NOW OPERATING · SAN JOSE · MUNICH · SINGAPORE · DUBAI · BANGALORE

Quantum Chip
Home / case-studies / Custom-7nm-AI-Accelerator-Card-with-HBM3-&-CXL-2.0
Case Studies AI Infrastructure

Custom 7nm AI accelerator card with HBM3 and CXL 2.0 — first silicon in under 14 months.

CUSTOMER

Hyperscaler-Track AI Startup
· Series C · Bay Area

PROGRAM LENGTH

14 months (Architecture
→ First Silicon) ·
22 months to GA card

SCOPE

SoC architecture · RTL
· Verification · PD · Tape-out ·
PCIe card · Firmware · Drivers

SION TEAM

42 engineers across Bangalore,
San Jose & Singapore

THE CHALLENGE

Beat the hyperscaler evaluation window.

A well-funded AI startup had a differentiated model-serving IP in FPGA form — but their hyperscaler customer had opened a 24-month evaluation window for custom accelerator cards. To be in the race, the startup needed a production ASIC, a PCIe Gen5 accelerator card, firmware, and a driver stack — all within that window. A missed window meant a missed multi-hundred-million-dollar opportunity.

• 7nm ASIC with high-speed SerDes (16 lanes) and CXL 2.0 support
• PCIe Gen5 x16 host interface with HBM3 memory subsystem
• Thermal envelope: 350W air-cooled, OCP-compliant form factor
• Driver stack integrated with customer's existing ML-serving runtime
• First silicon in hand within 14 months of program kickoff
WHY SION WAS CHOSEN

The only partner offering chip + card + firmware under one SOW.

The customer was a 90-engineer startup — exceptional at ML architecture, but with no in-house silicon, board, or firmware experience. They needed a partner who could own the entire hardware delivery path while they focused on ML IP. Sion's 24+ prior tape-outs at 7nm and proven high-speed SerDes and HBM3 integration expertise made the shortlist a formality.

7nm

Advanced-node execution

24+

Prior production tape-outs

What Sion Delivered —
End-to-End

From customer RTL to deployable card.

01

SoC Architecture & Integration

Full SoC architecture around the customer's ML compute IP, NoC fabric design, CXL 2.0 subsystem, HBM3 memory controller integration, and host-PCIe Gen5 subsystem.

02

RTL Integration & Verification

SoC-level integration, full UVM verification environment with 95%+ functional coverage, formal property verification on PCIe/CXL controllers, and gate-level signoff.

03

Physical Design & 7nm Tape-out

Full-chip 7nm physical design, high-speed SerDes integration, HBM3 PHY signoff, power-grid design for 350W, DFT insertion, and successful first-silicon tape-out.

04

Accelerator Card Design

OCP-compliant PCIe Gen5 x16 card with HBM3 integration, thermal solution for 350W air-cooling, power delivery for 8×48V rails, and manufacturability review.

05

Firmware, Drivers & BSP

Card-side firmware, Linux host drivers, DMA engines, customer-runtime integration plugins, and platform BSP for the target hyperscaler reference server.

06

Bring-up & Customer-Site Deployment

Lab bring-up, performance tuning, customer-site validation, co-deployment with the customer into the target hyperscaler reference environment, and issue resolution to GA.

First silicon on schedule. Into the customer's hyperscaler evaluation with weeks to spare.

14 mo

To first silicon

Architecture kickoff to working chip in hand

100%

First-silicon success

No respin required — PCIe, CXL, HBM3 all functional

3 WKS

Lab bring-up to customer sample

From first power-on to customer-site deployment

Series D

Customer Closed

On the back of a working, production-grade product

Why This
Mattered

— THE SION difference

Advanced-node execution discipline, not promises.

For an AI startup racing a hyperscaler evaluation window, "first silicon" is everything — it is the difference between a multi-hundred-million-dollar contract and a down round. Sion's ability to own every layer from SoC architecture through firmware drivers meant the customer's 90-person team could stay focused on ML differentiation while Sion carried the hardware execution. First silicon worked on the first try — a rare feat at 7nm with HBM3 and advanced CXL integration.

Confidentiality & Verification

Real programs.
Verifiable references.

Customer identities are held confidential under mutual NDA. Technical scopes, outcomes, and program structures shown on this page are representative of actual Sion engagements across storage, AI infrastructure, and networking. Full technical deep-dives, architecture reviews, and direct reference calls are available to qualified prospects under mutual NDA.

01

24+

PRODUCTION TAPE-OUTS

02

3nm

ADVANCED-NODE EXPERIENCE

03

Top 5

SEMI CUSTOMERS SINCE 2021

04

ISO

GRADE IP-PROTECTION PROTOCOLS

Hardware Left
Hardware Right

Your program. One accountable team.
Start the scoping call.

Under mutual NDA, we walk you through full technical deep-dives of the programs shown above, connect you with customer references, and map your hardware challenge to a Sion engagement model.