Alibaba Yitian 710 rated fastest Arm server CPU in the cloud (for now)

Researcher finds it beats Intel's Xeons for speed on one database-related tests, joins AWS Gravitons in cost-efficiency win

The homebrew Yitian 710 CPU developed in 2021 and deployed by Alibaba Cloud is the fastest Arm server processor for rent in hyperscale clouds when handling database-related tasks, according to research published this week in the Institute of Electrical and Electronics Engineers journal Transactions on Cloud Computing.

A paper titled "Are Arm Cloud Servers Ready for Database Workloads? An Experimental Study" finds in the affirmative, because Arm-CPU-equipped cloud servers rattle along and impressive and useful speeds – even outpacing Intel Xeons on some tasks.

The study – by Dumitrel Loghin, a research assistant professor in the School of Computing at the National University of Singapore – considers five Arm-powered server CPUs available in the cloud, plus the Intel Xeon Platinum 8488C (Sapphire Rapids silicon launched in Q1 2023) across eight clouds.

Here's a list of the CPUs Loghin tested, with info about where he ran them.

CPU Graviton 2 Graviton 3 Yitian 710 Kunpeng 920 Altra (3GHz) Xeon 8488C
Tested in AWS AWS Alibaba Cloud Huawei Cloud Azure, Google AWS
Instance tested c6g c7g c8y k1c Azure D32pls v5, Google T2A c7i
ISA ARMv8.2 ARMv8.4 ARMv9 ARMv8.2 ARMv8.2 x86-64
Max cores 64 64 64 60 80 32
Frequency 2.5GHz 2.6GHz 3.2GHz 2.4GHz 1–3GHz 3.2GHz
L1 data cache (core) 64KB 64KB 64KB 64KB 64KB 48KB
L2 cache (core) 1MB 1MB 1MB 512KB 1MB 2MB
L3 cache (shared) 32MB 32MB 64MB 32MB 32MB 105MB
RAM Type DDR4 DDR5 DDR5 DDR4 DDR4 DDR5

Loghin rented the servers mentioned above, then used them to conduct eight tests – namely the Dhrystone and Whetstone benchmarks, the number of kernel system calls and execl calls made by each system, file copy speeds, overall UnixBench score, RSA 2048 signing, and RSA 2048 verifications.

One all those tests, Alibaba's CPU led other Arm processors. On Whetstone it beat the Xeon Platinum, as did Altra's processors.

But Loghin points out that the Yitian has some inherent advantages: it uses a newer version of the Arm ISA, and speedy DDR5 RAM that some rival CPUs cant utilize. Yet some other silicon tested also has unique features: the kc1.15xlarge.2 VM type in Huawei cloud allows a pair of Kunpeng 920 CPUs to be run in harness using a NUMA topology that no other cloudy Arm server can match.

Graviton 3 beat the Yitian when running Redis, but the Chinese chip won out when measuring memcached throughput, and on two of three workloads used to test RocksDB throughput.

Arm server CPU benchmarks in the cloud

Arm server CPU benchmarks in the cloud – Click to enlarge

The paper considers the cost-efficiency of Arm servers compared to their Intel-powered siblings. In some circumstances, on some workloads, Graviton and Yitian represent better value – despite inefficiencies that see Arm CPUs make two times more system calls compared to Xeons.

That kind of factor is why Loghin asserts "Arm servers could be the ideal choice when the performance is not critical."

He thinks Arm servers could be better still with three actions – one of which is small increases in clockspeed, which Loghin believes would deliver "more performance at a negligible increase in cost."

He also wants Linux kernel developers to analyze and improve the performance of the system calls, which are around half as fast on Arm-based systems compared to x86-64 counterparts.

The researcher's last wish is for software developers to utilize Arm features such as Scalable Vector Extensions to improve the performance of their applications.

If those three suggestions are taken, Loghin feels cloudy Arm servers could be even stronger challengers to Xeons.

Or maybe they already are. Loghin acknowledges that AWS's Graviton 4, launched in December 2023, looks speedy and powerful on paper. But as it's currently only offered to preview customers, the rest of us must wait to put it to the test. ®

More about

TIP US OFF

Send us news


Other stories you might like