Most of Intel’s recent troubles can be attributed to its 10nm process node. Originally slated for a 2016 launch, it was delayed all the way to late 2019/early 2020. Poor yields, low operating clocks, and a slower ramp-up resulted in repeated delays and even discontinuation of the 10nm (Canon Lake) processors. The first server product line leveraging the process was launched in early 2021 in the form of Ice Lake-SP. While the 3rd Gen Xeon-SPs provide substantial gain over the preceding 14nm Skylake-based offerings (Cascade Lake-SP), the efficiency has deteriorated by quite a bit:
Running the SPECpower benchmark shows a notable regression in power efficiency upon going from Cascade Lake-SP to Ice Lake-SP. On average, the former is 60-80% more power-efficient than its more advanced 10nm successor. The efficiency gap between the two architectures decreases as you cross the 5 million IOP/sec mark, completely disappearing midway between 5 million and 6 million. This is due to the 14nm parts reaching their upper limit, losing a fair bit of efficiency at the cost of a relatively minor performance gain.
In a similar fashion, the 10nm Ice Lake-SPs see a rapid increase in the power draw north of 5 million IOPs per sec, growing by close to 100W with every million IOPs/sec gain. In comparison, going from 1 million to 4 million IOPs/sec saw an increase of just 100W on both systems.
While AMD’s Milan processors are also a tad bit less efficient than their predecessors, the two are based on the same process node (TSMC’s N7), and the delta is hardly appreciable. It’ll be interesting to see how the power efficiency improves as AMD shifts to TSMC’s 5nm node with Genoa, while Intel continues to use its 10nm ESF node for Sapphire Rapids-SP.