As far back as the late 1980s, chip pundits have predicted the demise of x86. Intel's CPU architecture was a dead end, they said, and the day was fast approaching when the chipmaker would reach the theoretical performance limit of CISC processor designs.
RISC CPUs were the future, they said. And sure enough, when the server market exploded in the 1990s, companies turned to RISC-based systems from the likes of IBM, Silicon Graphics, and Sun Microsystems to handle their heavy lifting.
Fast-forward to 2005, however, and an incredible turnabout has taken place. Not only does x86 still rule the desktop, but analysts now estimate that anywhere from 80 percent to 95 percent of all servers sold are based on the Intel architecture.
Meanwhile, RISC has fared far less favorably. Some designs, including DEC’s Alpha and Hewlett-Packard’s PA-RISC, have fallen by the wayside. Those still viable, such as IBM’s Power line and Sun’s Sparc, have been largely relegated to niche status as specialty chips for the high end. Even Apple, long a proponent of RISC chips on the desktop, has forsaken PowerPC in favor of Intel processors for its upcoming Macintosh designs.
The bottom line: You can’t escape x86. It’s everywhere, and the trend seems only to be accelerating.
So what happened? How could the pundits have been so wrong? And could we really be looking at the beginnings of an era dominated by a single processor architecture, one whose doom was first foretold almost two decades ago?
Intel architecture grows up
According to Nathan Brookwood, principal analyst at microprocessor consultancy Insight64, x86’s dominance is the result of a kind of perfect storm in the chip industry.
“No. 1, the computational performance of x86 processors has increased rather dramatically over the last 10 years,” Brookwood says. “No. 2, the software environments have gotten more robust.”
Brookwood says improvements include not just operating systems, but also application-level multiprocessing technologies such as Oracle RAC (Real Application Clusters).
At the chip level, one dramatic improvement has been the introduction of 64-bit addressing, which allows chips to access extremely large amounts of memory. AMD was first to market with 64-bit x86 chips, but Intel soon followed suit with its own EM64T (Extended Memory 64 Technology) architecture.
Dual core came next, with Intel now shipping dual-core Xeon chips to compete with AMD’s earlier dual-core Opterons. As opposed to clustering and grid technologies, which achieve mainframelike performance by linking banks of x86 servers, dual core brings true, RISC-style symmetric multiprocessing to the mainstream by building SMP directly onto the chip die. Similarly, vendors such as Dell and Sun have borrowed other engineering ideas from higher-end systems to build truly enterprise-ready x86 server hardware (see “Engineering sets x86 server products apart”).
Pat Patla, director for server product marketing at AMD, agrees that even the chip-level x86 innovations aren’t really new. “64-bit computing? We certainly didn’t invent that,” he says. “Neither did Intel. 64-bit addressing was simply something that lived on RISC platforms or even other proprietary CISC platforms. Our whole philosophy, our whole strategy with Opteron is to take all of the great features of higher-end systems, ... cherry-pick the best features, and put them in x86.”
Head-to-head with AMD
Indeed, so successful has AMD been at incorporating high-end chip engineering into x86 that for the first time Intel is in the position of playing catch-up on its own platform. Besides being first to market with both 64-bit and dual-core x86 processors, AMD has made other improvements at the chip level.
Unique to AMD’s design is a proprietary architecture called Direct Connect, which uses HyperTransport technology to create an architecture with extremely fast I/O performance between the CPU and the rest of the system. More importantly, Direct Connect allows the two cores of a dual-core chip to communicate with each other more efficiently than Intel’s current designs allow.
“Direct Connect architecture eliminated all the bottlenecks of a front-side bus. We integrated the memory controller directly onto the CPU core and then directly connected the CPU connectivity bus,” AMD’s Patla says.
But Shannon Poulin, enterprise marketing director at Intel, says many of AMD’s advantages are overstated and that Intel’s decision to stick with existing x86 technologies was a conscious one, designed to reduce customer expense.
