At processor giant Intel's (INTC 2.11%) developer forum back in 2014, a chip manufacturing technology guru with the company, Mark Bohr, spent quite a lot of time talking about Intel's 14-nanometer chip manufacturing technology.
A good deal of the presentation was spent discussing the chip area and, ultimately, cost-per-transistor reductions that this new technology enabled. This is what Intel seems to like to talk about quite a lot these days, but in Bohr's presentation was also information around the improvements in performance-per-watt that the new 14-nanometer technology brought.
Per Bohr, for server, laptop, and mobile applications, the 14-nanometer process technology is expected to bring a roughly 1.6-times improvement in performance-per-watt over the prior generation.
Image credit: Intel.
Following the launch of the company's first mainstream 14-nanometer server products, it's worth examining whether the company's 14-nanometer technology actually enabled a performance-per-watt improvement of this magnitude.
22-core Broadwell-EP versus 18 core Haswell-EP
The most powerful server processor that Intel built on its 22-nanometer chip manufacturing technology featured 18 of the company's Haswell CPU cores. The chip ran at a base frequency of 2.3 gigahertz, with maximum turbo (likely on a single core) of 3.6 gigahertz. Intel rated the chip at a 145 watt thermal design power.
Late last month, Intel released the successor to Haswell-EP known as Broadwell-EP. This chip is built on the company's 14-nanometer manufacturing technology and uses slightly improved versions of the Haswell CPU core. Broadwell-EP also scales up to a greater number of CPU cores than Haswell-EP does.
According to Intel, the successor to the 18-core Xeon E5 2699 v3 mentioned above is a 22-core chip known as the Xeon E5 2699 v4. The base frequency of the chip sits at 2.2 gigahertz (slightly lower than the base frequency of the corresponding Haswell part), with maximum single-core turbo once again at 3.6GHz.
The Xeon E5 2699 v4, like its Haswell counterpart, is rated at a 145 watt thermal design power.
If we take Intel at its word that the Broadwell CPU cores deliver an average of 5.5% better performance-per-clock relative to the Haswell CPU core, then we can start to do some quick-and-dirty performance-per-watt calculations.
Let's assume that when all 18 cores are fully loaded on the 18-core 22-nanometer part, the CPU cores stay at 2.3 gigahertz; for the 22-core 14-nanometer part, let's assume 2.2 gigahertz. If we multiply 2.2 gigahertz by 1.055 (5.5% performance-per-clock improvement), then the 22-core Broadwell-EP is really "equivalent" to a 2.32 gigahertz Haswell-EP part.
The 22-core Broadwell-EP part has 22.2% more cores than the 18-core Haswell-EP part.
If we assume equivalent power consumption for the two chips, then the 22-core 14-nanometer Broadwell part should be along the lines of 23.3% faster than the 18-core 22-nanometer Haswell-EP chip.
There is a clear improvement in performance-per-watt, but it looks to be more on the order of 23.3% rather than the approximately 60% generation-per-generation improvement that Intel claims for each new chip manufacturing technology.
Where does the 1.6x improvement come from?
Although real-world power consumption numbers could paint Intel's 14-nanometer in a more favorable light than the "paper" comparison done here, it's likely that the "real world" performance-per-watt increase attributable to 14-nanometer in fairly high performance server processors is closer to ~1.23x rather than the quoted 1.6x.
What's also interesting is that in Intel's paper describing its 14-nanometer process technology, the company claims that saturated drive currents increase by 15% and 41% for NMOS and PMOS transistors, respectively, over the prior generation 22-nanometer technology at the same voltage/leakage.
Perhaps transistor performance, rather than transistor density, should be the focus of Intel's next investor-targeted presentation on process technology?
