How much single-thread performance is enough in a server? Too much, and power efficiency declines; too little, and some applications slow down. AppliedMicro believes it has the right answer to that question with its new ARMv8 CPU design. At Hot Chips, the company disclosed microarchitecture details of its forthcoming X-Gene processors. Its disclosure is also the first for any 64-bit ARM design—ARM itself has yet to reveal details of its Atlas and Apollo CPU cores. By designing a custom CPU, AppliedMicro is on track to be first to market with an ARMv8 processor.

The X-Gene design uses a dual-CPU module as its basic element, and each module connects to a coherent fabric. The fabric is AppliedMicro’s own design, providing added performance and scalability compared with the standard fabrics available from ARM. The two CPUs share a level-two cache, whereas all modules share a single level-three cache. Each CPU includes integer, floating-point, and SIMD units as well as level-one instruction and data caches; there is no Bulldozer-style sharing of function units. The X-Gene design also includes hardware support for virtualization, enabling an efficient hypervisor implementation.

The X-Gene CPU is a single-threaded four-issue superscalar design with out-of-order execution. Although its design appears complex, the company avoided features that increase power dissipation more than performance. Rather than attempting to match Intel’s single-thread performance, AppliedMicro’s goal was to minimize power and die area while delivering what it feels is “good enough” single-thread performance. Still, on Google’s brawny-versus-wimpy scale, X-Gene sits on the brawny side. Although the company quotes a 3.0GHz target clock speed in 28nm technology, AppliedMicro’s first instantiation will be a 40nm processor targeting 2.4GHz; this chip is due to sample in 4Q12.

Following Cavium’s recent announcement, AppliedMicro is no longer alone in pursuing high-performance server designs using a custom ARMv8 core. By disclosing X-Gene’s microarchitecture, AppliedMicro is reinforcing its first-mover advantage and showing it understands the performance requirements of mainstream servers. Much now depends on the company’s execution, lest it allow competitors to close in.