In a world where semiconductor consolidation is the norm, it’s not often that a new player enters the field. Even fabless semiconductor companies have been the target of mergers and acquisitions (Qualcomm being the most recent and largest example) making the recent emergence of Ampere all the more interesting. Ampere is building a new Arm-based processor and platform architecture to address the hyperscale cloud compute demands of today and the future.
Though the name will be new to most of you, the background and history is not. Owned by the Carlyle Group, which purchased the Applied Micro CPU division from MACOM last year, Ampere has a solid collection of CPU design engineers and has put together a powerful executive leadership team. At the helm as CEO is Renee James, former President at Intel, leaving the company in 2015. She brings a massive amount of experience from the highest level of the world’s largest semiconductor company. Ampere also touts an ex-AMD Fellow, former head of all x86 architecture from Intel, ex-Intel head of platform engineering, and even an ex-Apple semiconductor group lead.
Architecturally, the Ampere platforms are built with a custom core design based on the Arm architecture, utilizing the ARMv8 instruction set. Currently shipping is the 16nm processor codenamed Skylark with 32-cores and a 3.0 GHz or higher clock speed. The platform includes eight DDR4 channels, 42 lanes of PCI Express 3.0, and a TDP of 125 watts. The focus of this design is on memory capacity and throughput, with competitive SPECint performance. In my conversation with James last week, the emphasis on memory and connectivity is a crucial component of targeting lower costs for the cloud infrastructure that demands it.
The second generation of Ampere’s product stack called Quicksilver is coming in mid-2019. It will move to the updated ARMv8.2 instruction set, increase core count, improve overall IPC, and add multi-socket capability. Memory speed will get a bump and connectivity gets moved to PCI Express 4.0. It will include CCIX support as well, an industry-standard cache coherent interface for connecting processors and accelerators from various vendors.
Interestingly, this part will be built on the TSMC 7nm process technology which Ampere CEO James says will have a “fighting chance” to compete or beat the capabilities provided to Intel by its own in-house developed process technology. That isn’t a statement to make lightly and puts in context the potential impact that Intel’s continued 10nm delays might have for the company long-term.
For systems partnership, Ampere is working with Lenovo. This is a strong move by both parties, as Lenovo has a significant OEM and ODM resources, along with worldwide distribution and support. If the Ampere parts do indeed have impact in the cloud server ecosystem, having a partner like Lenovo that is both capable and eager to grow in the space provides a lot of flexibility.
Hardware is one thing but solving the software puzzle around Ampere’s move into the hyperscale cloud server market is equally important. James told me that the team she has put together knows the importance of a strong software support system for enterprise developers and seeing that happen first hand at Intel gives her a distinct advantage. Even though other players like Arm and Qualcomm are already involved in the open source community, Ampere believes that it will be able to make a more significant impact in a shorter period, moving forward support for all Arm-processors in the server space. Migrating the key applications and workloads, like Apache, memcache, Hadoop, and Swift to native, and most importantly efficient, code paths is required for widescale adoption.
Followers of the space may be wondering why now is the right time for a company like Ampere to succeed. We have seen claims and dealt with false promises from numerous other Arm-based server platform providers, including AMD and the source of Ampere’s current team, Applied Micro. Are the processors that much different in 2018 from those that existed in 2013? At their core, no. But it’s the surrounding tentpoles that make it different this time.
“Five years ago, this couldn’t have happened,” said James in our interview. The Arm architecture and instruction set has changed, with a lot more emphasis on the 64-bit superset and expanding the capability for it to address larger and faster pools of memory. Third party foundries have caught up to Intel as well – remember that James believes TSMC’s 7nm node will rival Intel competitively for the first time. Finally, the workloads and demands from the datacenter have changed, moving even further away from the needs of “big cores” and towards the smaller, more power efficient cores Ampere and other Arm options provide.
Obviously, that doesn’t apply to ALL server workloads, but the growth in the market is in that single-socket, memory and connectivity focused segment. AMD backs up Ampere’s belief here, with its own focus on single-socket servers to combat the Intel dominated enterprise space, though EPYC still runs at higher power and performance levels than anything from the Arm ecosystem.
James ended our interview with a comparison of the Arm server options today to x86 servers more than 25 years ago. At the time, the datacenter was dominated by Sun and Sparc hardware, with Sun Microsystems running advertising claiming that Intel’s entry into the space with “toy” processors wasn’t possible. Fast forward to today and Intel has 99% market share in the server market with that fundamental architecture. James believes that same trajectory lies before the Arm-based counterparts rising today, including Ampere.
There is still a tremendous mountain to climb for both Ampere and the rest of the Arm ecosystem, and to be blunt, there is nothing that proves to me that any one company is committed completely. Qualcomm has announced its Centriq CPUs last year and Ampere claims to have started sampling in 2017 as well. We don’t yet have one single confirmed customer that has deployed Arm-based systems in a datacenter. Until that happens, and we see momentum pick up, Ampere remains in the position that previous and current Arm-based servers are found: behind.