look like container-sized ones. I bet we can find some use for that.
At these speeds clustering and distributed computing will be a thing of the past in supercomputing.
The supercomputer of next year may functionally be an enormous symmetric multiprocessor bottle-necked only by physical space and power.
The same code could be run on a laptop or supercomputer without having to optimize it. This will make development of large scale applications way more accessible.