$30M Raised to Redo the CPU
A startup has raised $30M to build a massively parallel, non–von Neumann processor with thousands of cores, paired with a new language called Vine designed around interaction nets and parallel-first execution. Commenters draw parallels to past ambitious architectures like Transputers, Cell, Mill, Tilera and various dataflow machines, noting that hardware is easier than building the compilers, tools and ecosystems needed to make such chips useful. Some see AI-assisted code generation and verification as a potential game-changer for overcoming historic software and portability barriers, while others doubt the commercial viability of yet another exotic architecture in a world already dominated by x86/ARM CPUs and GPU/ML accelerators.
Overall reaction
- Many are intrigued by a “redo the CPU” attempt but see it as revisiting old ideas rather than something clearly new.
- Sentiment ranges from “amazing / unfinished business” to “$30M to launder and lose” and “bad use” of public funds.
Comparison to prior architectures
- Frequently compared to GPUs, dataflow processors, systolic arrays, and past manycore projects: Transputer, Cell/PS3, GreenArrays, Tilera, Xeon Phi, Parallella, XMOS, Mill, Transmeta, Thinking Machines, TILE-Gx, etc.
- History cited as cautionary: many massively parallel or exotic CPUs under-delivered because software and ecosystems never caught up.
- Some note that RISC-V succeeded quickly by being intentionally boring and familiar, while “big swings” like Mill stalled.
Programming model & Vine language
- Tendrils relies on Vine and interaction nets; the chip itself is meant to do graph rewrites instead of running a traditional sequential ISA.
- Vine removes explicit threads/locks; parallelism emerges from graph structure, but developers still must choose parallel algorithms (e.g., divide-and-conquer factorial vs naive).
- Confusion that marketing language frames basic superscalar-style parallelism as novel; some find the docs and landing page too vague.
Compilers, LLMs, and software ecosystem
- Longstanding lesson: new CPUs live or die by compilers and tooling, not hardware alone.
- Some argue this concern is reduced now: LLMs, tight test loops, and formal methods could automate porting/optimization and make exotic hardware more viable.
- Others counter that CUDA-style ecosystem and network effects are still a huge barrier; AMD’s slow catch-up is used as evidence.
Market fit and customer value
- Big question: who wants an exotic, parallel-first CPU on an older node that requires a new language when GPUs/TPUs, RISC‑V, ARM NPUs, and existing dataflow IP already exist?
- For general-purpose computing, people point out: single-thread performance, memory bandwidth/latency, and compatibility often matter more than raw parallel FLOPs.
Hardware design & feasibility questions
- Doubts about abandoning out-of-order execution and cache coherency without heavy multithreading; otherwise cores may sit idle on memory stalls.
- Questions on how data movement and scheduling are handled at hardware speed, especially as graphs grow large.
- $30M is seen by some as barely enough for multiple tapeouts; others think old process nodes plus a focused scope might be sufficient.