Perhaps it is not fair to title an article in a way that suggests a company has disappeared from the market or requires a triumphant return. After all, it is not the case for Rambus, which was founded in 1990 to address a familiar and still current set of challenges around reducing memory bottlenecks. Rather, the real “return” of Rambus is more symbolic.
After years of extensive, expensive litigation with most of the memory industry, the company has settled its pending suits and is moving on with a fresh focus on delivering products, in addition to having a rich briefcase full of patents and license options. The concept of a return also implies that Rambus is ready to move past one of its biggest challenges—one that is more nebulous and non-technical. That would be the company’s reputation, something that Ely Tsern, the company’s chief technologist and vice president for memory interfaces says is not something it expects to gloss over.
In a chat with The Next Platform, Tsern made no bones about the company’s involvement with some particularly vicious suits with the likes of Micron Technology and Samsung, and further, he says, having watched the leadership evolution over his twenty-year career at Rambus, he sees how they are finally ready to move out of the past and roll into the future with a product emphasis—although of course, patent and technology licensing are still the keys to the Rambus business.
And make no mistake, it is quite a business. Tsern says Rambus is currently a $300 million company (yes) with around 500 employees, the majority of which are seated in the engineering camp. As part of its enhanced focus on actual products, Rambus will continue to have a division comprised of silicon designers who work on blocks that go into chips that are licensed and just this week, Rambus announced actual product in the form of buffer chipsets for DDR4 DIMMs for the server market, which have been in development for two years. While the chipsets are interesting in themselves, and there is some deep coverage of it here, the undercurrent here is that there must be some hidden value in a server main memory market that is, as Micron has outlined, just not exactly a smooth business to be in for a host of reasons.
The fact is, Rambus is pushing into a difficult DRAM market where it now will be more reliant than ever on the subjects of years of litigation. But that is all turned around now, Tsern asserts, noting that “what is interesting about this new chipset is that it bridges across multiple areas; it hits on the memory manufacturers that are our licensees—people we were in litigation with but who now we are working very closely with as our customers. This includes Samsung, SK Hynix, and Micron, the three big ones there—and they’re the ones who are delivering the majority of server memory modules.” Indeed, this is something close to 98 percent of the entire market, which is actually beneficial, says Tsern, to its customers, the OEMs, who further validate its products and push Rambus along.
But here is the real question, beyond anything we could dig up from the past that feeds the Rambus story with a great deal of negativity from the outside world. Why now? Why, given what we know about the state of the DRAM market, would Rambus choose this time to make inroads in server main memory, especially given its history with the key vendors (even though there is collaboration post-settlement).
“Memory today, broadly speaking, not just DRAM, is unlike previous years getting interesting. DRAM is the incumbent and there are all these other dynamics influencing memory close to the processor. There are people putting flash close to the processor not just for storage, the 3D XPoint announcement will get a lot of momentum. But we look at it as DRAM is still driving a lot of the volume. We want to establish with that volume with the incumbent—and that is making DRAM as fast as possible.”
“We chose to move into the chip market, and the server datacenter space, because it is dynamic from a hardware platform perspective. There’s a lot happening right now around data, and the applications being driven there, especially around real-time needs across market segments.” In addition to seeing increased demand in keeping the data and processor as hooked in as possible, Tsern says that there are also changes in the way systems are being designed, configured, and purchased with higher performance and density—both sweet spots for what Rambus is targeting with the new chips.
In terms of what the market looks like overall, it is clear that DDR4 will be around for some time, at least the next five years, but the transition to DDR4 within the last year has come with a steep adoption curve. Estimates put DDR4 adoption in the 20 percent range, but that is expected to skyrocket.
“In two to three years, most of the market will be at DDR4, so there’s a timing aspect to this we see, but also, DDR4 is very tough technically,” Tsern says. “It’s not like previous generations of DRAM. This is what we do, signal integrity and the way the channel electrically runs is really putting us against a wall electrically. It’s hard to get the speeds up there on DDR4 and while Intel is driving the speed to bump each year with new processors and we share these challenges and it’s aligned with the sweet spot of what we do.”
But let’s put this market niche in context. On the one side, there is the $40-plus billion dollar chip industry. On a related end, there is the memory business, which itself weighs in around $10 billion to $15 billion. But in between there, where there is that slim but important market for buffer chips for server DIMMs, which are only offered by a few relatively small companies. IDT, for instance, plays here, but it is only a smaller slice of their base, as is the case with M5. “This market has grown recently—a couple of years ago, it was maybe $100 million to $150 million a year, in about two to three years we see it growing over $400 million a year. It’s because DDR4 is running faster and Intel wants the server modules to use these chipsets, so there’s growth there as well as the fact that the higher capacity modules are increasing as well due to the data demands for a lot of emerging applications.”
All of this feeds into the approach with the new Rambus chipset, which is an electrical, high speed chip that can, in concept, talk to anything; from flash, to DRAM, to eventually 3D XPoint for the general memory space. This is an interesting addition, given the new options available, and is a springboard for new chips Rambus will consider developing once they stabilize their product line with the volume benefits of DRAM.
“We are working closely with Intel to get our chips validated. We are working closely with other memory manufacturers like Micron and SK Hynix to branch out. We’re sampling today with this, it’s moving forward,” says Tsern. “The general goal from a performance standpoint is allowing systems to have lots of modules and capacity and run fast, no matter what memory is on the other side.”
This is as close as Rambus came to explaining the company’s roadmap. There are very few companies that do what Rambus does, and no matter how it got there, after over 20 years of focusing on the exact problems that are most pressing architecturally going forward (and in the wake of so many new coming memory architectures) we might consider a full return of Rambus to the fore, although in the quiet, behind-the-scenes role that it has played in the last several years.