AMD plans to stretch its capabilities over the next few years, believing that flexibility will be a key part of its designs over the rest of the decade.
AMD on Thursday laid out plans to serve 30 percent of the market within the next two years, with new quad-core processor designs scheduled for 2007 and an acceleration of its manufacturing capabilities.
The company also talked about plans to build future processors with the ability to mix and match the building blocks of a chip to cater to different needs, and to allow its partners to add co-processors that can link directly to Opteron processors through AMD's Hypertransport links.
Executives speaking to analysts and press at the company's headquarters in Sunnyvale, California sought to maintain the momentum AMD has enjoyed over the last three years, gaining market share in important markets and giving Intel fits. "We want to open up our technology and unleash a completely new wave of innovation," said Hector Ruiz, chief executive of AMD.
The chipmaker plans to license its Hypertransport technology to allow customers and third-party chipmakers to build specialised processors and other chips that can connect directly to future Opteron processors, said Marty Seyer, senior vice president at AMD. Hewlett-Packard, Sun Microsystems, IBM and Cray have all agreed to participate in the program, which AMD is calling Torrenza.
Ultimately, AMD wants to separate the building blocks for its chips — such as processing cores, memory controllers, Hypertransport links and cache memory — into distinct parts that can be configured in multiple ways to meet changing workload requirements, said Phil Hester, AMD's chief technology officer. This also will allow customers to plug co-processors built specifically for certain workloads, such as Java or XML (Extensible Markup Language) traffic, right into Opteron chips, he said.
AMD's partners have been able to sign up for noncoherent licences to Hypertransport up until this point, Hester said in an interview after the briefings. The new licences provide a coherent link to the chipset and will allow server users to manage different co-processors with drivers, like PC peripherals, rather than having to use new applications for each co-processor, he said.
The difference between noncoherent links and coherent links is the difference between a co-processor being treated as an adjunct to the system and it being treated like another Opteron processor, said Nathan Brookwood, an analyst at Insight 64. A coherent link allows a specialised high-performance co-processor to access data stored in the cache memory of the Opteron processors. This could prove useful for applications such as cryptography or media processing, he said.
On the manufacturing front, AMD plans to introduce chips based on its 45-nanometre manufacturing technology by the middle of 2008, said Daryl Ostrander, senior vice president for logic technology and manufacturing at the company. That would mean a 1.5-year gap between the introduction of AMD's first 65-nanometre chips later this year and the volume production of 45-nanometre chips, he said. The number attached to the size of the manufacturing technology refers to the average size of features on the chip. Smaller features allow chipmakers to pack more transistors and more performance into their chips.
Intel has been losing market share to AMD in several areas over the last few years, but it has maintained an advantage in introducing new manufacturing technologies ahead of everyone else in the industry. Intel is already shipping a 65-nanometre chip, its Core Duo processor. Some in the chip-manufacturing industry have called for chipmakers to slow their cadence of shrinking transistors to every three years as the challenges become more daunting, but Intel has stuck to a two-year schedule. However, AMD's plan is to move from 65 nanometres to 45 nanometres in 18 months, which will allow it to chip away at Intel's advantage, Ostrander said.
Later in the day, Ostrander said that the 18-month turnaround for 45 nanometres was set because of AMD's confidence in that generation of its technology, but wouldn't necessarily serve as the cadence for future rollouts. Next on the agenda is 32 nanometres, which AMD will hopefully introduce 18 to 24 months after the 45-nanometre chips arrive in production volumes, he said.
By 2008, AMD will be ready to introduce something called Direct Connect 2.0 for server processors, Hester said. Direct Connect is the name AMD uses for its chip designs, which use an integrated memory controller to link directly to memory and Hypertransport links to connect to other processors or a system's I/O (input/output) controller. Details about Direct Connect 2.0 were not immediately available.
The company plans to introduce quad-core processors for servers and desktops in mid-2007, as it outlined at the recent Spring Processor Forum, and mobile chips based on a new power-optimised architecture. Hester revealed a few extra details about the quad-core chips on Thursday, disclosing that each core in a mid-2007 server processor will have either four 16-bit Hypertransport links or...
