AMD presents Naples server processors
AMD has released details about its next-generation server processors with Zen architecture. The processors are codenamed Naples and have a maximum of thirty-two physical cores and sixteen memory slots per processor.
Naples uses AMD’s new Zen architecture, which combines frugal cores with good performance: after years, the IPC of AMD’s architecture is again competitive with that of rival Intel. AMD pits its top model from the Naples series against Intel’s Xeon E5-2699A V4, a server processor for two-socket systems. That Xeon costs about five thousand euros, but the company does not want to mention a price for the AMD variant. It also doesn’t give details about power consumption, but does say that the AMD platform would be faster thanks to more cores and more addressable memory.
The thirty-two cores of the Naples flagship model are combined with a total of eight memory channels per processor, each good to drive two memory modules. This means that the maximum addressable memory of a system with two Naples processors is 512GB when using 16GB modules. In theory, even 4TB of memory can be addressed. The memory speed is 2400MT/s, where competitor Intel has four memory channels per processor with speeds of 1866MT/s, good for 384GB of addressable memory in a two-socket system.
In addition, the Naples processors have a large number of PCI Express lanes. 128 PCIe 3.0 lanes are available per processor, of which 128 lanes are also available in a two-socket system: the other lanes are used for the Infinity Fabric interconnect. That way graphics cards for gpgpu calculations or other accelerator cards can be connected directly to the processor, without slowing down plx chips to split pcie lanes. The first Naples processors should be available in the second quarter of this year, but general availability is not expected until the second half of 2017.
AMD demonstrated some benchmarks that pitted the 32core Naples processor against Intel’s E5-2699A V4, with different memory and core configurations. With equally clocked memory and 44 cores each active, the AMD system completed a benchmark to analyze seismic data with a sample grid of 1 billion samples and ten iterations almost twice as fast as the Intel system. With all AMD cores active and the memory at the maximum supported speed, the performance difference further increased to 2.5 times and when the dataset was expanded to 10 billion samples, the Intel system ran out of memory capacity to run the benchmark. Of course, the benchmarks have been carefully selected for the demonstration, but since Naples is based on the same Zen cores as Ryzen, AMD seems to have a good chance to play a full part in the lucrative server market again.