Slashdot Mirror
AMD Offers Full Details and Performance of Zen-Based Naples Server Platform (hothardware.com)
MojoKid writes: AMD lifted the veil this morning on architecture details and performance expectations of its next generation Zen-based server platform, codenamed Naples. Naples is an up to 32-core, 64-thread variant of Zen, targeted at enterprise and data center applications. The processors will feature eight-channel DDR4 memory controllers (with up to 16 DIMMs attached per CPU), with support for up to 4TB of memory and 128 lanes of on-chip PCI Express connectivity. In a 2P (dual processor/dual socket) configuration, Naples offers up to 64 physical cores (128 threads), access to 32 DIMM slots, and aggregate 16 memory channels. Versus a 2P Intel Xeon E5-2699A V4 based server, the 2P Naples setup ends up with double the memory channels, a higher total memory capacity, more cores (20 more physical cores, 40 more threads), and 48 more available PCI Express lanes. AMD's performance comparisons at its tech day event pit a 2P Naples server with 512GB of DDR4 RAM up against a 2P Intel Xeon E4-2699A V4 configuration with 384GB of RAM. The Naples system had a higher memory capacity and that memory was clocked much higher too -- 2400MHz versus 1866MHz. The Naples system has more cores, and with SMT on, can ultimately process more threads as a result. The AMD Naples system also has double the memory channels, further improving peak memory bandwidth. In its demos, AMD used a seismic analysis workload, which involved multiple iterations of 3D wave equations. According to AMD, the test taxes the entire system, including CPU cores, memory and I/O. In this demo, the AMD server system completed equations roughly 2.5x faster than the dual-socket Intel Xeon server. Expected price points weren't given, but Naples processors and servers should be available in Q2 this year.
Fun fact: Quad sockets are SUPER-RARE.
Go ahead. Try and Google for a quad motherboard. They haven't made them since like... the Athlon 64-era Opterons. The boards are MASSIVE. You can't fit 128 DRAM slots on a single board. (And who wants to BUY that many?) There's too many traces and the board becomes super expensive from the extra routing layers. The board is also going to flex under any kind of weight, and the larger the footprint the more important the mounting becomes (going from casual "bolt it to the wall." to only professionals can service it carefully lest they flex it and snap a delicate trace.)
Some of the larger boards (ala more sockets) are connected into sub-modules that stack perpendicularly into a specialized transport bus. The bus / message-passing chips are also pretty damn expensive.
The trend has clearly been moving away from MULTI-CPU, and toward MULTI-CORE. The same motherboard, you buy better and better CPUs with more cores as you need. All the cores are together. There's no motherboard manufacturer based chipset. No huge message piping. It's all there in the L1, L2, L3, (L4). And it's the SAME or MORE CPU cores than before. 4 sockets with 4 cores vs 1 CPU with 16 cores? That's a no-brainer. It's way more cost-effective for 99.99% of people. And you can upgrade ONE (or two) CPUs instead of four which adds up fast across many stations.