Environmental sustainability is not just a corporate social responsibility but a corporate imperative in many businesses. Regardless of your views on ESG, companies' boards of directors and their customers are citing this as necessary. As energy costs are skyrocketing, companies can be “green” and save money. This is important as few enterprises I have talked to outside the Nordics are willing to pay more for higher ESG capabilities. That was evident at the recent Open Compute Project (OCP) Global Summit, where sustainability was the theme in almost every booth.
We are all aware of the rising energy consumption of hyper-scale data centers the comparison to towns and even small countries. Of course, the solution is straightforward - more computing with less power. This opportunity has created a new wave of innovative startup chipset companies. I recently wrote an article about Groq, which focuses on discrete chips for specialized workloads such as artificial intelligence (AI).
Ampere is another rising star with the same goal but a different strategy. Ampere designs purpose-built, multi-core processors optimized for multi-tenant cloud workloads.
Time for a new type of CPU
I am an avid chip guy who has been in the industry for many years. This story reminded me of the transition from mainframes to a client-server model a couple of decades ago. First, the usage model changed, followed by the adaptation of the software to become higher performance more and more efficient, and then finally, the hardware changed. We are going through a similar transition right now. Cloud-native has happened from an architectural perspective and also from a software perspective. It is time for a cloud-native processor to drive higher performance and improved sustainability.
Enterprise-class processors – the wrong tool for the cloud
CSPs must have profitable cloud infrastructure while delivering the performance customers expect. Enterprise-class processors are not the right tool for cloud-native workloads and can negatively impact customer SLAs and add unpredictability.
Enterprise-class processors support applications written as monolithic blocks of code executed in a dedicated environment. Enterprise-class processors have evolved with increasing power and large amounts of memory with innovations such as high-capacity cores, higher CPU frequencies, and larger caches.
Simultaneous multithreading (SMT) and Turbo Boost are enhancements that can be effective for the correct use case but not for latency-sensitive, scalable cloud-native workloads.
SMT enables each physical core to split into two logical threads simultaneously executing separate instruction sequences. Turbo Boost automatically runs cores faster than the rated operating frequencies.
SMT can cause problems in a multi-tenant cloud environment. Multiple applications must share execution resources when a CPU core has two SMT threads. A resource-intensive workload, known as a noisy neighbor, will cause workloads on the other SMT thread to slow from lack of execution resources. Similarly, Turbo Boost can cause unpredictable performance depending on the type of workload.
In summary, enterprise-class cores perform well with computationally intensive applications but not in a shared cloud infrastructure.
Ampere taking a clean-slate approach to CPU design
Cloud-native applications have distributed components, such as micro-services, that perform specific tasks and collaborate to achieve an objective. A benefit is that lightweight applications are faster to develop, test, and integrate, leading to practices like continuous integration/continuous delivery (CI/CD).
Ampere took a clean-slate approach to CPU design that caters to this new software paradigm in multi-tenant environments. CSPs want to host more end users per server with dedicated physical cores. Ampere responded with a processor design that offers near-linear scaling across a 128-core Altra processor.
Ampere's processors are immune to the "noisy neighbor" issues that plague enterprise-class processors with SMT. Ampere's processor cores are single-threaded, resulting in no resource contention and more predictable performance. With Ampere, each thread runs solo on a single core.
Changing the game in terms of sustainability
Ampere starts with single-threaded cores - one thread, one process - different from the enterprise-class processors with multiple threads competing for processor resources. Multiple threads create an unpredictable environment, whereas a single-threaded core provides the same performance with every core resulting in a high degree of predictability.
Ampere uses a constant operating frequency delivering predictability as opposed to x86 processors that employ frequency scaling. Ampere uses power-efficient cores, which allow the stacking of many cores into a processor. The highest core counts in the industry.
Ampere pipelines cores with a large private L2 cache located right next to the core. Loading data and instruction sets into the L2 cache results in a predictable performance profile for each core.
The high number of cores makes it cost-effective for CSPs to rent each core to a single customer. Customers get scalability and predictable performance. CSPs can run at the lowest possible power with fully populated racks with no stranded capacity. Running at the lowest power reduces the thermal load. The net result is a quantum leap in efficiency and maximum performance per rack.
Microsoft now offers Azure Virtual Machines with Ampere Altra Arm-based processorsthat can run Linux workloads such as web servers, open-source databases, in-memory applications, big data analytics, gaming, and media.
An Ampere Altra VM is a high-performance compute alternative that scales up linearly, delivers predictable performance at full utilization, and is power efficient, directly reducing users’ overall carbon footprint.
Ampere has addressed the simple question posed at the beginning - more compute with less power. A cloud-native processor with single-threaded cores, consistent operating frequencies, and the most power-efficient cores in the industry, topped by a large L2 cache close to the processor resulting in a cloud-native processor built for the sustainable cloud.
Microsoft has proven this isn't a fantasy - the solution exists today. Instead of building out data centers with compute using enterprise-class processors switching to Ampere's cloud-native processors could reduce power consumption by 20 percent and still meet the compute needs.
Other major players have also embraced the Ampere technology. Hewlett Packard Enterprise announced in it will offer a cloud-native server in later in 2022 using Ampere chips. Google Cloud has launched virtual machines using Ampere Altra Arm-based processors. Oracle, in addition to being a major investor, offers a comprehensive line of Ampere platforms on Oracle Cloud Infrastructure (OCI).