It’s Not How Many Transistors, But How You Use Them
Putting this all together, it is clear that transistor count and transistor density are highly problematic metrics. Both are strongly influenced by the overall design and the ratio of critical blocks like computational logic, SRAM, and I/O. Of the three, SRAM is by far the densest, so a modest change in the size of a cache could produce big changes in transistor count with a minimal impact on performance and value. Moreover, not all layout transistors are equal. Active transistors are the fundamental building block for valuable components like CPUs and GPUs. On the other hand, decap and dummy transistors are more akin to overhead. Hopefully most companies are unlikely to conflate active and layout transistors, but it is important to distinguish the two when comparing designs.
Despite the problems with transistor count, it is potentially useful in a very coarse-grained fashion. It’s almost certainly true that a processor with 100-billion transistors is more complex and more valuable than one with 100-million. The analysis probably still holds for a 2X difference in transistor count – especially for something that is targeting explicitly parallel workloads like a GPU, or for two very similar processors (e.g., two smartphone SoCs or two server processors). But realistically, it’s hard to believe that a modest difference in transistor count necessarily translates into meaningful difference in value. In fact, AMD’s Radeon VII and RX 5700 provide a great counterpoint. The Radeon VII packs in 28% more transistors, but delivers fairly similar performance in part because the RX 5700 line uses a more advanced architecture. Moreover, the RX 5700 line is far less expensive since it uses GDDR6 instead of HBM2. When it comes to actual value to customers, it’s not about the transistor count, but how the transistors are used. Modest differences in transistor count just don’t matter compared to good architecture, feature selection, and other factors.
Many of these same criticisms hold for transistor density and process technology. If a modest increase in transistor count doesn’t meaningfully impact customer value, then it seems unlikely that a corresponding modest increase in density would tip the scales either. On the other hand, factors such as transistor performance, dynamic power, idle power, toolchain and third party IP support, wafer availability, and advanced features such as RF or optical devices can bring significant value to the table for the right products. Density is just one of many aspects of a process, and focusing on it too much risks missing the forest for the trees.