Reimagining the multi-tenant network - presented by Marvell.pdf

OCP Network

1. Reimagining
the Multi-tenant
Cloud Network
Nick Kucharewski
SVP & GM, Network Switching BU, Marvell
October 18, 2023

2. © 2023 Marvell. All rights reserved. 2
Forward-looking statements
Except for statements of historical fact, this presentation contains forward-looking statements (within the
meaning of the federal securities laws) including statements related to future revenue, future earnings,
and the success of our product releases that involve risks and uncertainties. Words such as “anticipates,”
“expects,” “intends,” “plans,” “projects,” “believes,” “seeks,” “estimates,” “can,” “may,” “will,” “would” and
similar expressions identify such forward-looking statements. These statements are not guarantees of
results and should not be considered as an indication of future activity or future performance. Actual
events or results may differ materially from those described in this presentation due to a number of risks
and uncertainties.
For other factors that could cause Marvell’s results to vary from expectations, please see the risk factors
identified in Marvell’s Quarterly Report on Form 10-Q for the fiscal quarter ended April 29, 2023as filed
with the SEC on May 26, 2023and other factors detailed from time to time in Marvell’s filings with
the SEC. Marvell undertakes no obligation to revise or update publicly any forward-looking statements.

3. © 2023 Marvell. All rights reserved. 3
2024: 1 year
After Generative AI
2023 = 0
A New Beginning
2022: 1 year
Before Generative AI

4. © 2023 Marvell. All rights reserved. 4
Two lasting impacts of 2023
Rapid expansion of AI
accelerator investments
Revolution in cloud
connectivityinfrastructure

5. © 2023 Marvell. All rights reserved. 5
Agenda
Multi-tenant cloud applications
The impact of Generative AI
Implications for network infrastructure
The future of networking equipment
1
2
3
4
© 2023 Marvell. All rights reserved. 5

6. © 2023 Marvell. All rights reserved. 6
Last 20 years ... evolution to microservices
UI
BUSINESS
LOGIC
DATA
ACCESS
LAYER
DATABASE
MonolithicMicroservices
MICROSERVICEMICROSERVICEMICROSERVICE
DATABASE
UI

7. © 2023 Marvell. All rights reserved. 7
Last 20 years … evolution to multi-tenant
Single tenant
APP APPAPP
Multi-tenant
IaaSPaaS

8. © 2023 Marvell. All rights reserved. 8
Last 20 years …hardware specialization
HW-specialized
compute nodes
General-purpose
compute nodes

9. © 2023 Marvell. All rights reserved. 9
Last 10 years … AI on the side
TRADITIONAL
CLOUD
DEDICATED
AI
Request
AI service
Response

10. © 2023 Marvell. All rights reserved. 10
Generative AI: pervasive integration into services
Salesforce Einstein Studio to help
enterprises train generative AI
models
Microsoft unveils Copilot pricing,
adds new generative AI
capabilities
Splunk’s new AI tools aim to
ease security, observability tasks
ServiceNow adds new features to
its Now Assist generative AI
assistant
Alteryx works in generative AI for
speedy analytics results
Informatica integrates
ClaireGPT, AI copilot to
augment its IDMC offering

11. © 2023 Marvell. All rights reserved. 11
GenAI drives new connectivity infrastructure
This changes everything
?

12. © 2023 Marvell. All rights reserved. 12
AI datasets exceed single-node capacity
AI compute
Node 1
AI compute
Node 2
AI compute
Node N
Dataset

13. © 2023 Marvell. All rights reserved. 13
AI accelerators require 6-8x more throughput
xPU
3000+ Gbps
CPU
400+ Gbps

14. © 2023 Marvell. All rights reserved. 14
Redesign of cloud fabric infrastructure
ConvergedFrontend / Backend

15. © 2023 Marvell. All rights reserved. 15
Pervasive-AI multi-tenant cloud networks
Low latency
and lossless
Instrumented
and open
PredictableTransparent
High
throughput
Fast

16. © 2023 Marvell. All rights reserved. 16
High throughput
Single 12.8T
ToR switch Multiple 51.2T
switches
Accelerated rackApplication rack
…

17. © 2023 Marvell. All rights reserved. 17
Low latency
Single-hop
latency
t (ns)
Queueing
delay
Traversal
time

18. © 2023 Marvell. All rights reserved. 18
Optimizing single-node latency
Low latency
under any conditions
is key to predictable
fabric performance
Switch silicon architecture matters
200
300
400
500
600
6412809216
Teralynx® 10
minmaxavg
t (ns)

19. © 2023 Marvell. All rights reserved. 19
Optimizing traversal time
… …
Flatten the network with high radix
256-radix switch: 5 hops512-radix switch: 3 hops
64K-port fabric

20. © 2023 Marvell. All rights reserved. 20
Optimizing queuing delay
Adaptive network tuning
zSwitch congestion controlTraffic engineering
Endpoint intelligence
Pipeline programmability enables emerging network features

21. © 2023 Marvell. All rights reserved. 21
Transparency: workload-aware orchestration
…

22. © 2023 Marvell. All rights reserved. 22
Open standards enable operator visibility
Proprietary fabricStandards-based fabric
Low visibilityHigh visibility

23. © 2023 Marvell. All rights reserved. 23
Open standards enable operator differentiation
Standards-based fabricProprietary fabric

24. © 2023 Marvell. All rights reserved. 24
Open standards enable supply diversity
Standards-based fabric
OCP provides the venue for HW and SW standards development
Single
vendor
Vendor W
Vendor Z
Vendors X & Y
Proprietary fabric

25. © 2023 Marvell. All rights reserved. 25
Marvell Teralynx Ethernet Cloud Fabric
51 Tbps Performance
Low Latency
Flexible Programmable Pipeline
Leading 512 Radix
SONiC Integration
The Clean-Sheet Approach for AI Multi-Tenant Cloud
Demo:
Booth B13

26. Thank you