Marvell got into the Arm server business with the 2018 acquisition of Cavium, maker of the ThunderX Arm server processor. Now the company is introducing the first major revision to the product line since the acquisition, and it’s a doozy.
The ThunderX3 line tops out at a stunning 96 cores with four threads per core, for a total of 384 threads per processor. Intel and AMD have only two threads per core, and the top-end Intel Xeon maxes out at 56 cores while the AMD Epyc is at 64 cores.
Marvell also bests Ampere, the startup run by former Intel exec Renee James that’s also working on an Arm-based server chip. Ampere recently announced the Altra Q80-30 processor a few weeks back sporting 80 cores, but no hyperthreading. Ampere’s strategy is cores over threads.
ThunderX3 supports eight channels of DDR4-3200 with two DIMMs per channel. The IO subsystem offers 64 lanes of PCIe Gen 4.0 with 16 controllers. The processor will be available in both single- and dual-socket configurations. And for floating point operations, ThunderX3 features four 128-bit SIMD (Neon) units per core.
Basic microarchitectural improvements mean the ThunderX3 will offer 25% greater IPC (instructions per clock) performance over ThunderX2. This plus the increase in the core and DDR frequencies will enable an overall gain of more than 60% in single-thread performance versus the previous generation. At the socket level, ThunderX3 provides more than 3x higher socket level integer performance and more than 5x socket level floating point performance over ThunderX2.
This thing is a beast.
And Marvell has decided to kick Intel in the shins by targeting it by name. In presentations sent to the tech press, the company claims that Intel has essentially lost its processor leadership after struggling to get to 10nm. ThunderX3 will be manufactured at a 7nm process, and Marvell is going for the single monolithic chip design rather than breaking it up into chiplets like AMD and Intel are doing. Marvell is also touting its support for PCIe Gen4, something AMD offers with Epyc but Intel does not, yet.
Marvell appears to be targeting all kinds of workloads, rather than picking a few. “Our target cloud workloads such as big data, databases, media streaming, web tier, elastic search and cloud storage are highly parallel in nature. ThunderX3 features 4 thread simultaneous multithreading (SMT) which enables significant throughput improvements for these workloads,” wrote Gopal Hegde, vice president and general manager of the server processor business unit at Marvell, in a blog post.
Hegde also said Marvell would target HPC workloads like quantum physics, quantum chemistry, computational fluid dynamics, genomics, and oil and gas workloads, as well as running native Arm applications that currently deploy on mobile phones and Arm end points as containers or virtual machines in the cloud or at the edge.
Before Marvell acquired it, Cavium was actually grabbing some share in the server space, something multiple Arm vendors attempted and failed. The highest profile gig for ThunderX2 is the Astra supercomputer at Sandia National Laboratories. Microsoft Azure offers ThunderX2-based clusters for internal purposes, and Marvell says it has deals with 20 other hyperscalers.
It’s an ambitious piece of engineering. The benchmarks provided by Marvell look good, but they come from Marvell, not an independent third party, so I will withhold judgement until that happens. Marvell is a $2.7 billion company in terms of annual revenue, taking on Intel, AMD, and to a degree Nvidia. Its primary business is in networking and storage controllers. So it has an uphill climb.
Copyright © 2020 IDG Communications, Inc.