#processors
Why Did Intel x86 Beat RISC Processors in the 1990s?
RISC workstations from companies such as Sun, Silicon Graphics and NeXT battled Intel in the 90s and lost. Will the same happen with Arm and RISC-V?
Whenever I write about the ascent of modern RISC processors such as Arm and RISC-V, I always get comments about how RISC processors failed to take on Intel in the past and that they are doomed to fail again.
It is true that there was a battle between RISC and CISC processors in the 1990s. At the time workstations from companies such as Sun, Silicon Graphics and NeXT were hot stuff. Computers magazines had lots of stories about them. I remember going to conferences back then and seeing mostly Unix workstations rather than Windows machines. However within relatively few years these Unix workstations were pretty much all gone.
The question people cannot stop asking is: Why would history not repeat itself? Will x86 not always end up victorious in the end?
To understand why this will not happen this time around, we need to understand why RISC workstations lost back in the day and why the everything will be different this time around.
Intel’s Volume Advantage in the 90s
Production volume matters a lot when it comes to cost of chips. Large and advanced independent chip manufacturers such as Taiwan Semiconductor Manufacturing Company (TSMC), which make chips for Apple and many others today, did not exist in the 90s.
The one who had the most advanced manufacturing capability and the highest volume was Intel. RISC processors at the time were cheaper to design as they were simpler. However lower volume meant higher unit price.
Intel had the opposite situation. They had much higher design costs as the x86 instruction-set made chips harder to design. However, due to their volume they could produce chips at lower cost per unit.
For high-end workstations which would be low volume, there was an obvious advantage with RISC chips. You could design higher performance chips at lower cost. For low volume workstations Intel would not be able to compete as the their design cost could not be amortized across enough volume of chips.
The Unix workstations ended up with advanced custom made hardware at every level which supported all sorts of fancy features such as hot-plugging hardware at runtime, extremely fast memory. I think some workstations even let you add a CPU while it was running but don’t quote me on that.
Combined this produced awesome RISC machines at eye-watering prices.
Open vs Closed Platform
Closed platforms may give better overall user experience, just look at Apple. The problem is that you end up with premium prices. All the RISC workstation vendors were like Apple: They sold premium products at premium prices.
The problem with such a business model is that the low-end competitor always end up improving and closing in on you. The low-end has the benefit of volume which means high-end products very often lose out to low-end products moving up the value chain.
The Threat of “Good Enough”
The problem for RISC based workstations was that eventually PCs started getting good enough. They still had flaws compared to RISC based workstations but the PCs could offer far more software and much lower prices. Thus Intel started gradually eating into the market of RISC workstations which then had to decide whether to fight in the low-end of their business or try to race ahead at the high-end.
With higher volume and higher total profits, Intel could afford to throw a lot more money at designing chips than the RISC guys. They could also use smaller node sizes to increase performance without actually having better designed chips than the RISC guys. This made Intel gradually erode the RISC advantage.
They never stood a chance because they began at the high-end not the low end.
Why RISC Today is Totally Different From the 90s
Today the chip game is completely reverse. Arm did not start by selling high-end RISC workstations. Instead they began by dominating the low-end chips market made up of micro-controllers, embedded devices, dumb phones and smart phones. In 2020 25 billion Arm based chips got shipped. In comparison the whole PC market in 2020 was 275 million units. Think about that! That is an order of magnitude difference.
Here is a crazy fact few seem to have picked up about the world we live in today. People are used to thinking about Apple as being the small player. Macs have about 10% marketshare of the computer market. So in people’s mind Apple is a small player. What people fail to take into account is how dominant smart phones and tablets have become. An iPhone today frequently costs significantly more than a desktop computer or laptop. You got plenty of smart phones capable of outperforming low-end PCs.
These devices have high end CPUs in them. Many of Apple’s iPad models have M1 chips inside. The very same chip used in many of their popular desktops and laptops. Thus in terms of CPU capability it does not make that much sense to separate these markets. So here is a simple statistics that will blow your mind: Apple shipped a total of 285 million iPhones, iPads and Macs in 2020.
Thus there are more CPUs going into Apple products than there are CPUs going into the whole PC market. Thus being king-of-the-hill in the PC market isn’t that big of a deal. This market is what Intel owns but it simply isn’t that big of a market. Apple alone are making of CPUs than Intel.
So if you look at the whole Arm market it is clear that the idea that x86 is still the dominant player is just ludicrous. Arm already won. Intel and AMD today are where the RISC workstation guys were in the 90s. They god the high high-end while facing a high-volume competitor which is rapidly catching up.
x86 Performance Edge is Not Going to Hold
The fastest super computers today are built using Arm not x86 chips. Intel and AMD are able to outperform the chips made by Apple but only at the cost of insane power usage. That is not a sustainable strategy to maintain an edge. Apple has the clear lead in terms of performance per watt. Now Apple cares more about slim design and long battery life than Intel and AMD which both make a lot of money on delivering to data centers. Thus they are not directly competing across the whole product lineup.
But other companies such as Amazon, Ampere and Nvidia are making Arm chips for data centers which will compete directly with AMD and Intel. Ampere has their Ampere Altra Max with 128 Arm Neoverse N1 cores. The Altra outperforms the AMD EPYC 7763 with 64 cores which is regarded as the fastest x86 CPU. Amazon has come out with their Graviton3 chip with 64 Neoverse V1 cores. Meanwhile Nvidia has announced their Grace super-chip which will have 144 Neoverse V2 cores.
It might be interesting to note the difference between the Neoverse N and V cores. The N variants are for traditional server workloads while the V variants are for high performance computing (HPC). With HPC we typically mean scientific oriented workloads such as machine learning, lots of matrix multiplications and other floating-point operations.
So you can imagine you might want the Neoverse N cores for hosting databases and web-servers while you would use Neoverse V for machine learning, data analysis, simulations etc.
The important point is that Arm chips exist for all markets. Apple make the Arm chips that can compete in the desktop and laptop consumer space. Amazon, Ampere and Nvidia supplies the Arm chips which can compete in data centers.
But x86 won in the 90s over RISC! Sorry but the rules have changed. Arm has higher volume today and Arm chip makers have access to equal or superior manufacturing through companies such as TSMC.
What About RISC-V?
It is not just when talking about Arm chips that the x86 victory over RISC in the 1990s is mentioned. It is also a point raised whenever somebody talks about the future of RISC-V. Again exactly the same story applies to RISC-V as with Arm. RISC-V chips can in principle be built by TSCM so there is no node size advantage for Intel.
What about volume? RISC-V like Arm is starting at the low end where volume is high. RISC-V processors are going into hard drive controllers, mice, keyboard, bluetooth devices, IoT and a dozen other things today. Thus RISC-V is a lot more like Arm than the UltraSPARC, MIPS and Alpha RISC processors of the 1990s.
Of course it will take time for RISC-V to gain the kind of popularity and performance Arm enjoys today, but there is nothing stopping RISC-V from beating x86 chips sometimes in the future. However, by the time RISC-V is competing in the high performance space I doubt there are any x86 chips around to compete against. The future will more likely be a battle between Arm and RISC-V. But you should give RISC-V at least another 5–6 years before you can expect it to be going up against Arm.
Chipmakers are notoriously paranoid: Silicon Valley was born in intrigue and suspicion. Despite Intel's iconic CEO Andy Grove making paranoia a corporate mantra, Intel became relatively relaxed. Qualcomm and Apple would throw you into their piranha pools merely for asking questions if they could, while Intel has learned to give as well as take. But it may be going back to bad habits.
https://www.theregister.com/2021/12/13/chipzillas_mystery_linux_muckabout_is/
