Founded back in 2017, Ampere Computing has in a short space of time established itself as a leading developer of Cloud Native Processors delivering the performance, scalability, security and power efficiency that are essential in delivering hyperscale cloud and edge computing workloads and applications.
The company is led by the industry semiconductor veteran Renee James who worked at Intel for 28 years and was President of the chip giant before retiring in 2016.
Today, Ampere operates out of nine global locations and since 2020 has been responsible for launching several Cloud Native Processor series with the aim of providing more sustainable server solutions for data centre applications. While being more sustainable they also enable cloud service providers to meet the growing compute demands of today’s systems by using a fraction of the power and real estate of many of its competitors.
“Ampere has designed its own cores from the ground up and the company’s goal is to bring a new standard to the hyperscale market - a market where highly-scalable processing capability and reduced power consumption can translate into hundreds of millions of dollars in profits and savings to customers,” explained Jeff Wittich, the company’s CPO. “We’re working with a large number of hyper-scalers and have developed an extensive partner ecosystem.”
Initial product launches included the Altra and Altra Max processors, which are Arm-based chips intended for large cloud service providers. Earlier this year Ampere launched the AmpereOne family, the company’s first set of processors with custom-designed chips which are highly optimised for data centre use cases.
“These new chips, which can support up to 192 cores, were designed in-house and built on a 5nm process. They remain compatible with the Arm instruction set, so there’s nothing new developers have to do to support them,” explained Wittich. “Ampere has been able to optimise these processors specifically for large cloud platforms.”
With the AmpereOne family, every core gets 2MB of L2 private cache and several additional features have been added, ensuring that every core gets access to the same amount of memory bandwidth and more fine-grained power management.
“We believe these devices are uniquely well suited for the cloud,” said Wittich.
Ampere’s focus on performance and power are a consequence of James’ belief that there was a need in the industry for what she described as, “something else”.
Speaking at the launch of the AmpereOne James made the point that new bars were being set for what was required of performance by artificial intelligence (AI) and the connected world, but warned that, “We cannot continue to use power as a proxy for performance in the data centre.”
A sustainable approach
“We need to design products that maximise performance but are more sustainable. Moore’s Law has slowed and while legacy processors have looked to improve performance by throwing more power at problems, that traditional form of scaling is proving to be no longer viable or practical, in many instances,” said Wittich. “Ampere looked at the constraints under which data centres were operating and the very scale of the Cloud and we concluded that a more sustainable and green approach would be needed.”
As Wittich explained data centres are now the foundation for all things to do with cloud computing and with AI software now being put to work at the data centre level, ever more processing capacity will be needed. Today, even smaller businesses are deploying high-end servers to run complex computing tasks.
To meet demand from data centres for more computing power but with a processor with lower power requirements, Ampere’s first 128-core data centre processor was launched back in 2020 stealing a march on the likes of AMD, Intel and Nvidia.
“Power is a serious constraint and we knew that something different was going to be required. Currently, data centres account for around 3 per cent of total global power consumption, so it’s imperative we look to curb growth in power consumption. It’s unsustainable. So, Ampere has taken a greener approach to CPU development, developing a core that is energy efficient and sustainable, which with energy prices as high as they are, has to be a key consideration.
“If you want to deliver more computing performance you need to design a CPU that is both efficient and that scales, but we didn’t want to take an incremental design approach – tweaking an existing building block, for example - but rather we took a clean sheet and built a processor from the ground up, with our own core, which would be inherently efficient and scalable. Scalability is crucial and not all processors can deliver that,” Wittich said.
That scalability of the latest Ampere architecture goes well beyond existing products such as the Altra Max but also exceeds many of what Ampere’s competitors are looking to offer.
“We’re operating in a highly competitive space but with the launch of the AmpereOne family we’ve gained an edge over our competitors,” Wittich suggested. Many market analysts would agree with that claim.
While both Intel and AMD are set to launch rival products – AMD’s Bergamo is just 128 cores while Intel’s Sierra Forest, set to arrive in 2024, is 144.
“Use cases for the general-purpose AmpereOne CPU will include AI inferencing, web servers, databases and video streaming and while this chip is intended for cloud providers, it will also be marketed to enterprises who build private clouds,” added Wittich
Wittich makes several important points regarding how Ampere has approached the development of these products.
“The first thing to say is that we’ve taken a back-to-basics approach from an efficiency perspective. Everything we do in terms of adding new features is considered in detail, and whether the changes being proposed will deliver incremental performance improvements. We use a very sophisticated emulation environment for any tweaks or changes to features early on, so that everything we do is focused on improving efficiency and performance.
“Also, we’re addressing an environment in which a lot of different applications will be fighting for resources – whether that’s the same user or multiple users – so we look to create as much resource isolation as we can. Our cores are single thread rather than multithreaded and that is essentially because of the variability in terms of performance that can occur with the latter approach.
“Finally, latency matters. There are a lot of users running applications in the cloud, in the network or at the edge. You’re dealing with a lot of data, so tasks are competing with one another, so latency and its distribution matters.”
According to Wittich the key for Ampere is designing for and delivering consistent predictability.
“Our main competitors Nvidia, Intel and AMD are responding, and our message and approach has certainly resonated with customers,” said Wittich.
“Our competitors are not going to radically overhaul their existing designs or mimic our attitude towards design, but we are seeing developments that are more akin to the approach we have taken.”
Wittich makes the point that Ampere has been delivering processors for several years – starting with 80 cores, then 128, and now its AmpereOne with 192 cores.
“We’ve been able to build our knowledge and have a better understanding of what our customers need. Consequently, based on customer feedback we’ve been able to adapt our designs,” explained Wittich.
“In the past few years there’s also been a better understanding about how power is a real constraint for data centres, so when we talk to customers it really isn’t a hard conversation.
“We have credibility in this market, with a very talented and experience team in place, and we can now point to some of the biggest data centres and enterprises that are using our processor technology.
“Ampere can deliver so much more in terms of performance and low power but crucially for data centres deploying our technology it doesn’t require any radical changes in data centre layouts, there are no additional thermal management issues, and they get significantly more capacity with the same footprint.”
According to Wittich, “That makes selecting our technology a pretty easy decision to make.”