Edit Values
TaiShan v110 µarch
General Info
Arch Type	CPU
Designer	HiSilicon
Manufacturer	TSMC
Introduction	2019
Process	7 nm
Core Configs	32, 48, 64
Pipeline
Type	Superscalar, Superpipeline
OoOE	Yes
Speculative	Yes
Reg Renaming	Yes
Decode	4-way
Instructions
ISA	ARMv8.2-A
Extensions	NEON
Cache
L1I Cache	64 KiB/core
L1D Cache	64 KiB/core
L2 Cache	512 KiB/core
L3 Cache	1 MiB/core
Succession
TaiShan v100

TaiShan v110 is the successor to the TaiShan v100, a high-performance ARM server microarchitecture designed by HiSilicon for Huawei's own TaiShan servers.

Brands

TaiShan-based CPUs are branded as the Kunpeng 920 series.

Release Dates

Kunpeng 920 CPUs were officially launched in early 2019.

Architecture

Overview

Key changes from TaiShan v100

• TSMC 7 nm HPC process (from 16 nm)
• 2x core count (64, up from 32)
  • Custom cores (from Cortex-A72)
• Memory
  • 2x memory channels (8, up from 4)
• I/O
  • PCIe Gen 4 (from Gen 3)

This list is incomplete; you can help by expanding it.

Block Diagram

Entire Chip

Memory Hierarchy

• Cache
  • L1I Cache
    • 64 KiB/core, private
  • L1D Cache
    • 64 KiB/core, private
  • L2 Cache
    • 512 KiB/core, private
  • L3 Cache
    • 1 MiB/core
    • Shared by all cores
  • System DRAM
    • 1 TiB Max Memory / socket
    • 8 Channels
    • DDR4, up to 2933 MT/s
      • 1 DPC and 2 DPC support
    • 8 B/cycle/channel (@ memory clock)
    • ECC, SDDC, DDDC

Overview

Overview

Though HiSilicon has a history of designing Arm processors. The TaiShan v110 core is HiSilicons' first custom homegrown high-performance ARM core and SoC design. The chip, which incorporates multiple compute dies and an I/O is a multi-chip package, is fabricated on TSMC's 7-nanometers HPC process and integrates up to 64 cores and up to 64 MiB of last level cache.

The SoC also incorporates a number of hardware accelerators. There is a crypto engine that supports AES, DES/3DES, MD5, SHA1, SHA2, HMAC, CMAC with throughputs of up to 100 Gbit/s. Additionally, there is also a compression engine supporting GZIP, LZS, LZ4 with compression throughputs of up to 40 Gbit/s and decompression of p to 100 Gbit/s.

Marketed as the Kunpeng 920, this SoC supports up to 4-way multiprocessing support through HiSilicon's Hydra interface. In order to keep the cores fed, eight DDR4 memory channels are incorporated per socket. Additionally, designed to facilitate an easy accelerator platform, there are 40 PCIe Gen 4 lanes provided per socket with CCIX support, enabling cache coherency.

Core

Each core is a 4-way out-of-order superscalar that implements the ARMv8.2-A ISA. Huawei stated that the core supports almost all the ARMv8.4 features with a few exceptions, including dot product and the FP16 FML extension. It features private 64 KiB L1 instruction and data caches as well as 512 KiB of private L2. Though light on details, Huawei says that compared to Arm's Cortex cores, their core features an improved memory subsystem, a larger number of execution units, and a better branch predictor.

ASIMD

Each core features a single 128-bit NEON unit. It is capable of executing single double-precision FMA vector instruction per cycle or two single-precision vector instructions per cycle. Operating at 2 GHz, a 64-core chip will have a peak compute of 512 GigaFLOPS of double-precision floating point.

MCP physical design

The SoC itself comprises 3 dies - two Super CPU Cluster (SCCL) compute dies and a Super IO Cluster (SICL). The SCCL compute dies contains 8 CPU Clusters (CCLs), memory controllers, and the L3 cache block. There are eight CCLs on each of the SICL dies for a total of 64 cores. The CCLs are TaiShan V110 quadplex along with the L3 cache tags partition. The Super IO Clusters include the various I/O peripherals including PCIe Gen 4, SAS, the network interface controllers, and the Hydra links.

Scalability

See also: Hydra Interface

Each chip incorporates three Hydra interface ports. The Hydra interface facilitates the cache coherency between the dies on the chip. Every link supports 240 Gb/s (30 GB/s) of peak bandwidth for a total aggregated bandwidth of 720 Gb/s (90 GB/s) in a 2-way symmetric multiprocessing configuration.

With all three links, there is also support for 4-way SMP. In this configuration, one link from each socket is connected to another socket for an all-for-all connection.

Die

• TSMC 7 nm HPC
• 20,000,000,000 transistors
  • 3-4 dies

All TaiShan Chips

This section is empty; you can help add the missing info by editing this page.

Bibliography

• Huawei. Personal Communication. 2019
• Huawei Connect 2018. October 2018
• HiSilicon Event. January 7, 2019