Graphics processing units (GPUs) are among the largest chips produced, and in just over a decade, the most powerful graphics processors have gone from billions of transistors to over 100. But according to TSMC, even these numbers pale in comparison to what lies ahead. Billions." ..... How about .1 trillion transistors?

Taiwan Semiconductor Manufacturing Corporation (TSMC) is the world's largest graphics processing unit manufacturer; AMD, Intel, Nvidia, Qualcomm, Tesla, and others are developing bigger and more powerful chips. Like Nvidia's recent Blackwell B200, we are already approaching the 100 billion transistor mark, and future GPUs will need to be much larger than this to continue to improve performance and capability.

However, certain elements (such as SRAM bit cells for caches) are not much smaller than they were five years ago, making GPUs increasingly difficult to manufacture. While logic circuits can still be miniaturized, constraints in other areas will only increase the overall die size of the chip in the end. Unfortunately, the maximum size of a single die is also quite marginal, with the maximum achievable area being just over 800 square millimeters.

This is known as the reticle limit, and unless costs skyrocket, there is no sign that this will improve significantly. But we are already seeing solutions to all of these, and in an IEEE Spectrum article via The Register, Mark Liu (TSMC Chairman) and Prof. H.-S. Philip Wong (TSMC's Chief Scientist), explained that by going the chiplet/tile and 3D stacking route, a 100 billion transistor GPU will soon look like nothing.

"As the trend of increasing transistor counts continues, multiple chips interconnected with 2.5D or 3D integration will be needed to perform the computation; by integrating multiple chips through CoWoS or SoIC and related advanced packaging techniques, the total transistor per system number of transistors is much larger than can be squeezed into a single chip. We predict that within a decade, the number of transistors in a multi-chip GPU will exceed one trillion. [Using the aforementioned Blackwell B200 as an example, that GPU may look like a huge chip with 208 billion transistors, but it is actually two 104B dies on a silicon interposer. AMD's MI300X 153B superchip is similar, with four compute dies all nestled together. However, these GPUs boast huge transistor counts by spreading outward, and the chips need to be stacked on top of each other to make the leap to a trillion transistors.

AMD has succeeded in significantly increasing the amount of L3 cache in the Ryzen 7 7800X3D and other CPUs because by designing the base chiplet in this way, it already has the connections needed to fit another slice of silicon on top of it. Of course, stacking chips creates new problems, such as how to keep heat from building up in the base layer, but TSMC sees significant room for growth in this area: "We see no reason why interconnect density should not grow by single digits or even more."

While the transistors themselves may not get much smaller in the next few years, the size and complexity of the chips themselves will clearly continue to increase; it is not hard to imagine what kind of processors the graphics cards of 2034 will have, but by comparison, the GPUs of today will look basic will no doubt look basic in comparison.