Alibaba's Yitian 710 is the fastest Arm-based CPU for cloud servers, study claims
An unexpected champion.
A recent study published in the IEEE's Transactions on Cloud Computing journal has demonstrated that Alibaba Cloud's 128-core Yitian 710 processor, developed in 2021, is currently the most efficient Arm-based server processor for database tasks in large-scale cloud environments, reports The Register. The research conducted by Dumitrel Loghin from the National University of Singapore, showcases the Yitian 710 outperforming other Arm-based processors and even surpassing Intel's Xeon Platinum in some specific tests.
The research assessed a variety of Arm server CPUs and compared them against the Intel Xeon Platinum 8488C, which was launched in the first quarter of 2023. This comparative analysis was performed across eight different cloud platforms to determine which CPU handles database-related tasks most effectively. Among the Arm processors evaluated — Amazon's 64-coreGraviton 2/3, Huawei's 60-core Kunpeng 920, and Ampere's Altra 80-core processor — the Yitian 710 emerged as a standout performer, especially noteworthy given its newer technological underpinnings. As for workloads, they included Dhrystone and Whetstone benchmarks, measuring the number of kernel system calls and execl calls each system made, testing file copy speeds, assessing the overall UnixBench score, and conducting RSA 2048 signing and verification tasks.
Alibaba's Yitian 710 was ahead of its rivals in synthetic Dhrystone and Whetstone benchmarks and significantly challenged Intel's Xeon 8488C, which demonstrates its huge potential. The Yitian 710 also outperformed its competitors in File Copy, UnixBench, and RSA 2048 signing and verification tests.
Despite the strengths of the Yitian 710, AWS's Graviton 3 surpassed it in the Redis benchmark. However, the Yitian 710 maintained its advantage in handling memcached throughput and in two out of three RocksDB throughput tasks, the report says.
The study also highlighted the Yitian 710's technological advantages, such as its use of the latest Arm instruction set architecture (ISA) and DDR5 SDRAM in 2021, which contributed significantly to its high performance. On the other hand, Huawei's cloud offers a unique feature where two Kunpeng 920 CPUs can run in a Non-Uniform Memory Access (NUMA) topology, a capability unmatched by any other cloud provider's Arm servers.
Cost efficiency was another focal point of the research. Although Arm servers require more system calls than their Intel counterparts (which is potentially inefficient), they still offer better value in some scenarios. This is particularly relevant as cloud providers continue to optimize for cost-effectiveness without compromising
The study suggests that with incremental improvements, such as slight increases in clock speeds and optimizations in Linux kernel system calls, Arm servers like the Yitian 710 could become even more competitive. The research recommends leveraging Arm-specific features like Scalable Vector Extensions to boost performance in respective workloads, potentially positioning Arm servers as a viable alternative to more established AMD EPYC and Intel Xeon servers in cloud computing scenarios.
Stay On the Cutting Edge: Get the Tom's Hardware Newsletter
Get Tom's Hardware's best news and in-depth reviews, straight to your inbox.
Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
-
Findecanor The actual article is paywalled, so I have to rely on the graphs in the Register.Reply
I think many of Alibaba Yitian's wins in some benchmarks can be attributed to it simply running at a higher clock than the ARM competition. -
Alvar "Miles" Udell Not entirely unsurprising, ARM vs x86 and all that, but how do they perform in real world workloads?Reply -
Amdlova On server side claims are "only show this perfect side". On real world it's another thing.Reply -
HaninTH
It is not our faults their government runs their op like they do. If they were a little more open with some stuff, maybe it would be easier to believe things like this are wholly legitimate.nookoool said:It will be okay guys. It will probably be an non-Chinese ARM processor next time.
Until the rest of the world can test these claims, we have to assume they're not entirely kosher.
I really do not care whom creates the technology that will drive our future. I just want it to be cost effective for end-users and deliver the performance we should expect from newer generations of technology.
The rich and powerful will get the backdoors or access they want into these systems regardless of who makes the physical kit. We need to let go of that, as the horse has long left the stable of personal privacy. Not like most of us care about it anyway, as long as we can get on the latest social media app and get in on the latest fad. -
nookoool HaninTH said:It is not our faults their government runs their op like they do. If they were a little more open with some stuff, maybe it would be easier to believe things like this are wholly legitimate.
Until the rest of the world can test these claims, we have to assume they're not entirely kosher.
I really do not care whom creates the technology that will drive our future. I just want it to be cost effective for end-users and deliver the performance we should expect from newer generations of technology.
The rich and powerful will get the backdoors or access they want into these systems regardless of who makes the physical kit. We need to let go of that, as the horse has long left the stable of personal privacy. Not like most of us care about it anyway, as long as we can get on the latest social media app and get in on the latest fad.
Did you see who ran the test? -
bit_user
The 8488C has only 48 cores (96 threads)! For sufficiently scalable tasks, it's better to have more cores, even if they're much simpler.The article said:Alibaba Cloud's 128-core Yitian 710 processor ... compared them against the Intel Xeon Platinum 8488C
synthetic Dhrystone and Whetstone benchmarks
These are no longer commonly used, since they don't tend to reflect common workloads.
File Copy
This just needs a cache optimization, similar to what Ampere's Altra had.
Source: https://www.anandtech.com/show/16979/the-ampere-altra-max-review-pushing-it-to-128-cores-per-socket/3
RSA 2048 signing and verification tests.
This could benefit from a hard-wired crypto accelerator. Does it have them?
two Kunpeng 920 CPUs can run in a Non-Uniform Memory Access (NUMA) topology, a capability unmatched by any other cloud provider's Arm servers.
Ampere's Altra can do that. -
HaninTH
Yes, and as bit_user has pointed out, their information is invalid, to say the least.nookoool said:Did you see who ran the test?
I have no problem believing a chinese firm can develop a chip that could rival western firms, but can they do it at scale reliably? This, i'm not confident in. -
bit_user
I was trying to make the case that it was irrelevant, not invalid. In other words, I don't dispute the results they found on their chosen test suite, but I consider the test suite either not to match modern cloud workloads or simply to be ones which favor a greater number of simple cores vs. fewer that are more sophisticated.HaninTH said:Yes, and as bit_user has pointed out, their information is invalid, to say the least.
The narrative of the piece seems to imply that this CPU was far ahead of its time, which is the main point I dispute. However, that doesn't mean it wasn't a milestone. If they managed to deploy 128-core ARM CPUs before Ampere, that would've been a nice achievement, in itself.