Over the years, we have written a lot about the software barrier to entry for Arm silicon in the data center. For 20 years, Intel played a big role in promoting the use of software running on x86 CPUs. As a result companies building Arm-based CPUs had to struggle to convince hundreds of infrastructure software makers to port and optimize their software for Arm. This pushed many out of the market, but ultimately the benefits of Arm (and Intel’s missteps) breached that barrier and we are now seeing a rapid expansion of Arm-based products throughout the Cloud. So now the question is can RISC V accomplish something similar and pry open the data center for its products? This is a story about individual chip vendors, but it is also a story about the power of ecosystems and maybe the power of open source.
The challenge any new Instruction Set Architecture (ISA, i.e. Arm, x86 or RISC V) faces is that moving software to a new ISA is a hassle. Modern compiler tools are sophisticated enough that they can with re-compile for each with a click of the button, but that does not ensure the best performance. Roughly speaking, merely re-compiling a piece of code will only yield 80% of the performance. That remaining 20% is enough to make it not worthwhile to port at all, but that 20% requires a lot of laborious, low-level optimization work. Coders spending hours mucking around in low-level languages and documentation. Intel’s advantage was the software industry spent decades doing that work because they wanted the best performance for their products when x86 was the only game in town. Arm was able to break through only after a dozen chip vendors spent billions of engineer hours doing much of this work, and then Amazon came along with Graviton to really drive the process home.
In all of this it is important to remember that as much as we talk about ISAs as agents in this process, it is not the ISA doing the work, it is engineers at actual companies. For x86 this largely meant thousands of engineers at Intel. If you speak to Arm True Believers, they will quickly point out that Arm has the advantage of an entire ecosystem to do all this work. Hundreds of Arm licensees could eventually overpower the work of an individual company, even one as big as Intel.
So will the RISC V ecosystem be able to prevail in a similar way?
There are now a dozen or so companies building advanced logic devices (CPUs, GPUs, AI accelerators, etc.) on RISC V. Most of these are in China, but not all of them. Some of these companies have very serious founding teams with immense engineering talent behind them. Despite this, they are starting from the same position as Arm CPU vendors 10 years ago. How will they get the big software infrastructure optimized for RISC V – all those databases, BiOS, firmware, VMs, analytics and the rest?
x86 has one company calling the shots. Arm had one company herding cats, but there is no similar central authority for RISC V. Yes, there is a central org behind RISC V but it too small and poorly resourced to provide a driving force.
That being said, RISC V’s completely decentralized model may prove the solution to this dilemma. If the difference between x86 and Arm was the distributed nature of the latter, then maybe even more decentralization will yield even faster results for RISC V. A quick Google search on the subject yields a surprising number of software projects and companies already doing the work needed to move to RISC V. Much of this optimization work is already taking place, sui generis. Admittedly, this is anecdotal, but there is no easy way to measure progress on this, and from our point of view it certainly looks like RISC V is showing real progress.
We think there are a few reasons for this. First, compute is changing. CPUs no longer dominate the data center they way they once did. So there is already a lot of work being done to port to new types of chips like GPUs. Secondly, the progress of Arm has shown it can be done raising the software industry’s awareness of what is possible, allowing RISC V to draft in that progress. The flexibility of RISC V also means that software companies likely feel much more comfortable investing in this effort. They do not have to sign any NDAs or licenses to get a hold of technical materials, and can tweak their software more as they see fit. Last, but not least, there are a lot of software people who Believe in Open Source, and want to see an Open Source semi project succeed.
We could very well end up in a world in which open source software runs on an open source semi, while closed software as in cars and mobile phones, runs on Arm.
None of this is to say that RISC V will have an easy time of it. As we have noted, RISC V itself faces a significant fragmentation process which very much runs counter to the economics of software. RISC V’s flexibility means that everyone may use their own flavor of it, reducing cross-compatibility Breaking into the data center will be a punishing process. On the other hand, as the data center becomes more heterogeneous companies may see the flexibility of RISC V as a way to differentiate their product.
This will not be a simple transition, but there are plenty of reasons to think it is at least feasible.