Cavium has used Computex to push out its next round of ARM server system on chips (SoCs), the ThunderX2.
The ARMv8.2 architecture, ARM's Server Base System Architecture chip, uses a 14nm process instead of its predecessor's 28 nm. The SoCs support as many as 54 cores per socket, which Cavium reckons will as much as triple the performance of the first-generation ThunderX.
Each core runs between 2.4 and 2.8 GHz in normal mode (3 GHz in turbo mode), with out-of-order execution support.
Cache features include 40K I-cache and a highly associative 64K D-cache.
The company says its interconnect offers more than 2.5 times the coherent bandwidth of ThunderX, with both single and dual-socket configurations supported.
Each socket can connect six DDR4 memory controllers and 12 DIMMs, with a dual socket configuration able to support up to 3 TB of memory.
The chip integrates Ethernet at 10, 25, 40, 50 and 100 Gbps Ethernet, and I/O also includes multiple SATAv3 and PCIe Gen3 (up to x16) interfaces.
You can peruse the full list of features in Cavium's media release.
As with ThunderX, the ThunderX2 family will have versions optimised for compute workloads (ThunderX2_CP); storage, big data, massively parallel processing databases and data warehousing (ThunderX2_ST); Web front-end and cloud RAN workloads (ThunderX2_SC); and media servers, embedded applications and NFV workloads (ThunderX2_NT).
How far this takes ARM towards being a genuine data centre competitor will become apparent once the chips start landing in designs. The Register's HPC sister publication, The Next Platform, has a detailed look at Intel's private benchmarking of a ThunderX-generation server here.
That generation still struggled against Intel's best, the benchmarks suggest.
"What everyone wants is for Cavium and the rest of the ARM collective to put their own best foot forward tests into the field on their current chips and the impending ones as they come to market," The Next Platform's Timothy Prickett Morgan notes, adding that "2017 could be shaping up to be a real battleground for compute." ®