“If you have an internal memory controller, every time you switch to a new memory architecture you have to re-qualify the processor, and thus the server,” Poulin says. In addition, he says Intel’s new E8500 chip set also addresses the advantage of Direct Connect by providing a separate front-side bus with dedicated bandwidth for each core of a dual-core Xeon processor.
According to Poulin, Intel customers also reap the benefits of its advanced manufacturing. Intel will be first to market later this year with production parts manufactured using a 65-nanometer process. In addition, Intel is now using 12-inch silicon wafers whereas some of its competitors are still using 8-inch wafers, giving Intel an advantage when producing larger, multicore chip dies in volume.
For its upcoming product releases, Intel will close the gap with AMD further by adapting lessons learned from its low-power Pentium M laptop CPUs. A new server chip code-named Woodcrest, due in late 2006, will incorporate dual low-power cores in a design expected to offer five times the performance of the first Xeon chips.
But not everyone takes Intel’s road map completely at face value. According to Graham Lovell, senior director for x64 servers at Sun Microsystems, lingering questions remain as to Intel’s ability to execute on its ambitious plans.
“When you go along to a meeting with AMD, there are usually no surprises in terms of the road map,” Lovell says. “There’s nothing going backwards. It’s not a case of, ‘I know we told you we were going to have this on our road map, but it’s slipped back by a quarter,’ and then you go to next quarter’s meeting, and it’s slipped back by another quarter. We don’t have those meetings with AMD.”
On that score, Insight64’s Brookwood advises a wait-and-see attitude. “Last year Intel had a number of execution problems; that’s 2004. 2005 has been relatively quiet. We’ll see whether 2006 is more like 2005 or more like 2004, but it’s a little soon to say.”
From benchmarks to the bench
The wild card in this deck is AMD’s ongoing litigation against Intel over alleged anti-competitive behavior. The suit, filed in June, accuses Intel of using threats and financial incentives to keep its customers from buying AMD chips.
Adding credence to AMD’s claims is the widespread belief that AMD’s technical advantages should have allowed its products to gain more market share than they have. But some analysts, including Insight64’s Brookwood, believe that other factors may be at play.
“I was talking to someone from AMD last month, and they said their biggest problem is still end-user awareness,” Brookwood says. “And even if the IT manager is aware, is the CEO aware?”
This is not the first time Intel has been accused of being a monopolist. AMD filed its first antitrust complaint against the chipmaker in 1991. That suit was resolved in a 1995 settlement that formally cleared the decks of all litigation between the two companies.
For its part, Intel admits to no wrongdoing. In a formal response to the suit issued in September, Intel blamed AMD’s woes on the company’s own business missteps, including failing to ramp up its manufacturing facilities to meet demand, and “AMD’s goal of shielding AMD from price competition”.
What is certain is that the industry cannot ignore the AMD suit. Many analysts speculate that the case will draw out for years. The full extent of the impact this will have on the x86 hardware market, however, is difficult to predict.
A niche for RISC
Litigation notwithstanding, x86 is likely to continue to dominate the volume server market for the foreseeable future. Yet AMD’s Patla says there’s still plenty of good engineering in RISC designs such as IBM’s Power architecture. And unlike AMD, Intel isn’t putting all its eggs in the x86 basket, either (see “Itanium: Not dead yet”).
Because current clustered environments don’t scale well beyond 16 or 32 CPUs, certain applications will always benefit more from the unique engineering of RISC architectures. Sun’s upcoming Niagara Sparc processor, for example, will support multiple processing threads inside each of its eight cores.
“I think it really depends on customer needs,” says Fadi Azhari, director of outbound marketing for the scalable systems group at Sun. “It could be a number of questions: operating environment, your existing infrastructure, the scale you need within a single image of the system.” There are many situations, he says, where large, RISC-based SMP systems may still be your best bet.
Insight64’s Brookwood agrees. “If you look at all server systems, you’ve got 95 percent being x86, 5 percent being ‘other.’ Is it going to get to 100 percent? No, for the same reason the guys using Apple aren’t going to throw in the towel and use PCs,” he says. “There are people using [RISC chips] where you’ll have to pry them out of their cold, dead hands.”