Maximizing heterogeneous system performance with ARM interconnect and CCIX

Transcription

1 Maximizing heterogeneous system performance with ARM interconnect and CCIX Neil Parris, Director of product marketing Systems and software group, ARM Teratec June 2017

2 Intelligent flexible cloud to enable new use cases Compute Acceleration 2

3 Distributing intelligence efficiently at scale Cortex-A75: Ground-breaking performance Cortex-A55: High-efficiency redefined 3

4 Unveiling DynamIQ New single cluster design New application processors Cortex-A75 and Cortex-A55 CPUs Built for DynamIQ technology Cortex-A75 32b/64b CPU.. Cortex-A55 32b/64b CPU Private L2 and shared L3 caches Local cache close to processors L3 cache shared between all cores DynamIQ shared unit Contains L3, Snoop Control Unit (SCU) and all cluster interfaces New AMBA 5 CHI bus interface Private L2 cache Private L2 cache SCU Peripheral Port Async Bridges ACP AMBA4 ACE or AMBA5 CHI Shared L3 cache DynamIQ Shared Unit (DSU) DDR DRAM 4

5 Build higher performance systems Boost performance Up to 7.5x more compute Linux Kernel ARM Trusted Firmware UEFI boot Fixed Virtual Platform (System Model) Innovations to increase throughput Up to 3x more packets per second Tailor solutions from edge to cloud Scale from 1 to 256 CPUs on a single chip Multichip connectivity supporting CCIX Scale to CPUs Cortex-A75 Cortex-A75 Cortex-A75 Cortex-A75 DynamIQ DSU & L3 Cache SCP MCP Peripherals Open Source GIC to 32 Clusters Cortex-A75 Cortex-A75 Cortex-A75 Cortex-A75 Cortex-A75 DynamIQ DSU & L3 Cache PCIe/25GbE Coherent Mesh Network CoreSight Non-ARM IP ARM IP MMU to 8 Controllers DDR DRAM Storage NIC-450 5

6 Relative Performance CMN-600: Delivering maximum compute density x compute Cortex-A57 CoreLink CCN-504 2x DMC Cortex-A72 CoreLink CCN-508 4x DMC Cortex-A75 8x 5x throughput 3x Compute = measured by specint2k6_rate Throughput = achieved requested bandwidth Same process node and test conditions. 6

7 Scalable solutions from core to edge Scalable system configurations Access point 1 Cortex-A CPUs 256 Data center compute PCIe 100GbE Cortex-A Accelerator IO IO NIC-450 0MB System cache 128MB CPU CPU CPU CPU CPU CPU CPU CPU System cache 20 GB/s Bandwidth 1 TB/s CPU CPU CPU CPU CPU CPU CPU CPU 1 DDR channels 8 7

8 ARM AMBA 5 CHI coherency protocol Credited, non-blocking, coherence protocol Protocol Messages Protocol Capable of achieving high frequencies (>2GHz) Routing Link Physical Packets Flits Phits Routing Link Physical Layered architecture definition Protocol/Routing/Link/Physical layers Topology agnostic Flexible topologies for target applications 8

9 Innovations to increase throughput Up to 3x more throughput with cache stashing Used for acceleration, network, storage use-cases Critical data stays on-chip Accelerator or IO selects data placement Low latency access by a CPU or group of CPUs Accelerator or IO Cortex-A CPU Cortex-A CPU CPU CPU CPU CPU L2 L2 Cache L2 L2 L2 Cache L2 Cache Cache L2 L2 Cache Cache L3 Cache L3 Cache Stash to any cache level CPU L2 cache private to a single CPU thread DynamIQ L3 Cache shared by CPUs in a cluster Agile System Cache shared by all DynamIQ clusters Stash critical data to any cache level Agile System Cache DDR4 Agile System Cache DDR4 9

10 Local Accelerator Customized, tiered acceleration platform Closely coupled cluster accelerator Low latency control and data exchange Global SoC accelerator or IO Direct, high bandwidth path to memory Cortex-A DSU Global Accelerator or IO Off-chip coherent accelerator High performance CCIX connectivity Agile System Cache Agile System Cache Remote Accelerator DDR4 DDR4 Agile System Cache Shared heterogenous system cache 10

