Larrabee "like a GPU from 2006"
NVISION: Nvidia makes clear it's not afraid of Intel's first foray into discrete graphics
Nvidia has delivered a scathing criticism of Intel's Larrabee, dismissing the multi-core CPU/GPU as wishful thinking - while admitting it needs to catch up with AMD's current Radeon graphics cards.
Andy Keane, general manager of the company's GPU computing group, spoke to reporters at the company's headquarters in Santa Clara, California on Friday, ahead of the opening of the company's NVISION expo.
"There's an incredible amount about Larrabee that's undefined," explained Keane, commenting on the specifications so far released. "You can't just say 'it's x86 so it's going to solve the massively parallel computing problem.'"
"Look at the PC," he continued. "With an OS they don't control, and applications coming from everywhere... to say arbitrarily that everything's going to scale to 32 cores seems to me to be a bit of a stretch. "
Keane dismissed the idea that the Intel Parallel Studio, announced last week at IDF, might solve the problem. "Multi-processing is a hard problem in computer science. It's been there for 30 years. It's not answered by software tools."
Unrealistic performance projections
John Montrym, chief architect for the company's GT200 core, raised further doubts about Larrabee's real-world performance, brushing aside Intel's announcements as marketing puff.
"They've put out a certain amount of technical disclosure in the past five weeks," he noted, "but although they make Larrabee sound like it's a fundamentally better approach, it sn't. They don't tell you the assumptions they made. They talk about scaling, but they disregard memory bandwidth. They make it sound good, but we say, you neglected half a dozen things."
"Every GPU we make, we always consider this type of design, we do a reasoned analysis, and we always conclude no. That's why we haven't built that type of machine."
"Intel is not a stupid company," he conceded. "They've put a lot of people behind this, so clearly they believe it's viable. But the products on our roadmap are competitive to this thing as they've painted it. And the reality is going to fall short of the optimistic way they've painted it."
"As [blogger and CPU architect] Peter Glaskowsky said, the 'large' Larrabee in 2010 will have roughly the same performance as a 2006 GPU from Nvidia or ATI."
Fighting back against ATI
Montrym then turned to the company's most direct competitor, ATI, the graphical division of AMD whose RV770 GPU has put Nvidia's GT200 based cards on the back foot.
"We underestimated ATI with respect to their product," he admitted. "We've looked very closely at this, and we know there are certain things we can do better. There will be improvements to things from all angles: there are some easy fixes in the software domain that will soon be forthcoming. Believe me, it's a very prime focus of ours."
Montrym ascribed the company's current embarrassment, at ATI's hands, to its earlier successes. "ATI has had the benefit for a long time of seeing Nvidia's products and having something to shoot for," he argued, "while Nvidia has not had the benefit of having someone to be shooting after."
He also predicted that ATI would regret its focus on raw graphical power at the expense of more general-purpose capabilities.
"ATI did not spend on things like PhysX and CUDA. But we believe that people value things beyond graphics. If you compare only on graphics, that's a relative disadvantage to us, but the notion of what you measure a GPU on will change and evolve," he argued.