Advanced Micro Devices on Monday unveiled the company's first graphics card with two processors that deliver more than 1 teraflop of graphics performance each. In addition, the company introduced a new version of its general-purpose graphics processor for high-performance computing.

The new discrete card, which doesn't have a brand name yet, was one of three new cards with AMD's fastest graphics-processing unit, or GPU. The cards reflect AMD's strategy of delivering cards for gaming at three separate prices: around $200 for the baseline card, about $300 for midlevel products, and $500-plus for the fastest GPUs. Power consumption would be 110 watts, 150 watts, and 250 watts, respectively.

At the low end of the new cards is the ATI Radeon HD 4850, which is scheduled for availability June 25 and will introduce AMD's first teraflop GPU. The faster, midlevel HD 4870 is scheduled for release July 8.

AMD's previous fastest card, the HD 3870, was 500 gigaflops. "We're now breaking into teraflop territory for $199," Rick Bergman, senior VP and general manager of AMD's graphics products group.

The third card, code-named the R700, will have two HD 4870 GPUs and will deliver the highest performance of any AMD ATI card. The product is scheduled to be available in eight weeks. In terms of performance to price, AMD is lining up the low-end card against Nvidia's 8800GT, 8800GTS, and 9600GT. The midlevel product faces off against Nvidia's 9800GTX, and the top-of-the-line card against the 9800GX2 and the GT200.

For the highest level of performance in a desktop, computer makers can drop in two R700 cards, one in a PCI Express slot each.

AMD also unveiled its second-generation general-purpose GPU. The FireStream 9250, which follows the original product by six months, delivers 1 teraflop of performance, and is made to fit into a PCIe slot in 1U servers or in desktops and workstations.

The FireStream would compete with Nvidia's Tesla line of GPUs. Nvidia on Monday launched its second-generation product, called the Tesla 10P, which also delivers a teraflop of performance. The GPU card for desktops and workstations costs $1,699. Nvidia also offers a four-GPU server for $8,000.

Nvidia provides its own technology called CUDA for developing applications for Tesla. CUDA leverages the C programming language. Development on FireStream requires the use of Open CL, an open source language that's an extension of C/C++, Bergman said. Both vendors provide compilers, debuggers, and other development tools.

The FireStream 9250, available in August, is priced at $999.

AMD and Nvidia are positioning their general-purpose GPUs as cost-effective alternatives to increasing the number of general-purpose CPUs in order to handle highly computational applications used in the oil and gas, financial, and medical industries, as well as in scientific research. General-purpose GPUs are marketed as providing multiples more horsepower for the same amount of energy consumption.

General-purpose GPUs are only about a year old, so it's unclear how well they'll be received in the market. AMD expects the products to start seeing broad adoption in 2009 or 2010, Bergman said.