In case you have not noticed, parallel buses have been going away, to be replaced by SERDES – USB, SATA, PCIe – and only DDR remains, and the writing is on the wall for DDR as we move to doing in-memory computing for lower power and latency.
Something that has always been serial is Ethernet, which has been evolving steadily at the hardware level over the years from coax, to twisted-pair, up to optical fiber, and has recently made a step advance from IPv4 to IPv6 at the software level.
This is interesting in that software is used to working with serial channels like Ethernet, and the IPv6 protocol can work equally well on other hardware like those at the top. Like PCIe you can also gang channels –
So, if you look at all your SERDES linked hardware as a bunch of small machines that can talk on IPv6, we already have a (network) computing paradigm that matches.
We also have most of our processors working with a 64-bit physical/virtual address space – a smaller space than the IPv6 address space. Therefore you can do a 1:1 mapping of virtual memory addresses to IPv6 addresses, i.e. if you want to map an NVMe storage card into an OS/application, you can do it by NVMe-oF regardless of where it is.
At this point you should remember that in machine virtualization the network stack can be short-circuited from the C code read/write level, so once your IPv6 channel is connected it can be re-assigned into different physical channels as needed for optimal performance.
How do you program a network of little computers? – that’s what ParC is for. How do you run legacy code on it? – that’s what wandering-threads is for.
Dear Kevin,
I listenend with interest to the interview you did with RunTime Recruitment, talking about EDA and AI, and I was just thinking again about these ideas this morning. I made this rambling video that might interest you. https://youtu.be/T-o_9rWGg8A Are you still based in CA? Do you come over to the London ever? I am here, for some terrible sins I must have committed at some time, …
I have a blog and this my most recent post, on MITs Matrix Calculus for Machine Learning course: https://eternaldoorman.blogspot.com/2023/10/steven-johnson-so-you-think-you-know.html
Best wishes
Ian Grant