Top researchers leave Intel to build startup with 'the biggest, baddest CPU'

Original Article ↗ Hacker News Discussion ↗

CPU vs GPU and ML Hardware

  • Multiple comments argue it’s far easier for a startup to ship a CPU than a GPU: CPU interfaces (compilers, OS, tools) are standardized, while GPUs need massive, evolving software stacks (graphics APIs, custom compilers, CUDA-like ecosystems).
  • Several people want affordable ML-capable hardware more than a “GPU” per se, but others note ML accelerators are even harder: you must match NVIDIA’s rapid cadence and CUDA lock-in, which most software assumes.
  • Discussion of GPU memory:
    • Request for GPUs with user-upgradable large RAM; countered that GDDR close to the die is essential for bandwidth, and any move to socketed or system RAM is a huge performance hit.
    • Techniques like GPU access to system RAM/storage exist, but are seen as last-resort tools that “all suck to different degrees.”
  • Debate over whether discrete GPUs/AI coprocessors will disappear like FPUs. Consensus: integrated NPUs/GPUs will dominate low-power devices, but high-end and datacenter workloads will continue to need large discrete accelerators.

RISC‑V, Openness, and Ecosystem

  • Some are excited by a “biggest, baddest” RISC‑V CPU and see room for a high-performance implementation, analogous to Apple’s use of ARM.
  • Others note RISC‑V’s main advantage is open licensing; it doesn’t prevent ME/AMT-style management engines, which are ISA-agnostic.
  • Ecosystem concerns:
    • Toolchains exist and are improving, but high-end microarchitecture-specific tuning is immature because there are few truly high-performance RISC‑V cores to target.
    • LLVM/GCC can and do optimize for particular cores via scheduling models, but this requires complex per-CPU descriptions and detailed vendor docs.
  • Some see starting at supercomputing/high-end servers and working downward as an unusual but potentially disruptive path for an ISA.

Startup, Article, and Intel Context

  • Commenters find the article vague on technical details, reading more like a local-business or investor pitch emphasizing founder pedigree rather than architecture specifics.
  • The piece is framed as regional news: Intel is a major Oregon employer, so a spinoff is notable to a non-technical audience that may barely recall what a CPU is.
  • Some see this as a bad look for Intel—loss of senior talent and continued disinvestment in Oregon—rather than clear evidence the startup is special.
  • There’s skepticism of “ex‑BigCo” branding in general; prior high-profile failures are cited as evidence that résumés and combined “X years of experience” are weak predictors of startup success.
  • A few expect brutal competition in AI/compute and predict that, if the company succeeds, it’s likely to be acquired by a larger player.