AMD is supposedly planning a 32-core CPU with an eight-channel DDR4 interface

Rumors have continued to swirl around AMD’s upcoming Zen processor and its performance capabilities, but a new slide from a CERN presentation is going to dump more fuel on the speculation fire. During a presentation on technology and market trends for the data center, one of the CERN presenters showed the following slide:

Most of this is information we already knew or have gleaned from software patch notes and other updates, but the claim that AMD was planning a 32-core with a whopping eight-channel DDR4 memory interface is sure to raise some eyebrows. That’s twice what Intel currently offers in high-end server architectures — not to mention 2x the number of cores.

The other way to think about this, on the other hand, is that if AMD wants to offer a 32-core CPU, it better be able to keep those cores fed from main memory. From this perspective, eight channels of DDR4 isn’t odd — it’s actually what you’d need to keep the CPU busy.

So far, the tidbits that have leaked out about Zen sound pretty reasonable. The new architecture is rumored to have a 64KB L1 cache (32d, 32i) and a 512KB L2 cache per core. The L3 cache is rumored to be 8MB and split between each group of four cores.

The high-level details look fine, but then, the high-level details and early reports on Bulldozer looked fine, too. A 32-core CPU design sounds like an amazing, Intel-destroying piece of silicon… until you recall that Bulldozer’s original plan was to compensate for low IPC by hurling insane amounts of cores at Intel.

Adding more cores may make more sense at 14/16nm, given that we know that these nodes are frequency and power-constrained, but Intel’s maximum clock speeds at 14nm drop off sharply as the CPU core count rises.

Intel’s Xeon E5v3 family rates a top-end quad-core with a base clock of 3.7GHz as a 140W TDP CPU (max clock is 3.8GHz). By 18 cores, the base clock has fallen to 2.3GHz. Still, that’s a 4.5x increase in parallelism for a 60% decrease in frequency — a net win, assuming your workload is parallelized enough to benefit.

Even if we assume that AMD has a superior performance curve thanks to the benefits of 14nm technology, a 32-core chip is going to be a low-clock endeavor. And given how the industry has shifted to per-core pricing for many applications and ecosystems, it may be hard to sell businesses on a giant CPU core. AMD always competed well on price when the Opteron family could keep up with its Xeon counterpart, but a 32-core CPU could never be cheap; Intel’s 18-core Xeons are 662mm sq, or roughly the size of a high-end GPU on 22nm. Since both GlobalFoundries and Samsung use a hybrid 14nm process with a 20nm BEOL, AMD’s 14nm CPUs won’t be as small as you might think. Even if the company uses dual 16-core chips and joins them on one package, the final product is going to be enormous.