Chapter 2
Open Chiplet Economy
OCP Sponsored Tutorial
Chiplet Summit Feb 6th 2024
1:00 pm to 5:00 pm
Santa Clara California, USA
The Open Chiplet Economy:�Driven by the OCP
Connect. Collaborate. Accelerate.
What Our Volunteers Delivered in 2023
Q1 | The Open Chiplet Economy (at the Chiplet Summit) Chiplets business workshop (to be repeated) Bunch of Wires 2.0 |
Q2 | Link Layer 1.0 ODSA Breakout Session at the OCP Regional Summit |
Q3 | OCP-LBL Chiplets Workshop JEDEC JEP30 Part Model standard (seeded/driven by OCP CDXML) |
Q4 | PHY workshop (to be repeated) First ever open chiplet economy experience center – real chiplets, workflow, IP MRHIEP Manufacturing Roadmap Final Report Packaging analysis paper at VLSI Design Conference ODSA Breakout Session at OCP Global Summit |
Connect. Collaborate. Accelerate.
OCP/ODSA Contribution to Standards/Tools�Engineer-Driven Multi-company Efforts
Area | Output | |
D2D PHY | Bunch of Wires 1.0 Bunch of Wires 2.0 2.1 in flight | 1st standard to scale from laminate to advanced packaging, 2-32 Gbps/lane Known implementations in 65, 22, 16, 12, 7, 6, 5, 4, 3nm – true heterogeneity 10+ products in flight Proven power of 0.3 to 0.5 pJ/bit, support for proprietary extensions |
D2D Spreadsheet | V3.0 in flight | Biennial release compares data on all PHYs |
D2D Link and Transaction | TLP 1.0 | Only known “streaming mode” open link layer |
D2D Transaction Profile | NXP DiPort/Other Profiles | Only known open maps of AXI SOC Traffic to D2D PHYs |
CDX | JEDEC JEP30 PM Open 3DK | Standard for physical chiplet description Workflow white papers |
Business | Chiplet cost model Business white paper | Open spreadsheet model to compare chiplet/mono Product planning assistance document |
Prototyping | Test Package | 1st open package to integrate chiplets across vendors Fully open package design and analysis Aim to build a common bring up infra First demo of chiplets/IP/workflow from 11 companies in an experience center |
All specifications available without an NDA, can be used without any restrictions
10+ refereed conference and journal papers
Community-driven hyperscale innovation for all.
Building a Chiplet
Connect. Collaborate. Accelerate.
One Example – designing D2D bandwidth�(Classification From Elad Alon BCA)
Category | Package | Per-Line Rate | Termination | Reach |
Ultra Wide (UW) | Advanced only�(<55 μm pitch) | < 8 Gb/s | No | <4mm |
Ultra Dense (UD) | Advanced only�(40-55 μm pitch) | 8 – 32 Gb/s | No | <2mm |
Full Reach (FR) | Standard or fanout�(55-130 μm pitch) | < 40 Gb/s | Supported | <25mm |
Cost-Optimized �Full Reach �(CO-FR) | Low-layer standard�(130-180 μm pitch) | < 40 Gb/s | Supported | <25mm |
Die Edge
Connect. Collaborate. Accelerate.
Bandwidth vs. Bump Density, Even within a Standard
From Elad Alon at Blue Cheetah�OCP Global Summit
Options at a performance target
The exact way to achieve a�combination of power,
performance, packaging �technology is and will need
to be domain-specific
Already true at the �Hyperscalers
Every chiplet your chiplet
talks to has to make the same�trade-offs your chiplet made
Connect. Collaborate. Accelerate.
Reduce Friction in Chiplet Integration�Bring Factors Forward in Design
Customer�Specification
Design &�Verification
Semiconductor�Foundry
Assembly�Packaging & Test
System�Software
OEM and�Product
codesign
physical modularity
Function modularity
additional constraints �with chiplets
Factors brought forward or made more complex
Large companies solve this by building chiplets in families and with engineering convention
https://www.opencompute.org/blog/building-an-open-chiplet-economy
Connect. Collaborate. Accelerate.
Large Companies Design Chiplets in Families
Area | Candidates |
D2D PHY | UCIe, BoW, SuperChips, XSR |
D2D Transactions | CXL, ODSA TLP, DiPort |
Chiplet and package test | IEEE 1838 |
Chiplet description | OCP-JEDEC CDXML |
Power models | ISIL |
Device traceability | SEMI |
Size guardrails | 🗶 |
Power delivery guardrails | 🗶 |
Thermal guardrails | 🗶 |
Wiring density guardrails | 🗶 |
Mechanical guardrails | 🗶 |
Bump and assembly pitch | 🗶 |
Connect. Collaborate. Accelerate.
Fugaku Thought Experiment
10
2/6/2024
Attribute | Detail |
Die size/ Power/Thermals | Approximately 400 mm2/�122 W/0.3 W per mm2 |
Aspect ratio | 20x20 |
Cores/ASIC | 52 |
Tofu, /HBM/PCIe/ area power | Tofu 25 mm2/9W, 100 mm2 for 4 HBM + 1 PCIe |
Ring bus extra area | Approximately 35 mm2, 10 mm2/complex |
Core complex | Approximately 240 mm2, 60 mm2/complex |
Core Area/Power | 4.6 mm2/2.2W |
Fab/Process | TSMC/7 nm Finfet |
ASICs/Node | 2 |
Nodes/System | 158,976 |
System power | 40 MW |
Connect. Collaborate. Accelerate.
1st We Chipletize into Standardized Modules
Community-driven hyperscale innovation for all.
Then We Create a Family of Variants
Community-driven hyperscale innovation for all.
Please join us in 2024
Connect. Collaborate. Accelerate.