Intel and Nvidia are locked in a war of words over processor speed. The battle appears to have been triggered by a paper Intel engineers presented last week at the International Symposium on Computer Architecture in France.
That paper, entitled “Debunking the 100x GPU vs CPU Myth: An Evaluation of Throughput Computing on CPU and GPU,” asserts that graphics processing units, or GPUs are, in essence, only 2.5 to 14 times faster than central processing units, or CPUs. Nvidia makes GPUs; Intel’s focus is CPUs.
That kicked off a blog post by Andy Keane, Nvidia’s general manager of GPU computing, claiming Intel’s example uses old equipment and was unclear about what sort of code was used in the test.
Nyah, Nyah, GPUs Aren’t That Fast
Many of today’s throughput computing kernels have “an ample amount of parallelism,” and that makes them suitable for today’s multi-core CPUs and GPUs, the Intel paper states. In recent years, reports have claimed that GPUs work 10 to 1,000 times faster than CPUs on these kernels, the paper’s authors said.
However, their performance analysis turned up different results. After applying optimizations appropriate for both CPUs and GPUs, the paper’s authors found the performance gap between an Nvidia GTX280 processor and the Intel Core i7 processor narrows to a factor of only 2.5 on average.
In his rebuttal, Keane said the testers used the Nvidia GeForce GTX 280, which is a previous-generation GPU.
“The GeForce GTX 285 was available, so by using the GeForce GTX 280, they still did not use the fastest available Nvidia hardware,” Keane told TechNewsWorld.
“Comparisons were made using the best available chips at the time of writing,” Intel spokesperson Nick Knupffer told TechNewsWorld. The paper had to be submitted in November 2009, and both Intel and Nvidia have launched new products since then, he pointed out. Further, the code was optimized and the details were presented in the paper, Knupffer said.
“No one knows what code was run or what data was used during the test,” Nvidia’s Keane added. “Intel didn’t release the source codes to either GPU or CPU kernels. The selection of what code and what data are used can be a significant factor.”
Picking the Right Frame
“Throughout this paper, Intel is trying to frame this as a CPU versus GPU debate,” Keane contended. “In fact, the developer community is moving towards heterogeneous computing, where a GPU is used to accelerate the application as a coprocessor to the CPU.”
The CPU runs the operating system, access to the input/output system, and other parts of an application that are sequential and latency-bound, while the GPU accelerates the compute- and data-intensive parts, Keane explained.
“Nvidia strongly believes that both the GPU and CPU are necessary for a high-performance computing system,” Keane stated.
“General-purpose processors such as the Intel Core i7 or the Intel Xeon are the best choice for the vast majority of applications,” Intel’s Knupffer pointed out. “While it’s possible to program a graphics processor to compute on non-graphics workloads, optimal performance is typically achieved only with a high amount of hand optimization.”
That optimization requires graphics languages similar to DirectX or OpenGL shader programs or non-industry-standard languages, Knuppfer said.
Flailing at Shadows?
In other words, Knupffer and Keane seem to agree that the CPU and GPU perform different tasks. Why, then, are Intel and Nvidia lashing out at each other?
“The fact that the GPU and CPU guys are going at each other just says that they’re both looking for new applications for their technology,” Carl Howe, director of anywhere research at the Yankee Group, told TechNewsWorld.
“They traditionally don’t compete,” Howe said. “The fact that the GPU and the CPU guys think they’re in a competition for garden-variety software code just says that, maybe, their current markets aren’t doing all that well.”
Launching a Preemptive Strike?
Perhaps Intel’s trying to set the stage in its favor in anticipation of an impending settlement of antitrust charges brought against it by the Federal Trade Commission (FTC). These focus on Intel’s alleged anticompetitive practices against AMD and Nvidia. The AMD part of the suit has already been settled.
The FTC’s charges follow on from Intel’s bid in 2009 to exclude its latest chipsets, such as the Nehalem, from a 2004 licensing deal with Nvidia. Under this deal, Nvidia developed chipsets for Intel processors in exchange for Intel’s licensing some of its graphics-processing patents.
That new approach hit Nvidia hard because its graphics card business is cyclical, and in 2009, chipsets were the healthiest part of its business. In October of 2009, Nvidia stopped developing chipsets that work with Intel’s microprocessors. In March of 2010, Nvidia filed a countersuit against Intel. This seeks to terminate Intel’s license to Nvidia’s patents for graphics processing and three-dimensional computing technologies.
Earlier this week, Intel and the FTC suspended trial proceedings to work out a settlement. If the lawsuit is settled, Nvidia may win the right to resume making chipsets for Intel’s microprocessors. That, perhaps, is what Nvidia is hoping for.
“We don’t yet know the details behind the FTC’s announcement, so it’s premature for us to comment,” Nvidia’s Keane said. “We remain hopeful that any settlement will recognize Intel’s history of impeding competition and innovation at the expense of consumers worldwide.”