At Computex 2018 in Taiwan, Cavium touted products that use their recently-released ThunderX2 ARM CPU for servers, while simultaneously offering assurances about the health of ARM for servers. That ecosystem is facing uncertainty as Anand Chandrasekher, head of Qualcomm's Centriq server operations, left the company last month following a Bloomberg report claiming the company was planning to sell or close that division.
Gigabyte and Cavium are offering the H621 2U server platform, which can scale to up to 4 dual-socket ThunderX2 nodes—with the highest tier ThunderX2 SKU at 32 cores and 4 threads per core, this works out to 1024 threads. Each node has 16 DIMM slots, and Cavium touts the system as being able to address up to 4TB of RAM. The H261 is available in two storage configurations, one with 12 3.5" drives, or 24 2.5" drives, both of which support two low profile x16 PCIe 3.0 Slots and 1 x OCP x16 Mezzanine Card.
SEE: New equipment budget policy (Tech Pro Research)
The two companies are pushing this solution for HPC use cases, much like HP Enterprise's Apollo 70 offering—which is also powered by the ThunderX2. Pricing information was not disclosed, though given that the ThunderX2 has more than 40 SKUs, pricing should be expected to vary. In terms of CPU cost, prices should be expected to be lower relative to competing processors from Intel.
At Computex, Cavium also unveiled a new platform for running Android VMs on ThunderX2, which necessarily has no emulation overhead as would be experienced when emulating ARM on x86-64.
While the ThunderX2 is competitive with Intel's Xeon line of server processors, there is headroom for optimization. To date, as most operations run on Intel systems, existing programs and applications are optimized for x86-64.
CloudFlare, which is using Qualcomm Centriq-powered servers (for now), has been working to optimize individual apps for ARM, and has published their results (and code) for others to use. A recent post about optimizing jpegtran indicates the program was 1.3x faster in NEON than the compared Xeon after optimization, though was only about half as fast as the same Xeon for the unoptimized program. (Much of this relates to gcc's handling of intrinsics on ARM.)
In an interview with TechRepublic, Cavium Marketing Director Surya Hotha expressed confidence in the performance of the ThunderX2, stating that: "The optimizations are still underway. We're working very closely with ARM to get the optimizations done on the gcc front, and the ARM compiler, and things like that, so you will see the numbers going higher as we move forward with the optimizations" but that at present, the performance is still good now.
For the future
Cavium started work on the original line of ThunderX processors in 2012, though the ThunderX2 is was originally developed at Broadcom as "Vulcan," which was axed after Broadcom was acquired by Avago, after which Cavium acquired the IP. Hotha dismissed concerns about Cavium's pending acquisition by Marvell negatively affecting the ARM-based server business, indicating that Marvell sees this as a growth market.
Building a slide deck, pitch, or presentation? Here are the big takeaways:
- Gigabyte and Cavium are offering the ThunderX2 ARM-based H621 2U server, which can reach up to 1024 cores per unit.
- Optimizations for developer toolchains are still ongoing, though current performance is competitive.
- Special report: The future of Everything as a Service (free PDF) (TechRepublic)
- Windows 10 on Arm, Android phones: Our new CPU will match Intel Core i7, says Arm (ZDNet)
- The Machine by Hewlett Packard Enterprise: The smart person's guide (TechRepublic)
- Cavium's ThunderX2 processor powers first 64-bit ARMv8 workstation (ZDNet)
- Windows 10 on Arm: What we learned at Build 2018 (TechRepublic)
James Sanders is a Tokyo-based programmer and technology journalist. Since 2013, he has been a regular contributor to TechRepublic and Tech Pro Research.