Professional Documents
Culture Documents
CPU
IO
TPU1, TPU2,TPU3 …
FPGA
FPGA AI training and Inference
AI Inference
100X improvement
in TOPS and TOPS/W
(CPU to GPU/FPGA to ASIC) GPU Auto AI Inference
Performance/power efficiency
th s TNG
th s
17B
on
Graphcore
on
A100 MK2
8 m m 54B 59B
n1 12
0
xi n
50 xi
Megatron 50
8.3B
GPT-2
BERT-Large 1.5B
340M IBM Power 7
1.2B
Rosset,MSFT
07/2018 01/2020 2010 2020
N7 N5
Marvell offers
Highest performance
technology 112G and parallel IOs
RTL
Netlist
Place
Layout
Customer Marvell
Switch / Routing IP
Chiplet
Encryption / Decryption Based
Accelerators
Compression/Decompression
Data Centers
Smart NICs
RDMA / iWarp / Network Offload Reliability and uptime 99.999%
Chiplet A Process
Safety Island / Safety Monitors ASIC
or I
based A B
Proven Interface IP accelerators B G I
G G A
A
B B B A B
I
High-bandwidth SRAM / LD / GD A A I
Secure PHY switches
Full turnkey capability Controllers Reliability and uptime 99.999%
Automotive network
© 2020 Marvell. All rights reserved. 12
Proactive measures to maintain reliability & availability
FAB & OSAT partnerships
Deep understanding of technology intrinsic failure mechanisms
Design for reliability
Test Pre-qualified reliability and quality screens from qualification test vehicles
(module that works as models
predict for supported lifetime) Enhanced escape prevention methodology for Automotive and Aerospace
2 25-year custom ASIC track record of execution with >2000 high-reliability production ASICs
4 Offering industry’s most flexible business models to enable next generation innovations