TDEB Top Bar

Trending
Semiconductor Industry News — The Digital Electronics Blog
Follow Us

TDEB Masthead

The Semiconductor Hub

Semiconductors.  Chips and Systems.  Innovation.

Indexed by leading Search Engines, AI and referenced by Engineers worldwide 500+ Articles 10M+ Pageviews 30+ Reports 50+ Citations

Configurable Multiprocessor Systems on Chip

Aarthi Chinappa
by
1
Share Facebook X LinkedIn WhatsApp Email

Multiple microprocessor cores that communicate over an on-chip network are increasingly found in integrated circuit designs. These systems offer very high performance while staying highly programmable and consuming modest power. Besides high-end CPUs, they suit high-performance embedded applications such as multimedia, digital signal processing and networking. Dr. Steve Guccione of Cmpware, Inc. presented an invited paper on this topic at ERSA 2005 (the International Conference on Engineering of Reconfigurable Systems and Algorithms). His paper and slides were posted online along with a free development kit at Cmpware (archived copy of the 2006 site; the company site is no longer online).

What is a configurable multiprocessor SoC?

A configurable multiprocessor system on chip puts several processor cores on one die and lets the designer choose how many cores there are, what each core looks like, and how they are connected. Instead of one large, fast processor, the chip has many smaller ones working in parallel, each running part of the job.

The cores need a way to exchange data. In a small system, a shared bus is enough. As the number of cores grows, a bus becomes a bottleneck, because only one transfer can happen at a time. An on-chip network solves this the way a computer network does: each core connects to a small router, and data travels as packets from router to router, so many transfers can happen at once.

Why designers use them

  • Performance per watt: several cores at a modest clock can do the work of one core at a much higher clock, and the lower clock lets each core run at a lower voltage.
  • Programmability: unlike fixed-function hardware, the same chip can be reprogrammed for a new codec, protocol or algorithm.
  • Fit to the problem: media, signal processing and packet processing break naturally into parallel tasks or pipeline stages, one per core.

The catch is that the software has to be split across the cores, and the designer has to choose the right number of cores, the right core configuration and the right network before the chip exists. That is where simulation tools come in.

What the Cmpware kit offered

According to the Cmpware product brief (version 2.0, January 2006, archived), the Cmpware Configurable Multiprocessor Development Kit (CMP-DK) was an Eclipse-based environment for architectural exploration and application programming of multiprocessor systems. Its listed features were fast processor simulation (2 million cycles per second), easy processor definition, configurable and user-definable networks, built-in profiling and performance monitoring, access to all system state, and custom hardware model definition. The brief describes it as usable for ASICs, SoCs and FPGAs.

Deep dive: the numbers and details

This part holds the bandwidth arithmetic, the speedup limit and what simulation speed means in practice. Skip it if the overview is all you need.

Bus versus network: a bandwidth example

Suppose 16 cores each need 400 MB/s of traffic to a neighbor, and each link, bus or network, carries 32 bits at 200 MHz, which is 800 MB/s.

InterconnectAggregate bandwidthDemandResult
Shared bus800 MB/s (one transfer at a time)16 x 400 = 6,400 MB/s8 times oversubscribed
4 x 4 mesh network24 links (12 horizontal, 12 vertical), 800 MB/s each per directionNeighbor traffic uses one link per transferEach link carries 400 MB/s of 800 MB/s available

A 4 x 4 mesh has 3 links per row times 4 rows, plus the same for the columns: 24 links that can carry traffic in both directions. With nearest-neighbor traffic, each transfer uses one link, so the network handles the load with headroom. The Wikipedia article on networks on chip summarizes the same point: links can operate at the same time on different packets, which gives high parallelism and throughput compared with buses and crossbars. The trade-off is router area and a few cycles of latency per hop.

How many cores help? Amdahl's law

Speedup(N) = 1 / ((1 - P) + P / N)

P is the fraction of the work that can run in parallel and N is the number of cores. With P = 0.9:

Cores (N)Speedup
21.82
43.08
84.71
166.40
Unlimited10

The serial 10 percent caps the speedup at 10, however many cores are added. This is why architectural exploration matters: a simulator lets you measure P for the real application and find the point where more cores stop paying for their area and power.

What 2 million cycles per second means

At the 2 million simulated cycles per second listed in the product brief, simulating one second of a chip clocked at 200 MHz takes 100 seconds of host time, if that rate applies to the whole system. If it applies per processor and the simulator runs the cores one after another, a 16-core system would take 16 times as long. Either way, it is fast enough to run real application code for many design variations in a day, which is the point of a cycle-level exploration tool, and much faster than simulating the same system at gate level.

Questions an exploration tool answers

  • How many cores, and of what type, meet the throughput target?
  • Which network topology and link width keep the busiest link below saturation?
  • Where does each task go, so that heavy traffic stays between neighbors?
  • How much local memory does each core need to avoid waiting on the network?
Share Facebook X LinkedIn WhatsApp Email

Post a Comment

1 Comments

Your comments will be moderated before it appears here.

Post a Comment

#buttons=(Ok, Go it!) #days=(20)

Our website uses cookies to enhance your experience. Learn more
Ok, Go it!