The document discusses CXL, a new open standard protocol for efficient CPU and memory connectivity. CXL allows for memory disaggregation and pooling across devices by enabling high-bandwidth, low-latency connections between CPUs, GPUs, accelerators, and memory. This helps address the growing CPU-memory bottleneck by allowing expansion of memory capacity beyond what can physically connect to the CPU. CXL also enables memory tiering by providing different performance and cost options for "near" directly attached memory versus "far" switched or fabric attached memory.
1. Tam Do, Product Marketing,
Microchip
Sanketh Srinivas, Product
Marketing, Microchip
CXL Memory Expansion, Pooling,
Sharing, FAM Enablement, and Switching
2. • Compute Memory Bottleneck
• Heterogenous architectures and Memory Efficiency Plateau
• CXL™ A Protocol For Memory
• Memory Tiering with CXLTM
• CXLTM Switching
Agenda
3. • CPU vendors are introducing powerful CPUs with higher core counts at a
rapid pace
• Amount of DDR memory that could be directly attached to CPU is limited
due to
• Number of parallel DDR interfaces/channels per CPU
• Requires large number of IO pins for each memory channel (~300
pins)
• Adding more interfaces will impact CPU Die and package size
Compute Memory Bottleneck
Host CPU DDR Memory
CORE CORE CORE CORE
Parallel DDR
Interface
CPU to Memory
Bottleneck
Memory
Controller
4. CPU
CPU
GPU
GPU
Accel
Accel
Heterogeneous Compute resources
Memory and Storage
resources
Data
Global data
Resources and
generation
CPUs, GPUs, Custom
Accelerators
• Analysis extraction
• Medical
outcomes
• Self-driving
vehicles
• Filtered and sorting
• Data analytics
• Cat photos
• Time critical data
• Stock market
decisions
• Task Prioritization
• Augmented reality
• Advertising
• Catastrophe avoidance
Heterogenous architectures and Memory
Efficiency Plateau
5. CXL Disaggregates Memory to Allow
Expansion
• Lowering pin-count allows
expansion
• More memory added to support
more processor cores
• Allow memory pooling
and multiple processors to share
memory and data more efficiently
6. • CXL enables Memory Diversity with the same look and feel
abstracting different memory types to be supported behind a
memory controller
Memory Tiering with CXLTM
7. Memory Tiering Model with CXL™
Storage
• On-package memory/High Bandwidth Memory (HBM)
• DDR direct-attached
• CXL direct-attached
• CXL Switch-attached
• CXL Fabric-attached or GFAM
OPM
DRAM (Hot)
DRAM (Warm)
SCM (Cool)
DRAM/SCM (Cold)
Direct attach
“Near” Memory
Switch/fabric attach
“Far” Memory
Microchip Memory Infrastructure Products will enable memory tiering to optimize performance
and cost for application-dependent workloads
8. Memory Tiering with CXL Switch-Attached
Memory Expansion
CPU
SSD
SSD
SSD HDD
HDD
HDD
CPU
CPU
CPU
IO Controller
I
O
C
T
R
L
M
E
M
C
T
R
L
L3 Cache
DDR5 DIMMs
L1 Cache
L2 Cache
DDR5 DIMMs
CXL
Switch
DDR5 DIMMs
M
E
M
C
T
R
L
DDR4 DIMMs
Low Latency NAND or PMEM
CXL
Flash
Controll
er
10. CXL™ Memory Performance
Comparison
CPU DDR
CXL
CXL
OS
70 ns 30 ns
100 ns
100 ns
100 ns
40 ns
70 ns
150 ns
Use-Case Load to Use
Direct DDR ~100 ns
Direct CXL ~170 ns
Pooled CXL ~200 ns
Switched CXL ~320 ns
Memory
Controller
DRAM
DRAM
DRAM
30 ns
DRAM
DRAM
DRAM
CXL
CXL Multi-head
Memory
Controller
30 ns
DRAM
DRAM
DRAM
Switch
40 ns
Memory
Controller
30 ns
DRAM
DRAM
DRAM
12. • Connects multiple hosts and
memory devices
• Provides low latency, high-
bandwidth interfaces between
various CXL devices, such as CPUs,
GPUs and memory modules
• It enables these devices to share
data through multiple host
platforms
• Supports downstream PCIe® links
Flexible CXLTM Switching
Root
Port
CXL Switch
CXL
Type 2
CXL
Type 3
CXL Switch
CXL
EP
PCIe
EP
13. • CXL 3.0 enables non-tree
topologies and peer-to-
peer communication
within a virtual hierarchy
of devices
• PCIe® is tree topology
CXLTM 3.0: Device to Device
Connectivity
16. • Using PCI Express® and (PCIe®) physical and electrical interface standard
• Enables efficient high bandwidth interface between CPU, memory and
accelerators
• Maintains memory coherency between CPU memory space and memory
on CXL attached devices that enables resource sharing, memory
disaggregation with memory pooling and sharing
• Open industry standard starting with CXL 1.0 to 3
CXL™ A Protocol For Memory
17. • How to get involved in the Project/Sub-Project Community
• Timeline for Contribution Availability
• Timeline for Product/Facility Availability
• Link to Contribution DB/OCP Marketplace
• Where to find additional information (URL links)
• [Example] Where to buy: https://www.opencompute.org/products
• [Example] Project Wiki with latest specification :
http://www.opencompute.org/wiki/Server/Mezz
• [Example] Mailing list: http://lists.opencompute.org/mailman/listinfo/opencompute-
mezz-card
[Call to Action] [MANDATORY]