Ultra-high Performance, Low Power, and Ubiquitous Data Centers on a Chip

Today’s computing has been a great success. Healthcare, financial, communication, and entertainment industries rely heavily on data centers. By some estimates, the amount of data processed in data centers is doubling every year. At this pace, there will be needs for many ultrahigh-performance data centers processing 1000 times more data in the next decade. However, today’s data centers typically consume ~ Megawatts of power, require > 100 racks, and weigh many tons. Scaling to a larger scale and higher performance data centers and supercomputing centers is fundamentally difficult even if the Moore’s law successfully continues to realize highly-integrated computing chips. In fact, the Moore’s law is already facing significant challenges due to the fundamental issues such as atomic dimensions, impedance, and power dissipation (the power-per-processing ratio is not improving). On the other hand, recent advances in nano- and micro-photonics promises great potentials for interconnecting and switching information on the same platform as electronics.

We propose to design, prototype, and test a revolutionarily new computing system with orders of magnitude reduction in power consumption and size compared to the current state of the art. Such a computing system can be realized on a ‘chip’ to enable a clustered super computing data center scalable to 100 Petaflop or to allow a high-performance data center to be at everyone’s homes, offices, hospitals, and automobiles. Ubiquitously networked supercomputing will not only revolutionarily transform our lives but also help solve grand-challenge problems. Proposed is a comprehensive research Center project which pursues a clean-slate approach to designing innovative hardware and software technologies for next generation computing systems.

2009 Update:

Healthcare relies heavily on data centers. We are seeing rapid transitions in data processing from one-dimensional text files to two-dimensional and three-dimensional images. In less than 10 years from now, our healthcare services may involve interactive medical imaging in 3-dimensions requiring ~ 1 Terabit/sec data (1000 x 1000 x 1000 pixels x 28 bits-color/pixel x 30 frames/second) processing in real time. The stored 3-D image data may need to be compared against genomic data, or may need to be processed for future 3-D searches through exabytes (1018 Byte) of data. Another aspect of next generation healthcare is very high-performance mobile telemedicine. For example, if today’s data center can fit on a desktop, a handheld device, or a wrist watch, we can envision revolutionary changes in the healthcare industry by offering a new very high-end healthcare services at everyone’s home or high-performance emergency healthcare on ambulances to address ‘the first minute’ medical emergency problems. By some estimates, our data centers must handle ~100 times more data in 10 years. However, today’s data centers typically consume megawatts of power and require massive power distribution and cooling infrastructure. The proposed project pursues next generation ubiquitous data center technologies and healthcare solutions.

Our Team has successfully demonstrated CMOS compatible silicon photonics 10x10 crossbar switches and athermal silicon photonic modulator and switches have been demonstrated. A new optically interconnected memory system (OCDIMM) has been designed and simulated to show significant improvements in throughput and latency overcoming typical latency-bandwidth limitations while supporting massively parallel concurrency. The seed project has attracted a new DARPA Si-PhASER project worth $1.732 million for the phase I (18 months) possibly more for subsequent funding. The seed project also spurred four workshops at HP, San Francisco, and Vienna, and allowed submission of a new NSF Engineering Research Center proposal ($37.5million /2) and a new NSF Science and Technology Center proposal ($50 million /2).