During the CXL Forum at OCP Global Summit, Eddy Hwang of Nvidia and Wai Kong Poon of Molex presented a next-gen architecture for enabling copper for AI computing.

1. Partnership to enable copper for the next
generation artificial intelligence computing

2. Partnership to enable copper for the next
generation artificial intelligence computing
[Seunghyun Eddy Hwang, Principal SI Lead, NVIDIA]
[Wai Kiong Poon, Global Product Manager, MOLEX]

3. Enabling Copper for Artificial Intelligence Computing
Next Generation Architecture

4. Voice of Customer
 High Speed Board-To-Board (112G PAM4 and beyond)
 Low Profile (5mm mated height as initial target)
 Surface Mount Termination (Solder Ball Attachment)
 Minimum PCB Real-Estate
 Mechanical Robust
 Simplicity in Design
 Ease of manufacturing; Cost Effective
High Speed Mezzanine Connector

5. MM
MM Pro
MMe
• Speed up to 112Gbps
• Speed up to 112Gbps (improved performance)
• Speed up to 224Gbps
Evolution of High-Speed Mezzanine
High Speed Mezzanine Connector

6. High Speed Mezzanine Connector
What have we learnt from the current Mezzanine Design?
 What have we learnt from the current Mezzanine Design?
— What are the Current Issues?
— What is the Signal Integrity Limitations?
 Moving from 112G to 224G:
— Understanding issues with current design
— Understanding limitations with current assembly process
 These understandings will enable us to design the next generation of High-
Speed Board to Board Connector

7. SI Performance influenced by Terminal & Assembly Process
 A good impedance control is critical for 224G application, but the impedance
optimization is limited by the following:
— Connector Design
— Assembly Process
 Due to the current stitched terminal design,
the flexibility and sensitivity of the terminal
width tuning becomes very challenging
Dimension Constraints due to Assembly Process.
High Speed Mezzanine Connector

8. Limitations of Current Design
 Minor variation in the dimensions will result in resonance
 This type of variation is very difficult to control during assembly process
High Speed Mezzanine Connector

9. TDR Connector only
Sensitivity of SI Performance due to Terminal Deflection
 Aside from dimension variation, the impedance is also sensitive to
deflection condition which affects the distance between terminal and its
nearby plastic
High Speed Mezzanine Connector

10. Limitations of Current Design
High Speed Mezzanine Connector

11. MM
MM Pro
MMe
• Speed up to 112Gbps
• Speed up to 112Gbps (improved performance)
• Speed up to 224Gbps
Evolution of High-Speed Mezzanine
High Speed Mezzanine Connector

12. Victim : Diff 4
Aggressor : others
#Rference impedance 86ohm
Comparing MM, MM Pro Vs. MMe (Signal Integrity)
High Speed Mezzanine Connector

13. Comparing MM, MM Pro Vs. MMe (Signal Integrity)
High Speed Mezzanine Connector

14. • It’s no secret that GPU accelerators now power many of the world’s fastest supercomputers and AI
systems
• NVLink is the world’s first proprietary system interconnect technology from Nvidia that allows multiple
GPUs to communicate directly via a high-speed interconnection
• NVLink connects the machines’ processors – CPUs and GPUs – so they can exchange data much faster
than CPU
What is NVLink?

15. NVIDIA DGX
• Nvidia Tesla V100 is the
world’s most advanced
data center GPU
• Supports AI, deep
learning, HPC, and
autonomous driving
• Tesla V100 offers the
• DGX H100, which uses 4th
generation of NVLINK, is the
world’s most advanced GPU
for large generative AI and
other transformer-based
workloads
• H100 contains 8 GPU modules
that communicate thru
NVLink needs high-speed
connector for baseboard
attachment

16. Improved H100 Performance Bring SI
Challenge
• Significant performance boost from A100
• Fourth-generation NVIDIA
NVLink provides a 3x bandwidth
increase on all-reduce operations and
multi-GPU IO operating at 7x the
bandwidth of PCIe Gen 5.
• Yet, no significant form factor change for
improved performance, which introduces
significant SI challenges
A100
H100

17. Mezzanine Connector Selection
Criteria
• SI performance
• As shown in crosstalk
comparison plot in left
figure, other vendor
crosstalk is much worse
than Molex
• Formfactor
• Molex pinout can route all
NVLink high-speed signals
in less number of layers
than other vendor
• Mechanical stability
Molex
Other vendor

18. Molex Connector Optimization
Optimized
Baseline
Optimized
Baseline
Optimized
Baseline
Resonance
free

19. SI Performance Comparison: H100 vs
A100
A100 H100

20. Thank you!