본문으로 건너뛰기Skip to content 1:1 문의 — [email protected]로 이메일 작성 Contact us — email [email protected]

Daily Silicon: Intel 18A, AI Silicon: 256 Cores, 16 Memory Channels

A practicing semiconductor engineer's daily read on global silicon news - Intel 18A, AI Silicon: 256 Cores, 16 Memory Channels

Daily Silicon: Intel 18A, AI Silicon: 256 Cores, 16 Memory Channels
VLSI Korea 오리지널 에디토리얼 표지
🆓 무료 글 · 누구나 무료로 읽을 수 있는 글입니다

A practicing semiconductor engineer's daily read on global silicon news:

Intel’s 18A Diamond Rapids puts 256 cores and 16 memory channels on the table, while the 775µm HBM4 ceiling and RISC-V CI show that package and toolchain qualification, not raw compute, are now the next constraint.

Chase's Take - I look at Intel’s 16 memory channels and 128 PCIe Gen6/CXL 3.0 lanes before its 256 cores. As the package grows, bump maps, power grids, and thermal escape routes collide with timing closure in the same table.

Teams attaching next-generation HBM cannot treat 775µm as one package-substrate spec; warpage, underfill, and the burn-in window need one qualification package. The next checkpoint is when 16-Hi HBM customer qualification turns into real production conditions.

1. Intel Hot Chips 2026: Agentic AI Architectures and 18A

Intel Hot Chips 2026: Agentic AI Architectures and 18A

TL;DR - Intel puts up to 256 cores, 16 12,800MT/s memory channels, and 128 PCIe Gen6/CXL 3.0 lanes on 18A-based Diamond Rapids, shifting AI inference constraints from compute to memory and I/O topology.

Source: Intel Newsroom - Intel Hot Chips 2026: Agentic AI Architectures and 18A


2. SK hynix Hot Chips 2026: Hybrid Bonding, HBM5, and the 775-Micron Ceiling

SK hynix Hot Chips 2026: Hybrid Bonding, HBM5, and the 775-Micron Ceiling

TL;DR - SK hynix’s 16-Hi HBM customer qualification and 775µm stack-height ceiling show that warpage, thermal paths, and bonding windows arrive before bandwidth headlines.

Source: Tom's Hardware - SK hynix Hot Chips 2026: Hybrid Bonding, HBM5, and the 775-Micron Ceiling


3. Infineon Acquires C2i Semiconductors: AI-Server Multiphase Control

TL;DR - Infineon’s C2i acquisition turns the AI-server board voltage regulator from a component choice into a software-defined multiphase-control problem.

Source: Infineon - Infineon Acquires C2i Semiconductors: AI-Server Multiphase Control


4. STMicroelectronics NUS HELIX: 4-Year Edge-AI Memory Architecture

TL;DR - The 4-year STMicroelectronics-NUS HELIX lab treats edge AI as a coupled problem of memory-centric architecture, in-memory compute, embedded memory, and chiplet integration.

Source: STMicroelectronics - STMicroelectronics NUS HELIX: 4-Year Edge-AI Memory Architecture


What to Watch Next

  • NVIDIA FY2027 Q2 earnings on 8/26
    The next P0 date for rechecking HBM and AI-accelerator demand in the numbers.
  • SK hynix 16-Hi HBM customer qualification
    Production transition for the 48GB cube must validate stack height, thermal path, and package yield together.
  • Infineon’s C2i deal close and product integration
    This shows whether software-defined multiphase control enters real AI-server reference designs.
  • CPython RISC-V upstream CI integration
    The test is whether Tier 3 support becomes a repeatable release process with buildbot and ABI tests.

Frequently Asked Questions

What did Intel disclose for Diamond Rapids at Hot Chips 2026?

Intel disclosed up to 256 cores, 16 memory channels at 12,800 MT/s, and 128 PCIe Gen6/CXL 3.0 lanes on 18A-based Diamond Rapids. The focus shifts AI inference constraints from compute to memory and I/O topology.

What is the 775-micron ceiling for HBM?

The 775µm stack-height ceiling is the limit SK hynix cited for 16-Hi HBM. Warpage, thermal paths, and bonding windows must be validated together before the 48GB cube moves to production.

What did Infineon acquire from C2i Semiconductors?

Infineon acquired C2i Semiconductors for AI-server multiphase control. The deal turns the AI-server board voltage regulator into a software-defined multiphase-control problem.

What is the STMicroelectronics-NUS HELIX lab?

HELIX is a 4-year STMicroelectronics-NUS lab. It treats edge AI as a coupled problem of memory-centric architecture, in-memory compute, embedded memory, and chiplet integration.

Related reading

Discussion

READ NEXT / CURATED

You may also like

Reader favourites from this week, alongside our latest semiconductor research.

VLSI KOREA BRIEFING

반도체를 읽는 안목, 메일함으로.

엔지니어의 시각과 산업의 흐름을 함께. 무료로 구독하세요.

이미 구독 중인가요? 로그인