AMD Zen and the Art of Microprocessor Maintenance

AMD is a fantastic company with highly talented people, but for some reason just hasn’t managed to put a winning streak of microprocessor architectures back-to-back. It’s frustrating to watch: they ride like mad to catch up to or even pull slightly ahead of Intel, then fall back in the pack when they have to make an extended pit stop, then ride like mad again to close the gap.

My first experience with this was the K5. In the same Socket 7 footprint, albeit about 2 years behind, AMD had a part that delivered slightly faster integer processing than an Intel Pentium and offered an extended lifecycle for embedded designs, something Intel was still struggling with philosophically at the time. The K6 arrived with a reengineered microarchitecture from the NexGen acquisition, and kept pace for three enhancements.

Until Intel got its tick-tock engine running. In another surge attempt, AMD introduced the K7, more well known as Athlon, on a copper-based interconnect process jointly developed with Motorola. Intel jammed its clock speeds on Pentium 4 from 2 GHz to 3.2 GHz, pushed SSE2 support hard, and added a dual-channel DRAM controller that blew the doors off AMD.

Hector Ruiz stepped in and took AMD in the direction of multicore – Barcelona. Battered by a stream of introduction delays, what would have put AMD in the lead when tipped in 2003 wasn’t fully ready until 2008, giving Intel plenty of time to get multicore religion. By 2009, “asset light” kicked in and AMD spun off its manufacturing operations into GLOBALFOUNDRIES.

I don’t have space for a comprehensive short-history-of-AMD post, and I don’t mean to sell their positive developments short. Through a lot of manufacturing and management turmoil, AMD did manage to carve out a nice position in processors for laptops and embedded use thanks to Intel’s misguided interpretation of “low power”. Much of this was based on stronger integration, such as the AMD Fusion family of APUs with CPU, northbridge, and GPU on the same part.

Point being, AMD is still trying to catch Intel in the mainstream. There was a family of cats – Bobcat, Jaguar, Puma – and a family of construction equipment – Bulldozer, Piledriver, Steamroller, Excavator. It’s kind of like watching that new and improved dishwashing liquid commercial; if the new stuff is so good, it implies the old stuff was pretty bad. It all depends what you are comparing to, but when you’re stuck at 28nm…

That leads us to this chart, shown at Hot Chips 2016. “Zen” again is a promise from AMD that this time, really, we’re gonna compete head-on with Intel. 40% more instructions per clock with reduced energy-per-cycle in GF 14nm FinFET. Ground-up architectural redesign – the complete Mike Clark presentation is over on Anandtech.

AMD has been riding really, really hard to get to Zen. In any other context, we’d be celebrating this level of effort and accomplishment. Much as Robert Pirsig was rejected dozens of times before publishing his classic, AMD has been spurned at nearly every turn in the road trying to get up and stay up alongside Intel. It’s very difficult to compete with the undisputed leader who has every possible advantage of time, money, and ecosystem support, especially when Intel keeps redrawing the map where the race has to go. AMD has no choice but to follow.

The AMD Zen story is here:

AMD Demonstrates Breakthrough Performance of Next-Generation “Zen” Processor Core

I think AMD just might pull even briefly with Zen at 14nm if delivered in mid 2017 as promised. The question is can they stay there, with Intel already eyeing 7nm? Or will this be a replay of past races where Intel just keeps on riding and AMD falls behind again?