Just a quick follow-up to our post yesterday about ARM making inroads into Intel’s strongholds. A few people asked about Intel in the data center. The assumption is that growth in mobile and web apps just drives demand for more compute in the Cloud. Shouldn’t that help Intel to sell more server CPUs and associated chips?
As with everything in compute, this used to be true, but there are major cracks appearing here as well.
Once upon a time, Intel had over 90% share of data center CPUs. These chips are very high performance, and it was believed Intel’s x86 chip architecture had a permanent advantage. Without getting into the technical weeds, the thinking was that a single x86 core offers more raw performance than a single ARM core. The perception was that ARM cores are more power efficient, making them suitable for mobile but not powerful enough for the data center. This no longer holds.
First, over the past 18 months, Intel has lost a lot of share in the data center to AMD. AMD also uses x86 architecture, but they use a 3rd party foundry to produce their chips. Intel has struggled with the last few process nodes in their own fabs, giving a big edge to fabless AMD.
But there is a bigger problem. The server CPU market is very concentrated, ten or so customers consume about 70% of the chips for this market. Those companies are used to getting their way with the supply chain, and Intel’s near monopoly on server CPUs was a major thorn in their operations planning. So in addition to shifting to AMD, they also started making their own CPUs. As we have noted in the past, the big companies are now capable of building their own chips, especially for chips that are perceived as bestowing a competitive advantage.
It turns out that the ARM architecture is flexible enough to overcome performance gaps. On a single core basis x86 wins, but designers can string together enough ARM cores that the whole chip has better performance than a whole Intel chip. There are now ARM chips on the market that offer better performance, both raw performance and when adjusted for power consumption (performance per watt).
Qualcomm was working on server CPUs which looked promising, but shut that effort down a few years back (for reasons best discussed over a beer). A few months after that shut down, Huawei announced a chip that takes a similar approach. We have discussed China’s ambitions to build its chip elsewhere. Huawei has not said much about this product lately, but it is definitely something to keep an eye on. Note that Huawei was a top ten Intel customer.
Even more concerning for Intel is Amazon’s Graviton server CPU. This product is clearly a test run, but points to a bigger effort. Amazon’s AWS is the biggest server CPU customer out there, they consume something like 25%-30% of the industry. There are also signs that other Cloud scale data center operators are designing some subset of their own chips.
As with PCs, none of these spell the end of Intel, but they are major holes in the company’s market share. They also point to the growing phenomenon of heterogeneous computing, the use of different chips for different components of a common task. For decades the watchword of compute has been compatibility. The modern computing industry is built on the notion of a common operating system running on common semiconductors. This led to Intel’s years of near monopoly status. But even this is changing. Compatibility still matters, but increasingly the tools that provide that compatibility exist in software whether in the cloud orchestration tools or in the low level operating system of mobile phones. For the data center this has allowed new types of chips to become important, notably GPUs and accelerators for AI and neural networks math. Owning the CPU no longer commands the strategic heights it once did. Customers increasingly mix and match, designing their own systems to meet there specific needs in very large volume.
Does this mean everyone should run out and short Intel stock? No. They still make a lot of money and have an immense well of talent. That being said, the conditions they face are steadily tightening.
Pingback: Apple Silicon – The Rest of Us | DIGITS to DOLLARS·
Pingback: Arm for sale? | DIGITS to DOLLARS·
Pingback: Trouble at Intel | DIGITS to DOLLARS·
Pingback: Start-up Semis Live | DIGITS to DOLLARS·
Pingback: Did Qualcomm Just Buy a Server CPU Company? – No | DIGITS to DOLLARS·