Nanotech - The Circuits Blog

Advanced Micro Devices' bet on a new approach to graphics chip design appears to be paying off, according to analyst Jon Peddie. This could put AMD's ATI graphics chip unit on top again--or at least on equal footing with Nvidia, the graphics leader over the last few years.

Peddie heads Tiburon, Calif.-based Jon Peddie Research, which specializes in graphics chip market research.

Test reports on AMD's and Nvidia's newest graphics chips are pouring in. Both companies are racking up good scores. (See Diamond Viper Radeon HD 4850 review here.)

But beyond the day-to-day test scores, AMD's ATI graphics chip unit may be winning the longer strategic battle. ATI has gambled on a radically different strategy for its latest series of chips--the HD 4850, HD 4870, and upcoming dual-chip R700.

"(AMD) is starting in the middle of the market and scaling up. That's a break with tradition," said Peddie. "We always started at the very, very tippy-top and build the most powerful thing you could and then let it scale down over time."

But Peddie said this traditional approach just isn't practical anymore. "The chipsets keep getting larger and larger despite the fact that we were going to smaller and smaller (manufacturing) process nodes. The chips grew faster than the process nodes shrunk and the consequences of that is that the power consumption went up, the costs went up, and it got to the point where it's kind of impractical to continue along that way," he said.

In essence, AMD's ATI unit strategy is to build smaller, less power-hungry chips and then gang them up to get better performance. Nvidia's strategy has been to build one large, extremely fast--and extremely power hungry--chip.

For ATI, the execution of this chip-ganging strategy is the key. And this is where ATI appears to have been successful. "The inter-processor communications. Getting that to work has been the trick. This is what ATI has done. They've come up with this stellar way of doing inter-processor communications so they can in fact get the scaling," according to Peddie.

And there's more than meets the eye. ATI has also cut in half the number of bits in the memory interface, Peddie said: down to 256 bits while Nvidia has remained at 512. "That has the benefit to ATI of reducing a big hunk of the power consumption."

Peddie said in the past this kind of approach would have been suicidal because it would have decimated ATI's test scores. "The argument against this is that graphics performance is a function of memory," he said. "Typically you want wider and wider (bit width)."

But ATI has countered this by using the fastest memory standard available. "So to compensate for shrinking down the bit width, ATI has jumped to the next-generation in memory design called GDDR5. GDDR5 is approximately three times faster than GDDR3--which is what Nvidia is still using and what ATI uses on their smaller cards," according to Peddie.

"So with three times the speed but half the width, they end up with 1.5 times the processing capability with the memory."

"A very clever thing that they did but mind you it was a gamble that looks like it's going to pay off," he said.

ATI has more processing units than Nvidia inside its chip too. "The other thing is that ATI has 800 processors in their chip and Nvidia has 240. That has a processor count advantage," Peddie said.

Though it remains to be seen if this advantage is borne out in testing over time, he added. "Nvidia and ATI keep improving their drivers so they'll seesaw back and forth with their scores, almost from week to week."

But in the long run, Nvidia may be forced to adopt ATI's strategy to keep pace in these week-to-week battles. "If ATI is successful, as we expect that to be, then Nvidia will have no choice but to adopt (ATI's) approach, just out of practicality," Peddie said. "It just makes a whole lot of sense."

AMD-ATI's upcoming R700 (rumored to be called the 4870 X2) two-chip graphics board will be the ultimate test of this strategy.

"It's a new proprietary inter-processor communication technology. If they put these two chips on one board and it does scale properly, then they have pulled off a coup," he said.

"When you gang up graphics chips (using the traditional Scalable Link Interface or CrossFire technologies) they roll off pretty fast. ("Roll off" implies that performance doesn't scale up well.) "So when you put two boards in, you don't get twice the performance but you (only) get one and a half. You put four boards in and you (only) get about 1.7, 1.8. What ATI is saying is that with two chips using (their) proprietary inter-bus, they will get 1.8 (the performance) with two chips. If that's true, you can expect to see four of them giving you something around 2.5."

Getting 2.5 times the performance from four boards would be a masterstroke for ATI.

The previous ATI dual-chip solution was very different, Peddie said. "The HD 3870 X2 was not a proprietary bus but a CrossFire connection. The CrossFire connection and the SLI connection are at the very, very end of the pipeline. Not the most efficient place to do an inter-processor communication. That's one of the reasons ATI has abandoned it."

