Qualcomm Centriq™ 2400

4 minutes read
iitk • computer-architecture

The Qualcomm Centriq™ 2400 is an ARM-based server processor developed by Qualcomm Datacenter Technologies, Inc. We had Dr. Niket Choudhary give a talk on the design on the processor and some of the decisions taken while developing it on 16th March 2018 at IIT Kanpur.

Motivation

In the 1990s, there were a lot of competing companies rolling out new processors all the time. After that, the market settled down a bit in the 2000s with Intel, AMD and ATI being the major competitors. With the emergence of Smartphones and the need for more power-efficient processors, a new race began. Right now, Qualcomm has firmly established itself in the mobile-market segment.

Qualcomm has a lot of experience on ARM based CPUs on the Snapdragon Processors:

  • Scorpion and Krait - ARMv7
  • Kryo - 64bit ARMv8


ARM is a pretty simple ISA which is easy to implement, improvise and is very popular on mobile devices. Its simple design allows for a power-efficient implementation ideal for mobiles.

Qualcomm Datacenter Technologies decided to use this vast experience in the mobile market in building server processors where the cost of running is also very dependent on the power requirement of the CPUs. This was the target of their attempt:

  • Build a power-efficient ARM-based many-core server chip
  • Scale-out performance/efficiency for cloud technologies
  • Disrupt the Datacenter Segment

Comparison with other Microprocessors

Microprocessor Report (Nov. 2017)

This new processor has been compared in the Microprocessor Report (Nov 2017) with competing state-of-the-art server processors like AMD EPYC 760 and Intel Xeon Platinum 8180.

As compared to the competitors, the chip has comparable performance but with huge savings on Chip Size, Cost and Power Consumption.

Cloudflare Blog

The folks at cloudflare compared the performance of this processor on benchmarks and have validated that the SoC performs better than Intel’s Broadwell and Skylake Architectures for Server Application loads like gzip, brotli compression and public key cryptography.

Qualcomm Centriq SoC Overview

Falkor CPU:

The Centriq SoC consists of 48 Falkor CPUs:

  • Fully ARMv8 Compliant
  • 64-bit only as there isn’t much legacy code in the cloud segment
  • Hypervisor Support (EL2) and TrustZone(EL3)
  • Cryptography Acceleration Instructions
  • Designed for Performance, Optimized for Power.

All Falkor CPUs occur as a duplex with a total of 24 duplexes.

Pipeline

Consists of a 4-Issue, 8-Dispatch Pipeline with variable length for different instructions. Maximum depth of 10. Uses register renaming.

Branch Prediction

The branch prediction algorithm used is state-of-the-art (Not mentioned, but probably some variant of TAGE). The penalty for a misprediction is 0-1 Cycle.

This penalty is reduced by using a Branch Target Instruction Cache (BTIC).

Caches

There is an unified L3 Cache, unified L2 Cache and L1 ICache and DCache. All CPUs have their own L1 caches. One L2 Cache is shared between a duplex while the L3 Cache is shared across all CPUs.

Dual tagging (virtual+physical address) is used to eliminate TLB from the critical path.

All caches are write-through, read-allocate and write-no-allocate caches.

The following TLBs are there:

  • 64-entry L1 Data TLB
  • 512-entry “final” L2 TLB
  • 64-entry “non-final” L2 TLB
Prefetching

Hardware stride prefetching is used in L1, L2 and L3 caches.

L1 Cache
  • 32KB 64-byte lines, 8-way associative
L2 Cache
  • 128 byte lines, 8-way associative
  • 15 cycle latency for L2 Hit
L3 Cache

Quality of Service Extensions:

  • There is a distributed L3 cache
  • Limited / No Allocation Policy Enforcement for Data
  • Fine-tuned Cache Allocation There is fine-tuned cache allocation where a QoSID is associate with cache entries allowing cache to be partitioned into more priority and lesser priority sections.
  • Memory Bandwidth Compression 128 byte -> 64 byte compression between memory controller and L3 cache Time taken for decompression is between 2-4 cycles. Compression is not used between other levels because then hit time would increase.

The Talk

The talk was appealing as I somehow managed to understand most of the terms being thrown around here and there. I have obtained knowledge about stuff like hardware prefetching, branch prediction from the CS422 course itself, so it was enlightening to see it being actually used in a real model instead of a theoretical one.

It did go a little hifi when Mainak Sir started asking questions and that was to be expected 😛. Also, I wasn’t very informed of the ring stuff with reference to which most of the questions were answered.

What Excites Me?

  • The comparable performance between ARM and x86 means that eventually we could move to RISC architectures giving us a more simplified instruction set. Most desktop processors still use x86.
  • Qualcomm using the knowledge gained from the mobile market to build cheaper and more efficient products in other markets.

Summary

The folks at QDT have targetted cloud applications and used a decade of their experience with ARM Chips to create a competitive server processor. They have focussed only on tackling issues with server applications and have managed to bring a custom-tuned processor for the datacenter market segment.