Intel's Sandy Bridge welcomes discrete graphics
Room for both Formula One and Toyota Prius
IDF The on-chip graphics of Intel's Sandy Bridge processor may be measurably ahead of Chipzilla's previous integrated graphics, but it's not intended to replace discrete graphics for high-end users and dedicated gamers.
"I don't see high-end discrete graphics cards going away," said Tom Piazza, Intel Fellow and graphics-architecture guru, speaking on Monday at the company's annual developer forum, "nor do I see Formula One race cars going away just because we built Priuses."
He also reminded his audience that the Sandy Bridge architecture has PCIe support on-chip — 1x16 or 2x8 — so support for the addition of discrete-graphics cards or GPUs is up to an OEM's decision concerning the inclusion of a PCI slot or on-board bus.
Piazza also attempted to deflect questions about Sandy Bridge's lack of DirectX 11 support. "There are no exclusive DirectX 11 games out today," he answered a questioner. "In fact, most [game developers] have actually skipped over DirectX 10, and most of the games fall all the way back to DirectX 9. I don't see the issue right now specifically about DirectX 11. DirectX 11 is, I'll just say, 'around the corner' — and on the Intel products as well."
When pressed about the ability of Sandy Bridge's integrated graphics to handle over-the-top gaming demands, he was honest about its limitations. "For the games that are designed for the highest-end, extreme-edition graphics, the answer is 'probably not,' unless you de-feature and run [them] on a smaller screen."
That said, he defended his baby. "But most of the mainstream games that are out there, I think the answer is going to be 'yes'."
Piazza said that in 2007, Intel CEO Paul Otellini asked if he could commit to a "10x in '10" promise — that is, an integrated graphics–performance improvement by 2010 of ten times over what was then available. Piazza said yes, Otellini announced that promise to the public, and Piazza and his team, he now claims, has upped that improvement to 25, rather than 10, times.
Not to put too fine a point on it, but claiming a 25X improvement over Intel's decidedly sub-optimal integrated-graphics performance of 2007 is not unlike comparing Intel's current financial performance to that which it was suffering during the darkest days of the global economic Meltdown.
That said, Piazza and his team took some ambitious risks to achieve that performance growth. One of the methods used to juice graphics performance was to fully exploit fixed-function processing units rather than force an over-reliance on programmable shaders — or, to use Intel parlance, execution units, or EUs.
He and his team decided to reverse course on programmability versus fixed-function units. "'Let's do pretty much a complete architectural break from what we were doing before'," he said he and his peep proposed. "And this is a significant change from previous generations."
The difference, he explained, was that "What we did here was — a simple theme — where before we would have said 'Can you program this piece of the pipeline?' Now we said 'Do you need to program this piece of the pipeline?' And if you don't need to program this piece of the pipeline, you push it back into a fixed function."
Piazza — who apparently prefers to disregard the memo from Intel's marketing department that a shader should be called an EU — summed up his team's change in focus by saying: "Let the shaders do the shader work, and nothing more."
He also argued — perhaps in an attempt to preempt attacks on the relatively low number of
shaders EUs in a Sandy Bridge graphics subsystem — that because of such enhancements as new instructions that allow for multiple operations per clock, merely counting EUs isn't a realistic way of measuring graphics performance. "It's not how many shader units we have in a system, it's how many yielded FLOPS we can get out of the shader unit."
The Reg would enjoy seeing that aphorism in needlepoint.
"This has 12 shader units," he continued, referring to the iteration of Sandy Bridge he was discussing, "compared to three years ago [when] we had eight shader units. Don't count shader units. These are totally different animals."
Thanks to the use of true fixed-function units rather than the emulation of fixed functions in shader units, along with additional enhancements in the shaders themselves, Piazza claims a doubling of per-shader performance: "So this 12 looks like a lot more than the eight we had before."
Hmm... Easy math: doubled performance, twelve shaders — that'd be an equivalent of 24 versus the previous eight.
But wait! There's more! "If we were doing a lot of transcendentals," Piazza claimed, "[it'd be] significantly more than that."
When Sandy Bridge systems begin to appear in early 2011, we'll be deeply curious to find out just how transcendental their graphics subsystems are — both in the mathematical and marketing senses. ®
Sponsored: Hyper-scale data management