Sonics and the Art of Making Chips Wait Less

Imagine buying a magnificent kitchen, then discovering that every cook must use the same narrow corridor. The oven is excellent. The knives are excellent. Dinner is late. A system-on-chip can suffer a similar indignity: capable processors, graphics engines and other blocks, all waiting for data. Sonics built a business in that corridor. Its customers made the chips. Sonics supplied designs for the connections, the queues and, eventually, the decisions about which parts could take a nap.

The traffic is inside the chip

Grant Pierce and Drew Wingard founded Sonics in 1996. The name is an acronym for Systems On ICs, which is admirably literal for a company selling something few people would ever see. The premise was that growing chips would contain more reusable intellectual property, or IP: predesigned functions assembled into a larger system. Someone still had to make those functions communicate.

Here lies the small trap in the word “integration.” Putting things together sounds like the last task on the list. In chip design, it can determine whether the list was sensible in the first place. Which block can reach which resource? How long does a request wait? What happens when several blocks want memory at once? A specification for each component does not settle the argument between them.

Sonics’ network-on-chip products made those connections configurable. The company also supplied tools to explore and test the resulting design. Its position in the market was system-level semiconductor IP: the infrastructure around computing functions, sold to the people assembling them. Arteris was a direct rival. Building the interconnect internally remained another option. The decision was partly whether a design team wanted to keep solving that class of problem itself.

A Sonics report on an October 2012 survey of 318 design and verification professionals put a number on the unglamorous work. Respondents who could estimate their time reported an average of 28 percent spent designing, modifying or verifying on-chip communications networks. That is a survey result from a particular period, not a universal law. It does, however, make “just connect the parts” sound rather expensive.

A memory queue with manners

A useful way into Sonics’ expertise is MemMax. Its job was to schedule requests between on-chip blocks and off-chip DRAM. Memory traffic does not arrive with identical needs. A stream that must keep a display supplied has different tolerances from work that can wait. Serving whoever shouts first is a policy, but seldom a satisfying one.

Sonics’ 2003 Toshiba announcement described MemMax examining access patterns and reordering requests to improve DRAM utilisation while preserving quality-of-service requirements for individual blocks. The clever part was the combination. Rearranging a queue can increase efficiency; respecting each participant’s needs keeps that efficiency from becoming somebody else’s missed deadline.

The distinction runs through the company’s technical work. A paper by Sonics engineers presented at DATE in 2005 described an arbitration scheme that balanced service latency, jitter and bandwidth guarantees. Jitter is variation in timing: the difference between a dependable delivery and one that arrives whenever it pleases. This was expertise in allocating a shared resource under competing demands. The wire was only part of the product.

“SoC design complexity has reached integration of 100 or more IP cores per chip in many cases.”

Drew Wingard · SonicsStudio announcement · 2014

By 2014, SonicsStudio Director gave customers a visual environment for configuring IP, analysing performance and working through the design flow. It also offered Tcl scripting for experienced users. That pairing makes commercial sense: a schematic helps you understand a design; automation helps you change it repeatedly without turning the afternoon into a clerical career.

The customer buys the next design, too

The Toshiba deal was corporate-wide. Announced in April 2003, it covered SiliconBackplane interconnect and MemMax scheduler IP for Toshiba business units, wholly owned subsidiaries and customers. Its promise extended beyond one chip: reuse the blocks and the integration work across future designs. A semiconductor company could concentrate its effort on what made its product different while buying some of the infrastructure it repeatedly needed.

This was a licensing business. Public historical contracts describe license rights, support obligations and royalties on devices sold. They also redact commercial rates. The practical cost was therefore negotiated around the agreement, rather than attached to a checkout button. SonicsStudio Director was announced on a subscription basis for IP customers. The economics joined the initial engineering decision to the eventual production volume.

In December 2005, Sonics announced four consecutive profitable quarters and said its growing royalty revenue accompanied increasing license revenue. Samsung Ventures invested in its Series D preferred stock, extending an existing Samsung licensing relationship. Dealroom records the round at roughly $12 million. That figure describes financing, not the cost of a customer’s chip design.

The customers were not merely names on a slide. In January 2017, MediaTek licensed multiple copies of SonicsSX and MemMax. In February 2018, Sonics announced that Starblaze’s STAR1000 SSD controller had reached volume production with SonicsGN connecting more than 60 cores. Here was the corridor in an actual building: a storage controller with many participants and a shared need to move information.

Sonics’ later announcements reported more than four billion customer SoCs shipped using its IP. Keep the noun in view: customer chips. Sonics was selling reusable designs, not operating a factory that produced four billion finished devices. The claim illustrates how an infrastructure supplier can have an enormous installed footprint while remaining largely unknown to the people using the products.

Giving idle circuits permission to sleep

Communication raised another question. If a block has no useful work to do, why keep paying to keep it awake? In May 2015, Sonics introduced its ICE-Grain Power Architecture. ICE meant Instant Control of Energy. The architecture combined configurable hardware, embedded control software and design tools, dividing a chip into smaller power-management “grains.”

In May 2016 came ICE-G1, an Energy Processing Unit. Its purpose was to exploit circuit idle periods using autonomous hardware control. Sonics positioned it for power-sensitive chips, including mobile, wearable and other constrained devices. It could work with third-party IP and without a Sonics network, giving the company an adjacent product rather than a feature available only to existing interconnect buyers.

The idea has an obvious condition: the idle interval must offer a worthwhile opportunity after the costs and delays of changing power state. A busy circuit cannot save energy by abandoning its work, and a sleeping one must return in time. The appeal of hardware control was responsiveness to those low-level events. The engineering task still required suitable partitioning, dependency handling and verification.

The product family expanded. ICE-P3 brought dynamic voltage and frequency scaling, with temperature-aware control, into the offering in 2017. The progression makes the strategy visible. Sonics had learned about communication and power constraints through chip integration; it packaged more of that neighbouring expertise for customers. The company broadened what it sold while staying close to the integration problems its customers already faced.

A chip company exits through a social network

The business also depended on helping customers implement the designs. A 2018 Inomize partnership addressed automotive functional safety work around configurable NoCs, including implementation, verification and safety documentation. A Synkom partnership brought NoC and EPU expertise into Japanese design-service projects. These arrangements show the limits of selling a reusable block: a customer still has a particular chip to finish.

Then the supplier itself changed. In March 2019, Facebook confirmed that it had acquired Sonics. Its stated initial focus was AR and VR. Sonics announced the wind-down of its business, and Wingard moved into silicon methodology at Facebook. The independent vendor’s chapter ended; its team and technology entered a company developing its own products.

For a reader building anything complicated, the portable lesson is to study the hand-offs. Measure waiting as well as working. Make repeated integration work reusable. Sell tools and implementation knowledge alongside the component. Those lessons are useful when coordination is a meaningful cost; a simple system with little contention may have little to gain from elaborate infrastructure. Sonics made its case where many capable parts had to share resources. The remarkable object was the chip. The business opportunity was getting its occupants to agree.