Exponential Efficiency Gains: Q&A with AMD’s Sam Naffziger

Energy efficiency gains are built right into Advanced Micro Devices’ microprocessors, known as the “brains” of computing devices. In 2014, AMD announced the 25×20 Energy Efficiency Initiative targeting a 25-fold improvement in the energy efficiency of its accelerated processing units by 2020. If the company built cars, that would be like turning a 100-horsepower, 30-MPG vehicle into a 500-horsepower one that gets 150 MPG within six short years.

“It’s been a good rallying point for the team,” says AMD corporate fellow Sam Naffziger, a top engineer who leads the company’s power management and power reduction efforts. “We’re making strong progress.” The 25×20 Initiative won an Environmental Leader Award in 2016. One of the judges called it an ambitious, cross-organizational approach to problem solving.

Recently we caught up with Naffziger to learn about his approach to balancing product performance with energy savings, and where AMD is in relation to their goal.

Where are you now relative to the 25×20 goal?

We typically introduce our products on an annual cadence in this market. We started out with a baseline of 1 in 2014. We were between 2 and 3 with the 2015 product, and made some more progress in 2016. This year, we’ve made great strides in our CPU performance and power efficiency.

We’re at a 5.9x improvement over 2014. Moore’s Law — a doubling in density and corollaries in performance every two years — is an exponential goal. Our goal is also an exponential one. At this point we’re about halfway to 2020 and the 5.9x is above where we actually need to be.

What was the predicted mark for this point in time?

It’s around 4.5. Next year, if we have a similar improvement, that 4.5 would be 8.5. Our commitment is to keep that kind of compounding improvement year over year.

How are you working to reach the goal, and does it get harder each year?

We’ve got a lot of technologies we’re excited about that cause the computer to rapidly reduce power to adapt its operation to the workload that happens to be running — and burn just enough power to get that task done. The challenge is that we’ve got to both increase performance and reduce power.

Yes, it gets harder, but I’m confident that with the tricks in our bag, we’ve got a path to getting there. It’s a combination of our engineering and innovations in the silicon, taking advantage of what remains of Moore’s Law and the processor architecture, and of course the software.

Are there innovations that were helpful in increasing the energy efficiency?

We have a very complex design — billions of transistors, sophisticated handling for display, graphics, and computing — and we’ve partitioned it into about a dozen different voltage domains. We have a dedicated processor that tracks all the activity and the workload, and it has tentacles everywhere to control these voltage domains to set them to just the right level.

Perhaps more important, for energy efficiency, it shuts them off very quickly. That’s done in a matter of microseconds. Between keystrokes, we can often shut silicon off. And then when a keystroke comes, we fire everything up, do whatever task it needs to do and then go back to sleep to save power.

What are the biggest challenges you face in improving the energy efficiency?

We wrapped both the performance dimension and the power consumption into the goal. It’s easy to deliver a low-power processor, something that does minimal computation and hardly burns any power. It’s also easier to deliver a very high-performance computer that’s running at max frequency all the time. We’re targeting something that’s both low-power and high performance.

The challenges we have are entering these low-power modes and keeping all the pieces of the chip at just the right operating level so they save power without hurting performance. That means we’ve got to, in real time, track and recognize what kind of program is running, and then configure our processor to use just enough power to make that perform well.

As you can imagine, there’s constant learning because those software programs are changing and evolving, too. We have adapted algorithms that track these behaviors and reconfigure on the fly. It keeps a lot of very smart engineers very busy.

How do you make sure those goals get balanced within your team?

The time challenge is we can’t do everything we’d like to do, and we have to rank order. These processors are super sophisticated and complicated. We can’t get everything squeezed out of the silicon that we know it’s capable of with the first iteration because we don’t have very many weeks from when we get the silicon to when we actually have to start shipping it to customers.

We prioritize based on the best energy efficiency at the time, and then the features that we don’t have the time to do, we leave them for the next version of the product next year. That’s actually worked well. We then get solid improvements year over year, and we can still deliver on time.

How do you and your team stay focused on a goal when it gets harder?

Innovation is borne out of having no other options. If we just did it the same way we did it last time, that’s a lot easier and lower stress, but we wouldn’t make the necessary gains. Having that aggressive goal means we’ve got to develop new approaches. We’ve got the right pipeline, and we have motivated and innovative engineers.

It’s a science-based target — I didn’t pull the number out of the air. I understand the technology well and had a reasonable idea of some of the things we’d be able to do. So I made a roadmap by product generation and the improvements we’d need to make at each of these dimensions. Then we stuck with that.

Is there a key lesson you’ve learned along the way?

You can’t just put out an aggressive goal that sounds good and say OK let’s find a way to make this happen. You’ve got to have a foundation in reality. The good news for us is that 25x is an impressive number, but it’s also based in what we can achieve.

What’s next for AMD?

We have a lot of work to do, but we’re looking at the rest of our product portfolio and a lot of those consume energy, too. We have graphics cards, gaming PCs, and server product lines. We’re looking at how to establish aggressive goals there in a similar manner to drive energy efficiency. We’re not going to stop at the 25x.

We’re now accepting submissions for the 2018 Environmental Leader Product and Project Awards. Early birds receive an entry fee discount when they submit by November 10. Learn more here.