Cloud computing adoption has been estimated to grow at a rapid clip, by as much as 30% annually. From billions of new, connected consumer and commercial devices placing demand on cloud resources, to the rise in AI machine learning and software as a service products (SaaS) from countless players, the modern data center is experiencing almost insatiable demand. As you might imagine, hyperscalers and modern enterprise data centers are also now commanding a huge energy draw, such that some estimates claim, if non-renewable electricity is the primary source for data centers, that cloud compute could actually surpass the aviation industry in terms of global CO2 emissions. It’s therefore also critical that data centers move to more efficient compute platforms, and that’s where relative newcomer Ampere Computing could make a major impact on a more efficient, sustainable cloud.
Former Intel Corporation President, Renee James founded Ampere Computing back in 2018. She currently serves as Chairman and CEO of this chip company that’s making significant headway with major hyperscalers like Microsoft Azure, Google Cloud, Oracle, cloud infrastructure OEMs like HPE, and content delivery networks like Cloudflare. The first product out of the gate was an 80-core chip called Ampere Altra, but the company quickly followed-up with the 128-core Ampere Altra Max, that brings the highest core density per socket of any data center server platform currently, but also newfound efficiency and performance in traditional cloud workloads.
Pioneering and ultimately defining the “cloud-native” processor category, Ampere’s first gen Arm Neoverse N1 cores sport dual 128-bit SIMD (Simultaneous Instruction Multiple Data) engines with some key advantages like larger, lower latency private caches and single thread execution. Single-threaded CPU cores have significantly more deterministic performance with dedicated resources, eliminating the “noisy neighbor” effect that can be a major issue for traditional X86 processors in cloud workloads. This common phenomenon occurs when a virtual machine(s) consumes an inordinate amount of available compute resources in a cloud multi-tenant infrastructure, thus degrading performance for other tenants. As such, modern hyperscalers must take steps to mitigate this issue with traditional multithreaded X86 platforms from Intel and AMD, though Ampere’s ultra-high core count single-threaded architecture doesn’t share this issue.
Multi-tenant cloud workloads are more efficiently served when each service is executed on a single thread/core, so as to ensure predictable performance and avoid resource conflicts. This hardware-level isolation is also generally better for implementing cloud service security protections. The net-net here is that Ampere claims its Altra and Altra Max processor platforms allow cloud service providers better concurrent workload throughput, more predictable performance, better resource utilization and most of all better performance-per-watt characteristics across all workloads, versus current X86 server solutions.
Above, Ampere is highlighting performance of its 128-core Altra Max 128-26 CPU versus Intel Xeon 40-core and 64-core AMD EPYC solutions in typical cloud workloads like NGINX, MySQL, Redis and h.264 media transcoding. As you can see, the company is claiming a 2X to almost 4X uplift in performance in some cases, versus these Intel and AMD X86 architectures, which results in higher throughput along with lower latency, and delivered with lower power consumption as well.
Further, Ampere’s own internal modeling claims its technology could cut compute power consumption by 20% and server real estate by 30%. These improved efficiencies could prove to be highly impactful in the effort to bring more sustainable data centers on line to meet the ever-growing demands for cloud-first and cloud-native solutions and services.
Ampere is also working on a custom CPU core architecture based on the Arm instruction set, and it’s expected to bring even better performance characteristics to the company’s many-core CPU portfolio.