CPU和内存之间用无线通信!梦想还是噱头?

无线化并不只在网络领域大有作为,它也可以大大提高CPU和内存之间的通信速度,Sun微系统正在研究一种无线主板,计划实现CPU和内存之间的无线通信。在今年11月份的IEEE Spectrum上曾讨论了多核心的缺点,IEEE的一篇文章引录Sandia National Labs科学家的论断,他们认为在很多情况下,多核心处理器对高性能计算反而是不利的。

  Sun表示他们有很多方案可以解决这个问题,尤其是一项他们已经进行了四年的计划,名为Proximity Communications,它是美国国防高级研究计划局(DARPA)的一项研究项目的一部分。

  Proximity Communications允许CPU和内存插槽不通过主板上的多层铜导线连接,而是通过一对完全不接触的金属板发送和接受数据。这将明显减小内存和CPU通信间的延迟,在多核心处理器时代,多个处理器使用相同的主线和内存通信时,这种延迟问题就更加突出了,Sun表示,利用Proximity Communications技术,可以实现数据在两个目标之间的直接、快速传递。

  Sun计划在2009年5月于圣地亚哥举办的国际电子元件与技术(ECTC)大会上详细介绍Proximity Communications技术,Sun高性能计算产品市场主管Bjorn Anderson称,“Proximity Communications技术的真正应用到产品中还需要几年的时间。虽然该技术不能完全消除延迟,但它可以将延迟时间减少一至两个数量级。”由于该技术是DARPA计划的一部分,所以Anderson表示,他不敢确定其他厂商能否得到许可权。

其实这个技术已经研究了很久了,在Sun网站上2004年就有介绍文章了,但是也许今年这个技术有了突破!
Sun网站有该技术的详细介绍:
Proximity Communication - the Technology

Cross-section diagram of Proximity Communication

September 20, 2004 - On the printed circuit boards of today's computers, information and electrical power travel over copper wires between CPUs, memory and I/O devices. The wires connect the devices using technologies such as pins, ball bonding and solder bumps which involve macroscopic conductors, massive in size compared to the submicron features on the chip itself. Because communication between chips does not rely on wired connections, the number of connections between chips is much higher than with ball bonds (about 100 times greater). The chips can talk at much higher speeds with lower latency and significantly less energy than using wires.

How Proximity Communication Works

Microscopic metal pads are constructed out of standard top-layer metal structures during chip fabrication. These pads are then sealed with the rest of the chip components under a micron-thin layer of insulator to protect the chip from static electricity. Two chips, with receiver and transmitter pads, are then placed facing each other such that the pads are only a few microns apart. Each transmitter-receiver pad pair forms a capacitor, and voltage changes on the transmitter pad cause voltage changes on the receiver pad despite the lack of a wired connection. It is much like the physical effect that causes touch lamps to light when a human touches the conductive base of the lamp. Another analogy is the synaptic connection of biological nervous systems, where signals jump from one neuron to another.

The concept is deceptively simple. The actual details are much more complex, because the chip must contain logic for driving and amplifying the signals, and the receiver circuit must tolerate far more variation than a wired connection. The voltages can vary widely, so Proximity Communication technology is engineered to work over about a factor of ten voltage variation. Because mechanical misalignment can and will occur, ingenious methods compensate dynamically for effects such as vibration and unequal thermal expansion. The communication continues to function via its large voltage tolerance and dynamic reconfiguration for misalignment.

Proximity Communication Advantages

Proximity Communication provides an order-of-magnitude improvement in each of several dimensions: density, cost, speed, latency, and power demand. Because the space taken up by the communication path, the power and the cost per bit transmitted all go down, it will be possible to get tens of terabytes per second in and out of a single VLSI chip. Current technologies are limited to a few hundred gigabytes per second. With all dimensions taken into consideration, the overall capability improvement is two orders-of-magnitude.

Proximity Communication Consequences

"Wafer scale integration"

Instead of trying to make processor chips ever larger, with resulting lower and lower yields, Proximity Communication lets us lay out a "checkerboard" of chips that all behave as a single integrated circuit. Wafer scale integration has historically failed because the yield drops to zero as the silicon area of a chip increases. With Proximity Communication, one can get the same performance advantage as wafer scale integration but with excellent yield.

Top view of wafer scale tiling arrangement of
chips using Proximity Communication

Improved Test and Assembly

When a flaw in a chip is discovered, Proximity Communication allows one to simply lift out the chip and drop in a new one (clearly with some level of clean room conditions). This is very expensive or impossible with current methods that connect chips to multi-chip modules, forcing replacement of the entire circuit module instead of just the defective part.

Increased Technologic Versatility

One can mix different technologies. "Processor in Memory" has been talked about as a way to put a complete computer system on a single chip....but the process technologies used to build CPUs are very different from the process technologies that are optimal for building dense memory like DRAM. Because the Proximity Communication lets each part be manufactured separately but then integrated using Proximity Communication as the universal interface, the constraint of using a single manufacturing technology vanishes. It is even possible to mix, say, gallium arsenide and silicon chips in a single array. This is made possible by the fact that Proximity Communication is inherently tolerant of different voltage levels needed for different semiconductor materials, and also by the fact that Sun's approach includes asynchronous logic to remove the need for a common clock between two circuit chips.

Dramatic cost savings

Sockets, pins, and circuit boards add cost to a system. Proximity Communication eliminates them. While it is not yet a commodity technology, once it is produced in quantity it can be less expensive to create systems with Proximity Communication since it eliminates so many of the parts. And with Proximity Communication, chips can be smaller than they are now, increasing yield, and thereby decreasing the cost of each component chip. It's like creating a five-carat perfect diamond out of five one-carat perfect diamonds -- the savings in cost is dramatic.

Proximity Communication Applications

Proximity Communication is part of Sun Microsystems Laboratories' 3-year High-Productivity Computer System Project, supported by a $50M contract from DARPA (Defense Advanced Research Projects Agency). In this application, Proximity Communication aids in the development of a scalable supercomputer, where the requirement of extreme speed overcomes the initial research and development cost of bringing the new interconnection method to the point of being used in a manufactured product.

Products of lower cost that use Proximity Communication may follow, or may precede the supercomputer. Proximity Communication is applicable to any product that now uses circuit boards to connect VLSI chips. It can make virtually any high-tech device smaller, faster, and cheaper, all at the same time.

The speed of many computer applications depends on the system bandwidth available to transmit data. Database searches, scientific simulation of weather, traffic, manufacturing, protein folding, gene sequencing, and signal processing are just a few examples of economically important computations that are limited not by the CPU speed but by our ability to move operands in and out of each CPU. It had been thought that this was an irreversible trend that would cause computer speed improvements to level off or cease entirely in a few years. Proximity Communication buys us at least another decade of Moore's law improvements in computing speed. Its initial application will be those areas where performance is a primary goal; as it becomes a commodity, it will become the preferred solution for areas where price-performance is also a primary goal.

Proximity Communication does for wires what integrated circuits did for discrete transistors. Proximity Communication makes possible the miniaturization of not just the transistors, but the ways we connect them together into extended systems. It has the potential to improve our lives profoundly as it increases the speed or decreases the cost of high-tech devices by orders-of-magnitude.