Google Axion Processors – Arm-based CPUs designed for the data center

Original Article ↗ Hacker News Discussion ↗

Lack of Technical Detail and Benchmarks

  • Many note the announcement is heavy on marketing (“X% better”) and light on hard data.
  • No public specs: no release dates, pricing, core counts, cache sizes, or exact x86 instances used in comparisons.
  • Several expect third‑party benchmarking eventually, but some wonder if terms of service might restrict it.
  • Testimonials from partner companies read as “we plan to evaluate,” suggesting limited real‑world testing so far.

Comparisons to Other ARM and x86 Offerings

  • Axion is described as Neoverse‑based rather than a custom core, similar in spirit to AWS Graviton and Azure’s ARM offerings.
  • One commenter expects performance roughly in line with Neoverse V2 systems like Graviton4, but details such as frequency and cache are unknown.
  • Some see this as Google “copying Apple’s playbook” on custom silicon, though others point out Apple at least ships detailed, demonstrably strong chips.

Implications for Intel, AMD, and Ampere

  • Hyperscalers building their own ARM chips are seen as pressure on Intel/AMD’s server CPU business.
  • Some argue this helps explain Intel’s push into foundry services; their big challenge is business/model and scaling, not raw process capability.
  • There’s speculation this could sideline Ampere on GCP, which has been offering Ampere-based ARM instances since 2022.

Lock‑in, Portability, and Availability

  • Axion appears cloud‑only, with no indication it will be sold as standalone hardware; some lament lack of purchasable modern ARM server CPUs beyond Ampere.
  • Others downplay ISA‑level lock‑in, noting these are standard ARM chips and ARM server standards like SystemReady/(U)EFI/ACPI exist.
  • Practical lock‑in still comes from cloud contracts (e.g., AWS savings plans) and proprietary services, making switching hard even if CPUs are “fungible.”

Cloud Experience, Pricing, and Reliability

  • Opinions on GCP vary: some find it simpler and better‑priced than AWS; others complain about degraded service, shifting pricing (e.g., BigQuery), and “best effort” support attitudes.
  • Multi‑cloud strategies are mentioned, but egress costs and committed discounts complicate this.

Efficiency, Emissions, and Accelerators

  • Some welcome any move toward lower‑power chips; others argue modern server CPUs are already relatively efficient per core, and total power comes from high core counts and I/O.
  • Several say hyperscalers chose ARM mainly because it’s licensable for custom SoCs, not because ARM is inherently more efficient than x86 in servers.
  • Google’s PUE and environmental reporting are cited as relatively transparent, but actual emissions numbers across providers remain hard to compare.
  • Axion’s offloads for networking and storage are viewed as Google catching up to AWS Nitro; some developers say they can’t rely on such vendor‑specific accelerators because they must run across multiple clouds.

Broader Industry Trends and Tone

  • Some see the announcement as “headline‑driven” or theater, timed to Google Cloud Next and possibly financial reporting.
  • Others frame it as another step in the broader trend: big clouds using ARM Neoverse + standard IP for differentiated but broadly compatible compute.
  • There is mild amusement at the physics/element naming trend (Axion, Graviton, Ampere, Cobalt) and speculation it’s mainly about sounding cool and avoiding trademarks.