AMD, Nvidia take graphics duel to TFlops

Archrivals Nvidia Corp. and Advanced Micro Devices Inc. have announced their competing next-generation graphics processors Monday (June 16), both claiming their massive multicore chips can deliver a TFlops for graphics and video as well as high-end technical and scientific applications.

But that's about where agreement ends. The two architectures take radically different approaches to multicore, memory, silicon integration, process technology and software.

The new chips come at a time when Intel is preparing its own next-generation architecture dubbed Larrabee, based on many streamlined x86 cores. At an Intel Research event last week (June 11), Intel chief technology officer Justin Rattner said Larrabee represents the start of a long term shift to a new graphics architecture based on ray tracing instead of the traditional graphics rasterization approach, a claim analysts and AMD rejected.

Going mainstream
In perhaps the most dramatic difference with past practices, AMD will break with the tradition of delivering a high-end part then evolving lower-end devices from it. Instead it will launch a part aimed at the graphics mainstream with a proprietary interconnect that links two chips on a board to deliver top performance.

"We didn't want to come out with one monolithic GPU and then disable parts of it for different markets," said an AMD spokesman prior to a full disclosure of the part in a briefing in San Francisco June 16.

The strategy makes sense for the financially troubled AMD, which also has laid out conservative roadmaps for its computer processors. The graphics choice reduces costs and risks while maximizing returns for the company, which has suffered through multiple loss-making quarters.

The decision to use a two-chip strategy for the high end was made more than two years ago, based on an analysis of yields and scalability. It was not related to AMD's recent financial woes, said Rick Bergman, general manager of AMD's graphics division.

"I predict our competitor will go down the same path for its next GPU once they see this," Bergman said. "They have made their last monolithic GPU."

'Good approach'
"On paper, the AMD approach looks good," said Jon Peddie, principal of graphics market watcher Jon Peddie Research. "If it works, it will be a significant shift in how GPUs are made, but we won't know until later this year" when customers can test the new parts, he said.

AMD says its 4850 device at about 110W and $199 will deliver about 75 percent of the performance of Nvidia's high-end GTX280 which costs $649 and dissipates 236W. Two of the AMD parts on a board will hit graphics benchmarks about 30 percent higher than the Nvidia device, the AMD spokesman added.

A 4870 version of the product sporting slightly higher performance will cost $299. Both AMD chips are made in a 55nm process, compared to 65nm for the Nvidia chips, and measure about 16mm x 16mm compared to about 24mm x 24mm for the Nvidia part. The approach gives Nvidia bragging rights for the single most powerful graphics processor, an edge that plays well to the gaming and technical communities who use the parts.

AMD will claim technology leadership in two areas. Its chips will use more than 500 cores, more than double the 240 cores on the new Nvidia parts. They will also use GDDR5 memory interfaces running at about 3.2Gbit/s or more. Nvidia will use the existing GDDR3 protocol running at up to 1.1GHz on a 512bit interface to deliver memory bandwidth up to about 102GBps on some versions.

Nvidia's strategy
Because memory bandwidth is not increasing as fast as processing resources on the latest Nvidia architecture, some applications may find they are memory bound, said Andy Keane, general manager of GPU computing at Nvidia.

Bergman said the AMD focus on a more mainstream design will enable it to roll out this fall a version for notebooks that consumes less than 70W. "There's no way this new Nvidia core will be in notebooks this fall," Bergman said.

The Nvidia chips include 1.4 billion transistors running at speeds up to 1.6GHz. "It's one of the largest chips in mass production and unlike some processors, it is not 60 percent cache," said Keane.

Nvidia said its latest chips can handle as many as 30,000 threads at a time, up from about 10,000 threads for its previous generation. It upgraded its latest chips to support double precision floating point math, something AMD already supports.

As with previous generations, the performance claims and definitions for the rival parts are not always clear. "There's a lot of specsmanship," said analyst Peddie.

However, Peddie gives Nvidia kudos for aggressively pursuing the emerging market for technical computing on graphics chips. The company is launching two board level products for such high-end apps, including a four GPU system in a 1U-sized rack-mounted device that delivers up to 4TFlops at 700W. It sells for $7,995.

Parallellism
In addition, Nvidia is updating its year-old Cuda development environment that helps programmers modify C programs to exploit parallelism. Cuda now supports all major 32bit and 64bit OS and includes a number of parallel algorithms and tools.On the roadmap for Cuda is support for Fortran and multiple GPU systems and clusters. Nvidia is also working on support for C++ and a hardware debugger.

"We have pulled in some basic C++ features but not more advanced ones like exception handling", said Ian Buck who heads up the Cuda development team. The hardware debugger presents "a challenge in figuring out a way to represent a system that might have thousands of threads running on it," he added.

For its part, AMD leans on third parties for some of the hardware and open source software to drive its graphics chips into technical computing apps. It is backing an emerging language called OpenCL as the programming environment for its technical computing initiative.

AMD has a compiler and tools for OpenCL in the works. "You can expect full support for the language in 2009," said Bergman.

Both companies are reaping the benefits of a move in recent years to programmable cores for graphics processing, resources they are leveraging to attack new markets in technical computing that range from astrophysics to oil and gas exploration.

Long road to ray tracing
Both Bergman and Peddie said even the new graphics processors are not up to the heavy processing demands of ray tracing, calling into question the viability of Intel's suggestion the industry is near a shift to the approach.

"The suggestion Intel can run ray tracing well with a 16 or even 80 core Larrabee chip is absurd," said Peddie, suggesting even the dozens of cores on today's AMD and Nvidia parts are not yet fully up to task. "Ray tracing is overhyped, and it's not going to happen in the next two to four years," Peddie said.

"Don't bet the farm on ray tracing changing the market because it is not going to happen in the next 2-5 years," agreed Bergman. "Ray Tracing is appropriate for some very niche markets where interactivity in real time is not a concern, but the vast majority of the graphics market is in games and video," he added.

- Rick Merritt
EE Times