Nvidia’s Pascal GPU family has been tremendously successful, but the company is already looking ahead to a new architecture, codenamed Volta. Details on the chip are scarce, though Nvidia has previously given guidance that Volta would be its first HBM2 product and would feature a stacked memory architecture. Now, rumors are flying that the chip will actually be built on TSMC’s “new” 12nm node.
The reason we’ve put the word in quotes is because 12nm isn’t really a new node at all. Here’s how TSMC’s CC Wei described it in TSMC’s last quarterly conference call.
[O]ur strategy is continuously to improve every node in the performance, such as 28 nanometer. And are continuing to improve the 16 nanometers technology. And we have some very good progress, and you might call it the 12 nanometer because we’re improve in the density, classical density, performance and power consumption. Yes, we have that.
TSMC appears to be performing the same kind of optimization and improvement to existing nodes that every foundry performs over time. It’s just that this time, they’re calling it something a little different.
This photo shows TSMC’s various types of 28nm technology and their respective characteristics at a very high level. The deeper you drill into the characteristics of each sub-type, the larger the differences become. A device built on TSMC’s 28LP process would have very different characteristics than a device built on 28HP, even though both are based on 28nm.
Over the last 20 years, it’s become increasingly difficult to cleanly define a node. In the old days, node size and gate length were synonymous. As time has passed, the metrics for a new node have grown more complex. Samsung, TSMC, and Intel all tend to brand their products as belonging to the same node (save for TSMC going with 16nm as opposed to 14nm), but that doesn’t mean they are identical, as the chart below demonstrates:
Over the last few years, we’ve started to see more and more manufacturers taking this route. Samsung’s 14nm technology is based on a respin of its original 20nm process, not the 20nm process it publicly reported when it first discussed the technology. It’s debuted multiple variants of 14nm, while TSMC has both 16nm and 16FF+. Intel has respun 14nm into 14nm+, and some rumors suggest the firm will do a third respin (14nm++). TSMC likely isolated some additional improvements it could make to its 16nm node and decided to adopt a new naming convention.
Several years ago, I predicted we would begin to see more firms debuting their own methods of identifying process nodes as advances became more difficult and the speed of improvement fell. Samsung is positioning its 10nm node as a long-lived node, while TSMC believes its 10nm will be a short-lived process. Calling its new 16nm refresh “12nm” allows the company to appear to stay in near-lockstep with the rest of the industry, as opposed to falling behind it (even if that ‘falling behind’ is illusory and based in marketing nomenclature, not fact). Intel’s 10nm node will be smaller than Samsung’s, but Intel generally marches to the beat of its own drum. But we’re seeing a bit of jockeying for position and even different plans around specific nodes today that didn’t used to be the case.
TSMC’s 12nm process will undoubtedly deliver some improvements over its 16FF+ node, but it’s not going to be the same as a full-node shrink, and the improvements may range from small to modest. I’m not too worried about how this could impact Volta, though. Nvidia has already demonstrated that it can deliver a new GPU on the same process node with significant performance improvements, and Volta could turn out to be a major advance over Pascal, even if its 12nm node is more of a 16nm tweak.