Well over 90% of cloud Infrastructure-as-a-Service (IaaS) instance types are based on Intel Xeon processors. Will AMD EPYC processors take significant cloud market share from Intel, or will some other server processor give both AMD and Intel substantial competition in the cloud?
Intel and AMD took over the server processor market in the mid-2000s. This happened after AMD introduced 64-bit instructions into the x86 instruction set with its Opteron processor architecture. Intel then followed by adding AMD 64 (now x86-64) instructions into its Core processor architecture and Xeon products. After that, AMD lost most of its server processor market share to Intel in the 2008-2018 timeframe.
AMD and Intel are currently the only server processor vendors cross-licensed to use each other’s critical x86-64 patents. Access to both AMD and Intel patents is required to build a modern x86-64 server processor.
A few years ago, Arm thought it was set to gain significant market share against the x86 instruction set. Arm’s business model is to license its Aarch64 instruction set to companies that then design their own server processors. Current licensed Arm server processors include Ampere’s eMAG, AWS Graviton, Huawei’s Kunpeng and Marvell’s ThunderX.
However, AMD’s introduction of its EPYC chiplet-based x86 design caught Arm and its licensees off-guard. Beyond Arm, both enterprise and cloud data center markets had discounted AMD’s ability to innovate.
Meanwhile, academics at Berkeley created the new RISC-V (pronounced “risk-five”) open source instruction set, which is being commercialized by the RISC-V Foundation. Many current and potential Arm server processor designers are evaluating the potential for RISC-V based server processors. They include some large IaaS clouds, like Alibaba Group and Google, plus some chip companies not currently in the server processor design business, like NVIDIA.
To gain share in the IaaS market, like AMD and Arm are both attempting, RISC-V will have to show both enterprise-class quality and cloud workload performance value. But RISC-V is still in an early stage of development and won’t be a mature contender for cloud data center deployments for many years.
IBM recently set its POWER instruction set free into the open source world. IBM’s POWER9 is demonstrably an enterprise-class server processor. IBM did a good job launching the OpenPOWER Foundation, which created a cloud server supply chain ecosystem for POWER9. Google has shown OpenPOWER-based POWER9 boards at events over the past couple of years.
We’ll have to wait and see if Google Cloud deploys POWER9 into its public IaaS. And if so, will anyone else follow?
While open source RISC-V and POWER instructions sets seem like wildcards, the ultimate wildcard may be other x86-64 server processor designs. Many of the patents needed to create a viable x86 server processor have either already expired or will soon expire. It has been 20 years since AMD designed AMD 64 and the Opteron processor! And Intel’s Pentium 4 and SSE instruction sets were started in 1998 and introduced to market in late 2000.
AMD launched its Opteron processor architecture in the spring of 2003 using Intel’s Pentium 4 32-bit instruction set, AMD’s AMD 64 64-bit instruction set extensions and Intel’s SSE2 multimedia instruction set extensions. Using that combination of soon-to-expire intellectual property today would enable new market entrants to bypass software ecosystem challenges faced by Arm, RISC-V and POWER instruction sets.
Remember that patent expiration is based on the date patents were filed, which is typically well in advance of product launch.
Validation is the biggest driver of software ecosystem growth for any processor instruction set. All processors claiming to be instruction set compatible must actually be compatible from both operating framework and applications development points of view. AMD and Intel have impressive x86-64 validation suites, but in the data center world that’s mostly to address three decades of legacy enterprise operating system and application compatibility. Cloud deployments might specify a much lighter validation suite, lowering the bar for new x86-64 server processor entrants.
While there are no public hints of companies designing new x86-64 based processor cores, it takes roughly three to four years to design and validate a new server processor core architecture. That means development should be underway now for 2022-ish server processor product launches, whether they are based on x86-64, Arm Aarch64, POWER9 or RISC-V.
It will take considerable funding to build-out and maintain competitive software development and operating frameworks for the non-x86-64 instruction sets. So it seems like a sound strategy to follow AMD’s lead, but just far enough into x86-64 software compatibility to please a few major cloud providers.
This article was written for Forbes, the original can be found here.