The Economics of Servers Could Soon Change
Marvell said today that it has built a chip designed for servers that uses the same architecture as chips inside cell phones. The chip's four 1.6 GHz processors aren't notable for their performance compared to today's server chips, which use the x86 architecture used by Intel and AMD.
Context & Ripple Effects
The x86 server market spent 2010 racing on core count — AMD's March launch of 12-core server chips against Intel capped a decade in which both vendors competed almost entirely on throughput per socket. Marvell's announcement breaks from that axis: it ports the ARM architecture that dominates cell phones into a four-core 1.6 GHz server part that concedes raw performance to Intel and AMD outright.
First-order effects
- Intel and AMD face the first credible non-x86 entrant in servers, forcing their duopoly to defend on a metric — energy cost per unit of work — where high-clock desktop-derived parts are weakest.
- Cloud and hosting operators gain a candidate chip whose value case rests on lower power draw and density, not benchmark speed, opening a purchasing criterion that did not exist at scale before.
Second-order effects
- The move only pays off if operating systems, middleware and applications get ported to ARM for servers, so the burden shifts to software vendors to decide whether a second server instruction set is worth supporting.
- Rivals are pushed toward low-power responses within x86 itself — lighter cores, micro-server form factors — rather than ceding the efficiency segment to an outside architecture.
Third-order effects
- If performance-per-watt keeps mattering more than peak performance for scale-out workloads, the server market fragments by job type: efficient ARM-class parts for web-scale serving, big x86 sockets for compute-heavy databases — ending one-size-fits-all server silicon.
- A viable second architecture would erode the pricing power of the Intel-AMD pair, since buyers gain a genuine alternative to negotiate against.
The trend: Server silicon is beginning its shift from a single x86 standard measured in clock speed toward workload-matched architectures where watts, not gigahertz, set the economics.