Marvell this week teased its forthcoming server-grade ThunderX3 processor, which can, we're told, sport up to 96 Arm CPU cores and 384 hardware threads per socket.
The component family isn't out yet, no pricing guides are publicly known, and some specs have been withheld. It is slated to officially launch in 2020, though.
El Reg reckons Marvell was ruffled by Ampere's Arm-N1-based 80-core Altra announcement, and just wanted to remind people it's still gunning for data center and supercomputer workloads. The timing could be a coincidence, of course.
We're told the ThunderX3, codenamed Triton, will use the homegrown 64-bit Arm-compatible CPU cores designed by Broadcom, and ultimately acquired by Marvell via Cavium, that support Armv8.3 instructions and some in 8.4 and 8.5.
The TSMC-fabbed 7nm Triton will include up to 96 of these CPU cores, each supporting four hardware threads, plus four 128-bit SIMD engines, eight DDR4-3200 memory controllers, and 64 lanes of PCIe 4.0, per socket. It can be used in single or dual-socket servers.
The processor family is aimed at supercomputing, cloud, and network edge applications.
You can get a full dive into the Triton chip over at our sister site, The Next Platform.
The ThunderX3 follows the ThunderX2, which has appeared in Microsoft Azure cloud and various supercomputers. ®
Cooper Lake is interesting because it includes, among other things, support in hardware for bfloat16, a 16-bit floating-point data format useful for artificial intelligence. We understand Facebook had early access to Cooper Lake chips to run machine-learning code on its servers.
It appears Intel still plans to ship four- and eight-socket 14nm Cooper-Lake-based Cedar Island Xeons this year, and one- and two-socket 10nm Ice Lake Whitley Xeons and four- and eight-socket 10nm Ice Lake Cedar Island Xeons in the second half of 2020.
So essentially, there won't be any single or dual-socket 14nm Cooper-Lake-based Whitley Xeons, as previously promised, but there will be 10nm Ice-Lake-based Whitley variants later this year, hopefully. This is due to Intel focusing on various things, such as actually shipping long-delayed 10nm processors to customers.