AMD isn’t just making a comeback in the consumer CPU space, but also using the new Zen core to help power their re-entrance into the enterprise market, something that will be vital for their business going forward. Zen, with Ryzen, is an overall spectacular architecture that improves greatly upon the construction equipment cores and provides some very nice competition for Intel with Naples.
Naples could be a force to reckon with
What makes Zen particularly exciting in the enterprise and data-center space is how well it scales with the number of threads and in how some workloads may benefit from some of the architectural changes that have been made, particularly cryptographic workloads. Naples has the potential to be a very powerful competitor, and likely for less money as well.
The new server CPUs, which as of yet don’t have an official name (though Opteron may be the obvious choice), are poised to enter the market in the second quarter of 2017 with up to 32 physical cores that have 64 PCIe 3.0 lanes and octa-channel DDR4 capability. AMD is building in a lot of capability with the Naples platform, hoping to entice those that need to accelerate their workloads with massive memory capacity, up to 4TB per 2P server with nearly 170GB/s, and with plenty of PCIe lanes to allow for GPUs and other accelerator cards to be used en masse. The future of computing is a heterogeneous approach that makes good, intelligent use of all resources, especially highly capable GPUs such as AMD’s own forthcoming Instinct line.
So what is Naples? It’s a collection of Zen CCX’s on one die that uses the Infinity Fabric to communicate. Though it seems that consumer Ryzen doesn’t use that fabric to communicate cache data between the complexes, here it may be different. What’s more, the Infinity Fabric should be around the speed of a typical PCIe 3.0 architecture, or higher. This allows the two separate CPUs to communicate everything and remain completely coherent, so that if a GPU attached to the PCIe lanes of one CPU needs to transfer data to the GPU of the other CPU, it can do that without a performance loss. Though we’re unsure just how the Infinity Fabric itself is setup, what topology, actual speeds and precisely how it communicates. But that aside, the limitations of the SMT approach we may have seen in limited quantity may not be much of an issue here. In fact, deep learning and other multi-threaded computational applications have seen great performance on Ryzen thus far.
How does Naples perform?
AMD was nice enough to include a few benchmarks, even if they’re slightly generic and we can’t quite confirm them as of yet, to illustrate their performance advantage. What we do know already is that deep learning, database and cryptography workloads do quite well on consumer Ryzen, and can only improve with increased parallelization. With the development of a compatible BLAS math library, such as their own BLIS library (which hasn’t been updated for Zen just yet), the numbers that you’ll see in Ryzen Part 2 will only get better. And Naples will become even more competitive. Our forthcoming numbers utilized Intel’s own Python distribution and their MKL library to ensure it performed as well as possible. You’ll see soon that the advantage isn’t much.
Of course we can’t wait to actually get to test their claims and see just how well the new CPUs and platform will perform given professional workloads. Already we’ve seen just how well Zen can do in the early stages, so it stands to reason that at the right price, Naples, too, will be a force to be reckoned with.