Samsung pushes ahead with 20nm DDR3 RAM, signaling uncertainty about DDR4

Samsung has announced that it is pushing forward 20nm DDR3 DRAM, and has already begun volume production on the next-generation lithography node. Some of you may recall that Samsung has been building NAND flash on 20nm for quite some time, but it’s significantly easier to build NAND flash on next-gen processes than it is to move DRAM to those same nodes. Samsung’s decision to deploy DDR3 at the 20nm node also says interesting things about its DDR4 timelines and confidence in those roadmaps.

Historically, DRAM has been stuck between a rock and a hard place. CPUs from AMD, Intel, and even ARM vendors can afford to be more aggressive with process node shifts; the chips command higher margins and they benefit from being “first” to a given node. NAND flash doesn’t necessarily have the margins of conventional microprocessors, but it’s much simpler to manufacture than DRAM and the company that secures an early lead in manufacturing tends to win the majority of the revenue in this space as well. DRAM, though, isn’t as easy to manufacture as NAND, but it’s nowhere near as profitable as CPUs.

Samsung’s first 20nm DDR3 product: SODIMMs based on a new 4Gb DRAM chip

By moving DDR3 to 20nm, Samsung is setting itself up for significant power savings (new modules supposedly draw 25% less power than previous designs), but it’s also a statement about where the company expects the memory market to go in the next few years. In the past, the memory market was driven by the demands of the desktop and server space first and foremost. True, a low-voltage option existed for smartphones and tablets, but these mobile devices themselves weren’t the driving force behind the space.

DDR3 20nm improves the DDR3 situation, but narrows the gap between itself and DDR4

Ideally (from Samsung’s perspective), it would reserve the 20nm node for its DDR4 products, ship those out in standard and LP form factors, and thus position itself to capture the vast majority of the revenue from this segment. Instead, the company is simultaneously transitioning both DDR3 and DDR4 to the 20nm node. That suggests Samsung doesn’t expect to see a strong shift to DDR4 in the near term and is hedging its bets on the technology accordingly.

Longer time frames, slower updates

Sebastian Anthony’s ugly mug reflected in a beautiful 16nm NAND wafer (Micron I think)

Slowing the DDR4 ramp-up only makes sense given the overall market. The vast majority of the market is no longer upgrading every 2-3 years, which means Samsung faces the possibility of amortizing the steep costs of a new new technology and production node over a long period of time. Putting DDR3 on 20nm may actually improve the economics of DDR4 — it makes it easier to bear the cost of the new RAM, even if bringing those same power savings over to the new platform makes DDR4 relatively less attractive.

We’ll undoubtedly see a few mobile companies start talking about LPDDR4 in 2014, but bulk shipments on the new RAM standard won’t happen immediately — for that, we’ll probably have to wait for the next consumer chipset after Haswell-E. Meanwhile, 20nm LPDDR3 will still offer manufacturers that haven’t designed new SoCs a way to continue using previous architectures but still offer customers a density improvement, cost savings, and power reduction.

Samsung still claims that DDR4 offers a performance and power reduction benefit, but the near-term benefits from shifting to the standard when it debuts later this year are likely to be small — and smaller still when compared against 20nm DDR3. Still, for companies struggling to deal with the high cost of next-generation foundry technologies, “better late than never” still definitely applies.