AMD's ATI unit is also better positioned than it was before in manufacturing. "Part of the reason that Nvidia has been ahead is that ATI has been suffering over the last three or four years with manufacturing problems. It's not that ATI didn't have a good chip, the problem was that ATI couldn't build enough of them."

This should change with the newest series of chips. "This (design)--so they say--will really go into high-volume production." Though he cautioned this still remains to be seen.

"The (new ATI chip) is a really efficient, tight design. They used to do this all the time but they kind of got off that trail. And now they're back on it."

UPDATE: On Monday, Advanced Micro Devices and Nvidia are launching graphics chips based on distinctly different design philosophies.

Nvidia chips are big and powerful

(Credit: Nvidia)

Nvidia's GTX 280 and GTX 260 are designed to deliver the biggest performance bang per chip. A so-called "monolithic" approach packs 1.4 billion transistors and 240 processing cores onto one piece of silicon.

(See Peter Glaskowsky's review of the GTX 280.)

AMD's modular approach tends toward less is more: smaller, less power-hungry chips that can be strung together to achieve higher performance. The company plans to implement this strategy with the HD 4850 and HD 4870 graphics processing units (GPUs) that are being introduced on Monday. (See "Notes" below.)

This design philosophy is based on the same multicore strategy now employed by AMD (and Intel) for the CPU, the main PC processor. Instead of making one chip bigger and faster, AMD boosts performance by stringing together multiple chips. This approach also allows a chipmaker to avoid the time and expense of designing separate processors for the midrange and very-high-end segments.

(Clarification: In the case of a CPU, a core refers to a "scalar" x86 processor core found, for example, in a quad-core Intel or AMD chip. The GPU core is a "parallel processing core." GPUs these days typically contain hundreds of these cores.)

"The beauty of this design is that it's scalable. You can put one or two (chips) on a board," said Matt Skynner, vice president of marketing at AMD's Graphics Products Group.

In the midrange segment ($200 to $300) AMD uses a single chip--for example, an HD 4870. At the high-end ($500 and above), it adds another chip to scale up to better performance. This dual-chip design--code-named the R700--will be marketed as the 4870 X2.

On the other hand, Nvidia says its emphasis on a single, very-high-performance chip is necessary to keep it out front. "At the high end, there is no prize for second place," Ujesh Desai, general manager for GeForce products at Nvidia, said in an interview with Nanotech: The Circuits Blog last month.

(Nvidia also offers multichip designs, but it puts each chip on a separate board, while AMD puts two chips on one board.)

And Nvidia is trying to raise the bar with GTX 280. "We're rendering about 3 million triangles per frame," Curtis Beason, an engineer at Nvidia, said last month at an event where Nvidia previewed the GTX 280 chip.

AMD targets smaller chips that can be strung together to get better performance. AMD chip on left, Nvidia chip on right.

(Credit: AMD)

"With (the previous-generation) GeForce 8800, what we achieved is a very photorealistic character. Very detailed skin. But it was a single character," Jason Paul, the GeForce product manager, said at the Nvidia event last month. "With GTX 200 what we're moving to is multiple highly realistic characters."

Nvidia is also boasting that a dual-core GTX 280 can convert a high-definition movie into iPod video format in 35 minutes, compared to about five hours for a quad-core CPU system with low-end integrated graphics.

Hewlett-Packard's Voodoo unit will be one of the first to adopt the new Nvidia GPUs. "We are excited to be one of the first companies in the world to offer the technology in the new Exhilaration Edition of the...HP Blackbird 002," said Rahul Sood, chief technology officer, HP Voodoo Business Unit.

Notes: updated 6/16, 12:10 PM:

--AMD 4800 series processors will be available starting next week at Besy Buy, according to AMD VP Rick Bergman, speaking Monday at an AMD event. "In just a little over a week from today. You'll be able to walk into a Best Buy and buy this chip (4800 series) on a graphics board for about $200. A teraflop for $200," Bergman said. He added that systems will also be available from Falcon Northwest, Velocity Micro, and ibuypower. "We're also introducing a system that can take four of these boards," he said. "That's almost five teraflops of performance in a personal computer."

--Both AMD and Nvidia say their GPUs can achieve about one teraflops (trillion floating point operations per second) of performance.

--Nvidia's GeForce GTX 280 will retail for $649 and be available on graphics boards starting Tuesday. The GeForce GTX 260 will be priced at $399, with availability slated for June 26.