Good news for AMD CTO Mark Papermaster: Moore's Law is not dead; CPUs and GPUs will continue to improve for the foreseeable future. But he also has some bad news. It is becoming increasingly expensive to keep everything on track, forcing innovative solutions like chiplet design.

Moore's law is, of course, the observation that the transistor density of integrated circuits doubles every two years; originally proposed by Intel co-founder Gordon Moore in 1965, the law initially stated that integrated circuit density would double every year, but in 1975 Moore In 1975, Moore revised the law to double every two years. Moore revised this schedule to every two years in 1975, but it has remained in place ever since, and to this day has proven to be highly prescient.

At a summit in Las Vegas (open in new tab), Paper Master explained that Moore's Law is still on track, but chip technology is becoming increasingly complex. [Moore's Law is slowing down. Moore's law is dead," Papermaster said. But according to AMD's technical guru, that is not actually the case.

"That's not to say that exciting new transistor technologies won't emerge. In fact, I can see exciting new transistor technologies emerging in the next six to eight years.

The difference today, Paper Master explains, is that while transistor density used to double in the same year, the cost of the same size chip remained almost the same, but the cost per area of silicon has increased with each additional production node. Computer chips of a given size have become much more expensive. [This was a key driver for AMD to move CPUs to chiplet designs a few years ago and again for GPUs with RDNA 3 earlier this year, he said. If chiplets are part of the solution to the rising cost of silicon manufacturing, combining older CPUs and GPUs with specialized accelerators will become increasingly important. [It's GPU acceleration, specialized function units, and adaptive compute like we've acquired with Xilinx GPU acceleration, specialized function units, and Adaptive compute," he said. [In other words, while transistor density continues to rise according to Moore's Law, chips are becoming more expensive, forcing companies like AMD to use chiplets to improve yields and thus stem the cost spike. Specialized circuits also tend to be much more compact than general-purpose CPU and GPU blocks, allowing for increased performance without the price penalty of traditional pure CPU and GPU designs.

In recent years, it has been widely reported that the end of Moore's Law is imminent. This is partly due to the fact that Intel, once the king of chip manufacturing technology, has been struggling. But Intel is not the only chipmaker. While certainly lagging behind, at the cutting edge, TSMC is fully on track with Moore's Law.

TSMC of course manufactures AMD's CPUs and GPUs, not to mention Nvidia's latest GPUs and Apple's M1 chips. TSMC's 5nm node, currently found in AMD's CPUs and GPUs, is approximately one generation ahead of Intel's 10nm node, recently renamed Intel 7.

TSMC has also just started 3nm production, and devices using TSMC's 3nm silicon are expected to hit the market in the first half of 2023, most likely in the form of Apple's Mac computers with 3nm derivatives of its M1 and M2 chips TSMC then expects 2nm production to begin in 2025. For now, Moore's Law looks sound enough. It just needs a little maintenance.