What's Next After the CPU?
转自：http://www.reed-electronics.com/electronicnews/article/CA474378主题：What's Next After the CPU?
概述：how distributed computing is changing the role of the Central Processing Unit (CPU) in computing
出自：ElectronicNews (http://www.businessweek.com/)
发表时间：October 22, 2004
内容：
What’s Next After the CPU?
By Ed Sperling -- Electronic News, 10/22/2004
Electronic News sat down to discuss the future of the microprocessor with Mark Aldering, VP of IP cores and the embedded processing division of Xilinx; Jeff Jussel, VP of marketing at Celoxica; Drew Wingard, CTO at Sonics; Thomas Peterson, director of product marketing at MIPS Technologies; and Malachy Devlin, CTO of Nallatech. What follows are excerpts of that conversation.
Electronic News: As we move into a world of distributed computing and sometimes ubiquitous computing, what exactly is a microprocessor and how has it changed?
Devlin: It’s any device that takes the data in, processes it and spits it back out.
Peterson: What may be more unique about processors is that they’re standardized software infrastructures that run on top of hardware. It’s the idea that there’s a broad base of software available that takes advantage of the structure.
Aldering: In Xilinx’s case, the hardware is reconfigurable, so there’s less of an issue with standardized software.
Wingard: I think we need to separate the microprocessor and processing. With the classic microprocessor, the hardware resources are visible to the people writing the software, and while there is some abstraction, the translation is relatively well understood and transparent. That means you can drop in assembly code to make sure your system runs right even though everyone tells us we shouldn’t do that.
Electronic News: But getting back to the issue here, isn’t there a shift away from a central processing unit?
Peterson: I think you have to talk about why are things changing. There is change going on. But what are the forces behind it? The key is integration. You can look at a digital cable set-top box and see one piece of silicon. It’s a processor and it integrates some signal processing hardware. That trend will accelerate toward micro-architectural platforms, which will include the processing unit.
Jussel: We’re seeing a move away from the silicon architecture as the point of value in a chip toward the algorithm as the point of value. Whether you can do a function better, faster and cheaper than the other guy is where the value is now. That’s a move away from a fixed processor. Yes, there will be a CPU in there. There will be reconfigurable logic and custom co-processors that are there to accelerate portions of the algorithm.
Aldering: The belief that we are always going to get faster, more capable CPUs every six months has been shattered. The twin-stage pipeline running at 10 GHz doesn’t look like it’s going to happen. You have to look at cost and things like partitioning the problem into a dual-core, a custom co-processing fabric and dedicated logic. The new Cisco router has 176 CPUs on a chip. It’s all about the best way to solve the problem.
Electronic News: What does that mean for chip designers?
Wingard: We are moving away from processor-dominated designs. In the early world of the SoC, people were surprised you could put a CPU on an ASIC. We’re now at the point where it’s trivial to implement multiple CPUs from a process technology perspective. From a design complexity management perspective, it’s far from trivial. The high volume applications are almost all embedded processors-digital set-top boxes, phones, MP3 players. The world of embedded systems programmers is primitive compared with the people who program supercomputers. No one has come up with an abstraction that allows you to manage more than one program at a time. The only approach that we have come up with that makes sense is to partition the design into multiple CPUs.
Peterson: You have to go back to the market drivers for these technologies. In our space, which is consumer electronics-and that market is different from other markets -- cost is key. We have to accomplish some specific, predetermined algorithmic function. Performance is useful in that, but if the cost of that performance is more complexity and more infrastructure it’s probably a net loss.
Aldering: In an ASIC or a structured ASIC, you can get the performance but the bar continues to go up and up and up.
Electronic News: How is all of this affected by system-level design?
Devlin: At the system level, this is nothing new. It’s been the same problem for the past 20 years. We use multiple processors because one can’t handle it. That’s true even with the new chip that looks like a Formula One racer. It doesn’t take long before it ends up in a Ford because no one can handle it all.
Peterson: Where is the target from a performance perspective? In raw processing, you have the same feeds and you can make it run faster. In the consumer space, the loads are different and you have to balance out what you can with software.
Electronic News: Looking at all the changes we’ve discussed, who will lead this market? Will it still be Intel or will it be other players?
Wingard: Intel’s attempts to move into the communications field have been surprisingly ineffective. In that market, you can’t just say you have a processor. The communications and the consumer market are more complex. This is so much more complex than it was several years ago because when we integrate something we have this extra constraint. When we get to the point when there’s a significant amount of data storage, no one is interested in buying a single chip when there are six extra banks of DRAM on it.
Aldering: To a hammer, everything looks like a nail. The uniprocessor at 20GHz doesn’t have a bright future at this time. Whether you partition into multiple processors or partition in multiple gates, the cost of gates is getting cheaper and cheaper.
Peterson: The real winners are going to be the integrators. The folks who can take an integrated digital TV and compete with a turnkey solution will include a CPU that can adjust to migrating standards. Companies like Freescale, with its integrated general processors, and PMC will continue to win.
Aldering: But to integrate, you have to be right. You can integrate 20 things that are of interest to customers. But inevitably you’re going to have customers balking about features they didn’t want and they don’t want to pay for.
Wingard: I’d like to challenge that. The $40 million, 90-nanometer design is an often-quoted number. A thorough analysis of where that $40 million goes shows much of that money is spent in ineffective integration of technology. The abstractions we have for putting systems together is very ineffective. When people are wiring things together, when any one element changes, everyone else has to change. There are strong benefits to abstraction. One of our customers is Broadcom, and what they do is tell their customers, ‘You can have three of these and four of those, and you can have it in six weeks. And it’s not going to cost you $40 million.’ They have put in place the design techniques and the design technologies that allow us to react to these markets.
Electronic News: But isn’t this a matter of ‘good enough’ versus best in class?
Wingard: There’s an aspect of ‘good enough’ and there’s an aspect of targeting a specific market space. They went from non-existent in the Wi-Fi space to the number one player by essentially stealing the market with a family of chips. None of those designs cost them $10 million.
Devlin: Traditionally, we’ve had a central processing unit where everything is sequential. The hardest thing is getting people to think in parallel.
Wingard: Sequential thinking is very much a software approach. The hardware people who are building the box have been thinking in parallel from day one. We’re seeing a switch where the algorithm people have to think more like hardware guys even though their implementation is programmable.
Jussel: We have to move away from hardware toward software. In some cases, it’s taking software and thinking about it in a parallel fashion.
[ Last edited by 碧城仙 on 2004-11-17 at 12:34 PM ]
页:
[1]
