A while ago, I wrote about Heatsinks Hitting a Wall. The gist of it was that we seem to be hitting air cooling’s performance limits – there’s only so much that this technology can achieve, and current designs may be about as good as it gets.
An additional factor facing heatsink designers is the diminishing size of CPU cores, due to decreasing line sizes and die shrinks, and the consequent rise in watts/mm² radiating off the CPU core.
CPU manufacturers make money by getting as many good CPU cores from a silicon wafer as possible. In addition, the more cores/wafer, the lower the cost/wafer. Ergo, their incentive is to shrink die size to get more cores/wafer. The enabling technology is line size – the smaller the line size, the smaller the CPU core, the more cores/wafer etc.
An AMD presentation (HERE) laid out these factors very nicely – I have extracted the essentials in the following table:
As the first three rows show, as line widths decrease, CPU core sizes decrease and cores/wafer increase. The net result is cheaper CPU cores and enhanced profitability.
The flip side of all this is what happens to watts/mm². Assuming power requirements stay the same for each generation (more on this later), say 50 watts, smaller cores dissipate more heat per area each generation than the last.
The “so what” of this is that it becomes more difficult to move the same amount of heat from an increasingly smaller surface area. The cooling challenge is to find the low cost solution to quickly spread heat from the core to a larger surface area (for a theoretical discussion, go HERE.).
A presentation I found at Princeton University (HERE) succinctly summarized the relationship between “Moore’s Law and Power Dissipation”:
“To get the performance improvements we’re accustomed to, CPU Power consumption will increase exponentially”
This is captured in a generalized formula for Dynamic CMOS Power Dissipation:
C = Capacitance (decreasing as line width shrinks)
V = Voltage (increasing as frequencies increase)
A = Activity (relatively constant)
F = Frequency (increasing)
Looking at AMD and Intel trends gives a good picture of how this balancing act is playing out:
|Intel Die Size|
|AMD Die Size|
|AMD 50 watts/mm²|
|Intel 50 watts/mm²|
Even assuming voltages “toe the line” and remain constant, watts/mm² inevitably increases. If we look at current CPU cores at the same frequency from AMD and Intel, we see the increase in watts/mm² even though CPU voltages have decreased.
Comparing CPUs at the same frequencies clearly shows the increase in watts/mm²:
0.13 Line Size
0.09 Line Size
0.07? Line Size
|Intel Watts/mm² @ 1.6 MHz|
|AMD Watts/mm² @ 1.46 MHz|
Taking some liberties, I projected out the next generation CPU Watts based on probable watts/mm², assuming a doubling of frequencies to around 3 MHz:
AMD: 0.7 x 64 x 2 = 89.6 watts
Intel: 0.4 x 98 x 2 intel = 78.4
I think it’s reasonable to expect CPU voltages in the 70-80 watt range as the “norm”. More important, the cooling challenge as expressed in watts/mm² is increasing, and the probability of more exotic approaches (peltiers, water, etc.) is increasing. Expect to see more CPU cores with heat spreaders (eg., P4s) as one approach.
Also, don’t discount software approaches such as Intel’s “Thermal Throttling”. Trading off CPU cycles for cooling is seductively easy and, for the average user, probably transparent on screen. How this squares with truth-in-advertising is an open question.
I would not be surprised to see a CPU heat spreader using an exotic material or technology (heat pipe) as standard.
For overclockers, more exotic solutions will be required to keep CPU heat under control. Should be an interesting challenge all around.
As always, opinions and comments from those more knowledgeable than I on this issue are welcome.