Skip to content

Brief · 8 September 2026

What changed

Phoronix reported a mysterious new x86 processor that adds Intel‑style APX extensions and supports up to 32 AMX tiles, a claim that would outpace Intel's current matrix‑extension lineup. https://www.phoronix.com/news/x86-CPU-With-APX-x86S

One number

32tiles

Maximum AMX tiles claimed for the new mysterious x86 CPU, exceeding Intel's current offerings.

source ↗

Still vapor

The vendor’s press‑release‑style teaser promises “legacy‑free x86” performance, yet provides no silicon roadmap, silicon‑validation data, or pricing—typical hype that rarely translates into production AI servers.

Today’s AI‑hardware signal is quiet. No new GPU, accelerator, or server rig entered our catalog; the roster of 51 verified machines remains static. The only concrete shift is a speculative CPU announcement that could reshape AI‑workload economics if it materializes.

Phoronix’s deep‑dive uncovers a start‑up‑origin x86 chip that adds the APX instruction set and, more intriguingly, claims up to 32 AMX tiles per core. Intel’s current AMX implementation tops out at 16 tiles, so the advertised doubling would theoretically double matrix‑multiply throughput for transformer inference. The article notes the design “surpasses Intel’s current Advanced Matrix Extensions,” but offers no silicon samples, production timeline, or power envelope.

For operators, the relevance hinges on three questions: Will the chip ship in a data‑center‑ready form factor? Can it be paired with existing DDR5‑ECC or HBM memory stacks without a custom motherboard? And at what price point will it compete with NVIDIA’s Blackwell GPUs or AMD’s MI300X accelerators?

Until silicon is validated, the claim remains a headline‑grabber. Teams planning new AI clusters should treat the 32‑tile figure as a speculative upper bound, not a procurement target. Keep an eye on any follow‑up disclosures that include silicon die shots, TDP numbers, or OEM partnerships.

If the processor does appear, its real impact will be measured against real‑world benchmarks like TensorRT‑LLM throughput, not the theoretical tile count.

Composed by the MadCoolStuff editor pipeline · Groq · openai/gpt-oss-120b · 2026-09-08

Tags

What we read