SAN FRANCISCO—ARM Ltd. executives sketched the broad outlines of an ambitious but potentially risky move into high-performance computing Wednesday (Sept. 8), disclosing some details of the 2.5-GHz Cortex A15 processor, codenamed Eagle, at an event here.
The long-rumored move sets the stage for potential head-to-head competition against Intel in server and other high-performance computing markets. Intel’s relentless drive for performance has yielded increasingly power-hungry processors and prompted concern among OEMs and consumers about everything from power dissipation to cost to environmental impact.
“I’m very excited to see someone who knows power come up the performance curve,” said Mitch Markow, strategic processor technologist with Dell who participated in the event as a panelist.
Markow wouldn’t say whether Dell was committed to using the eight-way multicore capable A15 in a future Dell server product, but he said the company is looking at the device “very closely.”
There were precious few technical details of the Cortex A15 Eagle during the event; company executives said more details would be forthcoming late this year. But it is possible to piece together enough information to get some idea of the scope of the processor and to infer what ARM is planning. The bottom line is that the Eagle is not your smartphone’s ARM core.
A technology departure
To start at the fundamentals, the A15 core has been designed in conjunction with the development of the 32nm and 28nm process technologies that IBM, GlobalFoundries, and Samsung are taking into production.
This in itself is something of a departure for ARM — there was no mention of a TSMC version of the core. One presumes that market pressure will demand one, but the absence was striking, and perhaps is a signal of the market presence that GlobalFoundries and Samsung now command in the foundry space.
On this process roadmap, ARM has constructed a 2.5 GHz-capable superscalar core. These figures in themselves depict high single-thread performance, which ARM did not quantify beyond saying the single-thread performance of the A15 would be about five times that of the previous top-of-the-range Cortex A9.
There has also been exacting attention to energy conservation, including the ability to power-reduce individual stages of the execution pipelines on the fly and the ability to get the entire CPU into sleep mode in 10 microseconds, allowing a far more aggressive use of sleep than is possible with existing large-scale CPUs. Thus used either as a single CPU or in multicore configurations, the A15 should hit a very aggressive energy/operation point.
ARM in the cloud
From here on up, the Eagle is clearly aimed not at advanced smartphones or mobile media devices, but at the network infrastructure, server, and cloud-computing space. To begin with, the A15 architecture is designed for big-machine problems. The total address space is 1 terabyte, far in excess of the needs of mobile or embedded computing.
This space is augmented by hardware virtualization support. ARM did not give details, other than to say the hardware virtualization features would be supported by existing virtualization middleware products such as VMWare (see ARMv7 gets 40-bit, virtualization support)
Virtualization hardware will work hand-in-hand with ARM’s TrustZone security system to provide both the task mobility and the intertask security necessary for cloud-like computing environments and massive server farms.
The CPU core also is intended for multiprocessing. ARM showed illustrations of the core deployed in clusters of four CPUs sharing an integral L2 cache, and two of these clusters linked by the previously-announced AMBA-4 bus architecture.
Such a configuration, ARM said, would be fully coherent, easing virtualization and allowing threads to be distributed widely across multiple cores. Again, these are server-class features, not the stuff of embedded processing.
ARM CEO Warren East describes the Cortex A15 CPU
A further hint at thinking about very large systems is ARM’s statement that the core includes soft-error detection and recovery — presumably on the caches, rather than the logic circuitry — so that even as the total amount of memory in the system becomes enormous, up-time can remain reasonable.
The three ARM silicon partners working on the first implementations of the A15 — Samsung, ST-Ericsson, and Texas Instruments -- focused their presentations on high-end mobile devices where single- or dual-core implementations could move a hand-held computer out of the smartphone category and into the realm of serious local computing.
But the overall feature set of the Eagle architecture — virtualization support, the enlarged address space, hardware-supported coherent eight-way multiprocessing, and soft-error contro l— suggest another vision entirely.
That said, Eric Klein, vice president of technical planning for Nokia, said his handset company is keenly interested in the increased performance the A15 brings.
“When you talk about 5x (performance increase), we’re going to use every bit of that,” he said.
A lesson from history?
For all the buzz leading up to the event, some in the audience were left unimpressed.
“It’s predictable. It’s the next logical thing to do,” said Jeff Bier, president and founder of the technical consultancy BDTI. “They’ve got a chance (to make inroads in the server market)…but they were very careful about” not making aggressive claims about their server-market opportunities, he added.
Several observers compared ARM’s potential charge into servers with Intel’s 1990s-era attempts to own the communications-silicon space, which turned into a strategic retreat after several years.
Bier said what would have been more interesting would be to hear ARM CEO Warren East and Mike Inglis, general manager of the company’s processor division, describe real potential for innovative applications that take advantage of low-power / high-performance cores such as the A15.
ARM is holding in early November in Santa Clara, Calif. ARM Technology Conference during which ARM is expected to disclose more details.