10/4/2025 - research

started work on researching how to make my gpu with my main source being bitlui's video + github. it's a pretty awesome video so i would recommend checking it out.

i also spent some time looking at the schematic and pcb design behind it.

will likely be using risc-v's since they are cost effective yet powerful.

2/4/2026 - more research

it's been a bit of a while since the last devlog but i've tried to do a ton of research over that time to try to make the project more feasible software wise as well. right now flashing firmware still looks like it's going to be a big pain which is something that i think i will need to figure out when i get the board. the main change i want to highlight is the shift from a gpu towards a more clustered style board.

part 1: architecture

so i think i've figured out a general architecture on how i want the board to work. the plan is to have 16 modular slots where i can add different chips/designs i have. in addition, each one of these modules will not have 16 chips like i was initially thinking but to instead have 17. the reasoning here being that we have 16 compute chips and then 1 low power chip which contains information about the board and will send that to the main chip allowing for the computer to know what all the different modules are. this should also allow for me to make more modules and might even make it easier firmware which is something i am concerned about.

4 layer pcb. i'm concerned about going for 1 layer for power and one for gnd since that will take away a large portion of the area for traces. i do also need to stay within the bounds of the standard manufacturing for pcbs for stuff like traces and vias otherwise it will get very expensive.

part 2: chips

chip wise there's a few different considerations. i haven't fully decided on which chip i will be mainly using to communicate with all of the modules and the computer but that's what research is for ig. module wise my main comparison was initially between the esp32, stm32, and riscv. i have the most experience with the stm32 which does come out to being the best in terms of computing power/performance especially with the n6 for npu related tasks and h7 for general stuff. i am leaning towards the h7 however, i could easily make a configuration where i have two modules of n6's and six modules of h7's allowing for me to get the best of both worlds. the v8 is an upcoming upgrade to h7 but isn't out yet.
the esp32 does have bluetooth which i don't need for the modules but might be helpful for a secondary chip in tandem with the primary chip on the mainboard to make use of the peripherals. by contrast, riscv has an open architecture and is more suitable to clustering. it does have a lower price point that still gets you pretty good performance but you can also buy more expensive chips that might be better.

see i was pretty happy with this but then Kai mentioned that i might want to take a look at rockchip. rockchip manufacture mpus which would add extra complexity as i would also need to figure out how to work with the different chipsets above while still supporting them. now he specifically mentioned their RK3588 as it's pretty powerful and also accessible. the only problem is that they are expensive by itself but they are also currently out of stock. i instead looked a bit more and came upon the RK3576 which is half the price and still pretty powerful. at the same time, i realized the RK3588 might be a good option for the main chip to communicate with the computer if i got an interface working.

2/6/2026 - research part 3

https://docs.google.com/spreadsheets/d/1Cx6SLtbQfrMjaOoha3WtSbsGnKbudi3LlXvoRu4ph1w/edit?gid=0#gid=0

compiled a document with info on all four of the main chips i have in mind right now. thinking of taking bitluni’s design, increasing each card by 4x and then adding more cards

take some chips ig