by Tets Maniwa

March 24, 2010—ISQED, San Jose,CA—Ramanan Thiagarajah, senior director of product and test engineering at Inphi Corp. described the overall process Inphi uses to design and build very high performance memory solutions for server-class computer systems.

He started by describing the transition from centralized mainframe computers to distributed arrays of smaller computers like PCs to client-server architectures and back to a form of centralized computing, now called cloud computing. Cloud applications are now everywhere; YouTube, Twitter, Picassa, … and include applications, on-demand compute power, and unlimited storage. The drivers for the move to cloud computing include technology, energy efficiency, and economics.

The common threads for the cloud environment include data and computation anywhere and any time along with the availability of virtually unlimited on-line storage makes these functions possible. Because the CPUs have run into physical limits in their ability to scale clock frequencies any further, the advent of multi-core processors and virtualization software has enabled the use of massive compute resources for anyone one and for everyone.

As the servers and storage networks grow, however, they consume increasing amounts of power. Operating and cooling account for the largest single percentage of costs in IT centers and is growing at a faster rate than any other resource. The very large server centers house hundreds of thousands of CPUs and hundreds of gigabytes of memory to handle the vast amount of data pouring through the Internet.

In terms of economics, cloud computing offers the opportunity for significant savings in capital expenditures and also in the costs of operating these large systems. Instead of collecting resources and staffing for peak loads, eases can access computer and storage resources on demand. Current forecasts are for the aggregate cloud computing to become a $100 billion market 2013. In addition, the market for cloud-based services is expected to grow over 25 percent in the same timeframe.

One might ask, is there a real need for this much hardware and storage. Requirements for greater bandwidth are evident in many areas. For example, the downloads at Apple Computer have increased by a factor of 10 in the past year. Internet-based traffic is expected to exceed 20 exabytes in early 2012. This amount of data traffic requires improvements in servers, storage, network infrastructure, and access. The data volume is user driven, everyone wants to watch movies, look at pictures, and interact with their many friends all the time.

The opportunity space for cloud computing resides in the market sector in terms of time to market, differentiation, and cost and quality concerns. At the same time, users are asking for highest performance at the lowest power with the greatest reliability possible. To address these requirements, servers are migrating to a multi-core architecture in a virtualized operating system environment. In 2007, computers provided a base level of service with four cores in two sockets. By 2017, servers will have 32 cores in four sockets, and will provide 16 times as much computational resources as 2007 versions of servers.

A growing bottleneck in server systems is the memory. DRAM density and performance is not scaling to match CPU performance. In one of them latest memory technologies, DDR3, frequency scaling is a significant challenge. Due to the parallel bus structures loading and crosstalk become insurmountable in the very high capacity memory systems. The metric of performance per watt per dollar is adversely affected. Those users depending on server throughput see large negative financial impacts due to the increased latency of memory-constrained systems.

To scale system performance in line with the increases in computer throughput requires innovative changes in memory bus architectures. Normally, larger and faster memory arrays draw more power. Enterprise class servers are now allocating almost half of the power budget just for the memory, and the next generation of servers will incorporate larger memory arrays to improve throughput.

Inphi has created memory modules with twice the memory capacity of standard modules and operate on one third less power. In a typical data center, this could amount to a savings of 30,000,000 MW of power a year and a cost savings in power and cooling costs of $6 million.

Cloud computing represents a very large opportunity for the semiconductor industry, because semiconductors are a key enabler of innovative solutions. Large data centers are optimized for reliability and availability due to the number and types of applications running on the systems. Users need dependable up time and data integrity in their systems so the redundancy available through multi-cores and virtualization are essential characteristics of any large data center. This level of reliability must be designed into the system from the beginning and must exist throughout all components in those systems.

In summary, everyone is using cloud computing today. The primary characteristics driving the adoption of cloud computing are energy, performance, and reliability but the cost savings of cloud type data center are not insignificant. Because data volume continues to grow at an exponential rate, hardware performance must be continually upgraded. Semiconductors are enablers to major improvements in performance. And finally, the market is ripe for a new generation of innovative semiconductor solutions.