Cisco has decided to buy memory-controller specialist Memoir Systems and absorb the technology into its Insieme business unit, which specializes in data-center switch technologies, although the communications company has indicated that the memory controllers may be used in switch ASICs made by other divisions.
The move underlines the fragility of the IP startup ecosystem in the way that a large chipmaker or system house can suddenly lock technologies away from competitors. Cisco’s move comes only a month after memory architectures hit the news.
The mini-meltdown by the internet in mid-August was a symptom of how important smart memory can be. Some smaller routers – primarily made by Cisco – at the edge of the core network suffered a hiccup when the number of possible routes on the internet exceeded the amount of memory available in their content-addressable memory (CAM). They had to overflow into the standard memory with additional processing by the CPUs, which slowed the router by orders of magnitude, resulting in lost data or heavily delayed packets.
Although the problem was fixed temporarily by allocating more CAM to regular IPv4 addresses, the number of discrete routes continues to increase and will eventually exceed the physical memory in these routers. The affected ISPs will ultimately need new hardware. But simply adding more CAM is an expensive option, so router and switch makers are turning to other forms of smart memory and processing to provide hardware that can scale with the still-growing internet. However, it looks as though networking design teams outside Cisco will have to look elsewhere for lower-cost smart memory techniques.
One of the arguments that Cadence Design Systems has used for buying heavily into the IP business, not unreasonably, is that as a much larger neutral supplier of tools to chipmakers it represents not only a relatively secure home for IP but a more attractive one for smaller chipmakers. More established IP suppliers such as ARM and Imagination Technologies can simply use their market capitalisation as a defence, with a possible backup of antitrust legislation if a chipmaker or systems company with a particularly strong market position decided to choke off the supply of IP to others.
The full lock-up doesn’t always happen. When Qualcomm bought SoC interconnect technology from Arteris, it did not get the company in its entirety although it did acquire key engineers forcing the startup to rebuild its senior engineering team. However, the deal did, naturally, unsettle potential customers, focusing attention on who owns which bit of IP and how that might affect future relationships.
The issue is unlikely to go away anytime soon. The desires of VCs and startups are somewhat misaligned. VCs have the job of selling a company not its products and, very often, the value of a technology company in its early days is almost entirely at odds with its actual sales. That has a knock-on effect on the willingness of the chipmakers and systems companies to engage with startups as they have to consider the risk of depending on technology that may not be pulled from their hands immediately but will not be a dependable foundation for future designs. There does not seem to be any easy way out of the dilemma as deals designed to lock out potential purchasers simply lowers the market value of the company at a sensitive stage, although a more open deal would most likely increase the value of the technology.