It is curious that if we take a list of specs from Intel and AMD in their desktop and server processors for the same architecture and see the differences, we will realize two quite interesting factors: There is a big gap. in frequency and consumption. Why does this happen if the efficiency and knots are equal?
The game-changing factors: server vs desktop, frequencies and cores
There are many factors that make this the case, so we will try to break them down into several key points and comment on them, because as we well know these are totally different product lines and although it does not seem being the case, Intel and AMD follow different and very curious paths towards this approach.
Let’s start from the bottom, since EPYC vs Ryzen and Core vs Xeon are two independent fights that are designed from the second zero to approach the markets from the wafers. And here’s the first palpable difference for us: Intel has different fabs for Core and Xeon processors, while AMD orders everything from TSMC and uses the same wafers for EPYC and Ryzen.
Why this representative? Well, because we know it, the server lines use the best chips to have the least amount of current leakage. And here we relate to the second representative factor: energy efficiency and its relation to thermal density.
Again, the question would be why? Well, because of the type of use and the loads that both will carry. Server processors are standardized with sustained loads of 100% 24/7 utilization, which means the processor will consume all of its TDP all the time. Therefore, a more efficient ASIC in terms of current leakage is a great improvement which will save electricity, voltage, consumption and hence temperature.
The ratio of efficiency, frequency and nuclei
With this in mind, it makes sense to think that the number of cores should be the maximum that allows both the architecture, chip, wafers and platform, because we strive for the highest performance per package. But then the frequencies will suffer from the scalability of voltage, efficiency and power.
And is that the cores as such are not only to blame for the increase in watt-to-core ratio, but also counting on the number of buses and caches that include this class of processors. Buses are a big source of power consumption, but they are necessary to increase for example the number of RAM channels or just some like PCIe or USB.
Finally, we take into account the stability factor. A server processor must be completely stable over time, because it surely will not rest until its mission ends several years later, in some cases even a decade. Therefore, a server processor will always strive for the best ratio of frequencies / core / consumption / temperature / stabilityTherefore, the best way to do this is to lower frequencies and voltage without losing sight of as many cores as possible for the platform they are intended for.