11 Coherent multichip 11

12 Interconnects needs at different scale SoC interconnect Connectivity for on-chip processors, accelerator, IO and memory elements. Server node interconnect - scale-up Simple multichip interconnect (typically PCIe) topology on a PCB motherboard with simple switches and expansion connectors Rack interconnect - scale-out Scale-out capabilities with complex topologies connecting 1000 s of server nodes and storage elements. 12

13 Scale-up server node compute and acceleration Accelerator Smart Network Persistent Memory Shared virtual memory system 13

14 Cache coherent interconnect for accelerators Extends cache coherency to the multi-chip use cases Intelligent network/storage, machine learning, image/video analytics, search and 4G/5G wireless Open standard to foster innovation, collaboration and choice to solve emerging workload challenges Hardware specification available for designs start for member companies 14

15 CCIX multichip connectivity and topologies New class of interconnect providing high performance, low latency for new accelerators use cases CCIX defines up to 25GT/s (3x performance*) Examining 56GT/s (7x performance*) and beyond Enabling low latency via light transaction layer Compute Node Switch Accelerator Smart Network Flexible, scalable interconnect topologies Persistent Memory Flexible point-to-point, daisy chained and switched topologies Simplified deployment by leveraging existing PCIe hardware and software infrastructure Runs on existing PCIe transport layer and management stack Coexist with legacy PCIe designs * Note: Based on PCIe Gen3 Performance 15

16 Extending the benefits of cache coherency CCIX provides true peer processing with shared memory Accelerators may cache data from processor memory Coherency eliminates the software and DMA overhead of transferring data between devices Compute Node DDR Memory shared virtual memory Accelerator Cache Coherency protocol easily adaptable to other transport in the future shared virtual memory Supports all major instruction set architectures (ISA) enabling innovation and flexibility in accelerator system design Compute Node DDR Memory Accelerator Cache DDR Memory 16

17 CCIX example request to home data flows Accelerator shares processor memory Shared processor and accelerator memory Requestor Requestor Cache Cache Cache Cache Home Memory Memory Memory Daisy chain to shared processor memory Shared memory with aggregation Cache Cache Cache Cache Cache Memory Memory Memory 17

18 RNI PCIe Transaction Layer Data Link Layer PHY (16Gpbs) Today s existing PCIe integration Example CMN-600 mesh design PCIe IP connects to a CMN IO interface via AXI Cadence IP XP 16 Lanes AXI 18

19 RNI PCIe Transaction Layer Data Link Layer PHY (16Gpbs) CML Coherent Multi-chip Link (CML) Example CMN-600 mesh design CML converts CHI to CCIX messages CXS - CCIX stream interface Cadence IP CXS XP 16 Lanes AXI 19

20 RNI PCIe Transaction Layer Data Link Layer CCIX Transaction Layer PHY (up to 25Gpbs) CML Cadence CCIX integration Example CMN-600 mesh design Low latency CCIX transaction layer Cadence IP Support up to 25Gbps vs 16Gbps PCIe Gen4 CXS XP 16 Lanes AXI 20

21 CCIX 25Gbps technology demonstration 3X faster transfer speed with CCIX vs existing PCIe Gen3 solutions Transferring of a data pattern at 25 Gbps between two FPGAs Channel comprised of an Amphenol/FCI PCI Express CEM connector and a trace card Transceivers are electrically compliant with CCIX Fastest data transfer between accelerators over PCI Express connections Xilinx and Amphenol FCI first public CCIX technology demo 25Gbps-Performance-over-PCIe/ba-p/

22 Scale-up server node performance with CCIX CCIX is a class of interconnect providing high performance, low latency for new accelerator use cases Easy adoption and simplified development by leveraging today s data centers infrastructure Optimize CCIX SoCs with ARM and CCIX controller IP Server, FPGAs, GPUs, network/storage adapters, intelligent networks and custom ASICs Accelerator Smart Network 22 Persistent Memory

23 The trademarks featured in this presentation are registered and/or unregistered trademarks of ARM Limited (or its subsidiaries) in the EU and/or elsewhere. All rights reserved. All other marks featured may be trademarks of their respective owners. Copyright 2017 ARM Limited