Gridless Clock Distribution

As you push the frequency limits in large, high performance CPUs like the Core 2, you need to minimize overhead in each clock period so you can do the maximum amount of work. Timing is key.

All functional blocks within a microprocessor operate on the same clock signal. You can think of the clock signal like a metronome or a heart beat; work only happens at every tick of the metronome, every beat of the heart or whenever the clock is logically "1" or "high".

There's a single clock signal that's distributed across the chip, but the more complex the chip and the more times the clock changes per second, the tougher it becomes to just make sure the clock gets to all parts of the chip.

In order to reach the absolute maximum frequencies possible, making sure the clock signal arrives at all parts of the chip at the same time with very little variation becomes increasingly important.

The most effective method of distributing this single clock across the chip is by using a metal grid to distribute this critical signal across a large, complex die. The problem with this metal grid is that although it makes sure the clock gets to all parts of the chip with very little skew, it has a huge capacitance and consumes tremendous amounts of power. Large CPU cores can spend 30 - 35% of their power budget simply on clock distribution. This would obviously not work for Atom.

While the Atom processor needs to run at relatively high frequencies, we're not talking about pushing 4GHz here. Without strenuous frequency requirements, a simpler form of clock distribution is ok. Atom uses a binary tree for clock distribution; the clock signal is sent from the PLL to the point of divergence, then fanned out into multiple trees of inverters and finally distributed vertically from each point of divergence to each FUB.

The binary tree approach significantly reduces the amount of current consumed to the point that less than 10% of Atom's power budget is spent on distributing the clock.

Building by FUBs Atom's FSB and Multi-Core Aspirations
Comments Locked

46 Comments

View All Comments

Log in

Don't have an account? Sign up now