I write posts and try to do them daily or at least weekly. Check out my site (tristanxr.com/) for my latest posts. I often write about topics like ARM, RISC-V, NixOS, and things I am interested in or involved in.

Boing ball on Glacier

Got an Artix 7 board from Alinx, PCIe is like 80% working in a test bitstream with openXC7. Hoping to get Glacier to be usable under Linux with this.

GLM-5.3 is good at removing Claudisms

Claude added a dedicates port from an SDRAM controller to the scanout module. It then wonders why PnR is becoming difficult to solve. I tell it its because of that long bus between the SDRAM to the scanout. It pushes back but checks, it realizes I am right. I tell it to go from wishbone to AXI4 and use the different channels to help with access. It does that and timing increases beyond the pre-second SDRAM read port. LLM's need to be trained more on hardware design, they don't understand timing. I do think it's somewhat because they don't imagine things. I definitely visualize the congestion and can rationally think about it. But at least Claude wants to be evidence driven and doesn't believe things unless it checks. But it wants to check absolutely everything.

Glacier is properly running a starfield animation.

It's funny when Claude uses flip flops for buffering instead of BRAM and wonders why routing struggles.

Claude: This change is adds an extra 270 cycles of latency. *Runs PnR, Fmax up by 40%* Claude: I was wrong, this is faster.

I might get a Kintex 7 FPGA, it looks like openXC7 has done a decent job getting those supported.

Why is PayPal closing my account that I use actively and have logged into over the past 3 years? I literally last used it just a few hours ago. I've had this account for years.

Glacier is coming online. I can only issue 1 command to the GPU until it dies. However, progress is being made. Once things are fixed, going to try rendering a test pattern.

Putting a RISC-V GPU onto the ULX3S at NixCamp.

Chock now has a comparison with Claude Code and Codex. chock.ws/comparison/

Managing to run GLM-5.3 Flash Heretic at 4 t/s on an Ampere Altra box without using the GPU. Hoping to push it to 8 - 12 t/s soon. This is using @LilithSemi's Vulcan codegen library. Once the inference engine is optimized enough on CPU alone, will be working on getting fast GPU inference on an RTX 5070 (12GB). AI is at the point you can have it design its own harness and inference engine and work with it to optimize. You can feed it papers that show different quantizations, caching methodologies, and have it self optimize.

ai& is the best open weight model provider out there that you can trust. I've been using GLM-5.3 from them and it's been excellent. Chock has first class support for ai& and was tested with models ai& hosts. I'm looking forward to seeing them grow and expand their capabilities.