Fred Weber, CTO of AMD, is a man on a mission. Over the last several years he has helped to lead AMD out of the wilderness and into the promised land of 64-bits and multiple cores. I caught up with Weber last week at the Future in Review conference, and asked him about AMD's product direction, focus on spreading x86 across product categories and recent processor introductions. His remarks follow:
Geode LX800 (An energy-efficient processor for ultraportable, set-tops, TVs and handhelds, with a built-in encryption engine and 2D graphics, and optional 3D graphics.)
The LX800 is first Geode processor since the team was acquired from National Semiconductor. It extends the line to higher performance and lower power, and potentially lower cost. So far, 17 major beta customers --names that one will recognize--will announce product for IP set tops, printers, ultra-thin wall-mount displays, thin clients and other devices over next three to six months. It brings a unique level of power at one-watt power consumption. Transmeta was probably the closest competitor but future availability [of Transmeta chips] seems extremely limited. Our two-chip solution creates the whole PC, such as an ultra portable, other than connectors to standard I/O devices. In retrospect, it's the first down-payment on our "x86 everywhere," bringing x86 to areas like consumer electronics and storage.
Speed and multicore architecture
From one point of view we have to be cautious--multiprocessing is hard. All things being equal, you would always rather have once processor twice as fast rather than two processors at the same speed. That's a no-brainer. For the last 20 years, we have been able to add more pipelining, caching, branch prediction, and super scaling to increase performance as you have more transistors. It's like building ever taller skyscrapers—at some point doubling the height is problematic. You reach a point where if you put on twice as many transistors, you get may 1.2X performance, not 2X. If you use the transistor budget to put two processors down, you don't get 2X performance because of the inefficiencies in a two-processor system. You could get 1.6X performance improvement, which is better than 1.2X, as long as you can match power characteristics.
For those reasons of architecture—nothing to do with Moore's Law—we have reached a point where multiprocessing makes sense. The problem space is the software trying to tackle searching, video transcoding and parallelizing. There are just enough applications to get over the hump now, beyond the chicken-and-egg problem. As AMD and Intel figure out how to make dual-core processors higher performance, then more software will come, and when we come out with quad-cores and then eight, you get a stair-step effect that motivates developers to make their software better. Over the next five to ten years there is a virtual circle that at a measured but relatively fast pace will bring more and more levels of parallelism.
Server and desktop futures
We have a huge upside in the server space, with one-two-three punch. One is the right 64-bit approach, number two is Direct Connect and three is dual core, which is a good six months ahead of the competition and because of Direct Connect [formerly HyperTransport] offers better scalability. We should be able to maintain our leadership for the foreseeable future. Rumors are that Intel will have a Direct Connect in 2006 or 2007, but we are already on our third version. [Editor's note: After two years on the market, AMD's 64-bit Opteron processor has more than 5 percent of the x86 server market worldwide, according to Gartner Dataquest. In North America, AMD's four-way Opteron servers claim about 20 percent market share. Both AMD and Intel are prepping new virtualization technology for running multiple operating systems on a chip.]
The client space is more hard to predict. We have fabulous product, but it’s hard to differentiate. A lot is driven by other factors. We have done well in consumer and retail, but have a tougher time penetrating commercial accounts. There's no reason not to use us--it's more about what are the reasons to move. In the last two years we have established credibility and reliability with the Opteron, removing barriers and concerns so now more [vendors] are tempted to use us. [Editor's note: Both AMD and Intel have launched dual-core desktop processors. Acer, Hewlett-Packard, and Lenovo are among the manufacturers planning products based on the dual core AMD Athlon dual cores. In April Intel debuted its Extreme Edition Pentium 4, and this week is introducing its new dual-core Pentium D processors and supporting chipsets.]
Assessing the Cell processor [Editor's note: The multicore chip developed by Sony, IBM and Toshiba that will run the next version of the PlayStation video game machine.]
It's an interesting direction—a very application specific design that is more programmable than solutions in past. It's a nice move from fixed function to general purpose, but it is still very specialized. It points the way. I am interested in heterogeneous multiprocessing, in which all the processors don't all have the same characteristics, just as in housing with huts, hotels, apartments and houses. Heterogeneous multiprocessing does the same thing. We want to use different processors for different uses. If you are doing something massively parallel, you want them all working, or if you are doing serial work, you want a couple of big processors turned on. The Cell is a beginning in that direction, but it's program model is still too difficult. I am a big believer in a simplified programming model--if you have many x86 processors, porting code is not a problem, it's more of performance tuning issue.