Advanced Micro Devices Monday unveiled a new graphics chip designed to be less power hungry than its predecessors by allowing it to be doubled up on high-end graphics cards.
The chipmaker is slated to introduce a new single-chip graphics card - the ATI Radeon HD 4850 -- based on the new chip on June 25. Then in August, AMD is expected to release a graphics card -- codenamed R700 -- that will include two of the new chips.
"We're [at] a turning point in the way we design our graphics chips," said Matt Skinner, a spokesman for AMD. "As they get bigger and bigger, they use more power and we're coming up on power constraints, as well as how many transistors you can fit on a certain dye size.
"For the last several years, we've designed large monolithic chips for the enthusiast or high-end market," he added. He said that as new technology is added to the monolithic chips, they require "more and more power, which makes it impractical for gamers."
To deal with these issues, Skinner said AMD engineers designed one chip for users in the performance segment, which is a step below the enthusiast market. For high-end game enthusiasts, they'll simply put two chips on one graphics card.
That way, the new cards use less energy hungry and users don't have to wait six months or so for a new lower-end chip to come out.
The HD 4800 family of chips is designed to be the first teraflop-level graphics cards, according to Skinner. "In the last generation of GPUs, we got to half a teraflop of performance running 110 watts at peak," he added. "We're getting a teraflop of performance in the new graphics cards with 110 watts. It's double the performance for the same wattage."
He noted it was AMD's biggest single-chip performance increase in seven years.
Dan Olds, principal analyst with the Gabriel Consulting Group, said designing a chip that can be used alone or with another chip is a smart move for AMD.
"GPU guys have to deal with the same laws of physics that confront CPU guys," Olds said. "You can only get so small and cram so many things into a reasonably sized package. In other words, you've got a limited set of silicon real estate and have to figure out how to best utilize it. Turns out that multiple computational units give the best overall performance, while not overheating. The choice is to either speed up one big chip, or have two chips running at regular speed. Two chips win right now."