Compressing ADCs defuses data rate explosion in Data Acquisition Systems
If there were a DAS Olympics, algorithms competing in the "Compression for Medical/Wireless" event would be among the fastest.
Daniel explains why wireless and medical imaging systems cannot simply appropriate data compression technologies, such as MP3, used in other fields, elaborating on why compression technologies fall short for higher-performance applications. He then describes a new compression algorithm that can keep pace with sample rates of up to 40 gigasamples per second.
Today’s Data Acquisition Systems (DAS) using Analog-to-Digital Converters (ADCs) in wireless infrastructure and medical imaging applications are adding more channels while at the same time increasing both bit-resolutions and sampling rates, causing a data explosion at the system level. All this data must be transported from the DAS, where it is acquired, to processing devices such as FPGAs or CPUs, where the raw data is processed. This “data pipe” uses many device I/Os, crosses various interfaces and busses, and often includes different forms of storage along the way such as DRAM or disk drives and RAID. Over the years, the brute force approach to increasing the size and capacity of this pipe was to throw more hardware at the problem. Indeed, ADCs grew wider in bits; FPGAs grew larger in density, and memory requirements increased accordingly, with the same thing happening to a lesser extent in I/O as well. Systems grew more complex, used more power, and saw overall higher bill of material costs.
Dealing with this data explosion without increasing complexity and cost in FPGA-based designs means adopting some form of data compression in the FPGA, usually with IP algorithms. Real-time signal compression of the data sampled in systems such as wireless and medical imaging can take on the data geyser, at the same time reducing system complexity, power, and costs.
Over the years, various compression technologies have been developed to deal with the bandwidth and storage demands of sampled data. For example, telecom applications use Adaptive Differential Pulse-Code Modulation (ADPCM) to compress speech from 64 kbps to 24 kbps. Audio applications have moved from CDs at 700 kbps (44.1 ksamples/sec) to MP3 files at 64 kbps.
But while many industries and applications have adopted compression, high-speed sampled data in such applications as medical imaging and wireless infrastructure traditionally could not use these technologies. MP3 compression, for example, achieves its high compression ratios of 10:1 by exploiting well-known characteristics of human hearing known as psychoacoustics. The distortions introduced by the lossy MP3 compression algorithm occur in frequency bands that have no effect on the frequency response of the human ear. Unlike how it behaves with regard to the audio intended for human ears, MP3 will seriously distort the spectrum of an RF signal coming from an ultrasound transducer or from a wireless base station antenna. Signal properties such as center frequency, bandwidth, and SNR are simply unrelated to the frequency response of human hearing. Similarly, other compression standards such as those for images and video like JPEG and MPEG take advantage of limitations in human vision such as spatial frequency resolving and color discrimination to achieve their lossy high compression ratios.
Other compression technologies also fall short for higher-performance applications. Lossless compression products (such as WinZip) using variants of the Lempel-Ziv-Welsh (LZW) algorithm, for example, can reduce the size of computer files. These will not introduce any distortion of course, but would achieve low (less than 10 percent) compression ratios on signals with widely varying characteristics. Medical and wireless applications don’t use these compression algorithms, both lossy and lossless, because they are slow. The fastest of these algorithms hits a brick wall at around 50 Msamples/sec, and most stop at much lower than that, unable to keep up with the ADC’s fire hose blast of data. And consider that many systems, such as wireless, tightly specify latency. These complex algorithms introduce latencies that may be larger by an order of magnitude or more than existing standards dictate.
It’s important to note that key attributes shared by all these algorithms include the use of a priori information regarding the properties of the data they compress, non real-time operation at high sample rates, and offering of either lossless or lossy functionality but not necessarily both.
For the fastest growing segments of the high-speed ADC market, medical imaging and wireless, new solutions based on compressing ADCs to support their high-bandwidth, high-performance application are required. FPGA designers are now seeing new compression algorithms coming on board to alleviate the data-bottleneck issues.
One such compression algorithm makes it possible to target ADC applications where compression hasn’t been used previously due to the limitations of the existing solutions just described. This new real-time compression algorithm – called Prism – from Samplify Systems (see Figures 1-3) keeps up with sample rates of up to 40 Gsps, has extremely low latency, supports both lossless and lossy modes, and does not introduce strange distortion artifacts in its lossy modes. Furthermore, the amount of loss introduced is controllable, thus allowing the user to make the trade-off between compression ratio (bit rate) and signal distortion. And, because of the compression efficiency of this algorithm, designers can use lower-end FPGAs with simpler (or no) transceivers to decrease end-system costs and complexity.
A three-pronged approach to the sampled data that wireless, medical, and other systems (such as radar) produce so generously could include:
· Reducing the aggregate bit rates with this new signal compression technology
· Transporting the compressed data across high-speed serial interfaces to shrink the number of I/Os
· Using port concentration to efficiently maximize the use of those I/Os
Port concentration does not employ compression and is a lossless operation. Port concentration multiplexes the bit stream into fewer I/O channels when the full channel bandwidth is not being used.
From a system-level perspective, to maximize compression’s ability to reduce I/O and thus lower complexity and power consumption, the acquired data should be compressed as soon as it is sampled, and decompressed far down the data pipe, closest to where data is processed.
A compressing ADC
The SAM1610 device from Samplify is a compressing ADC. This single-chip device integrates Samplify’s Prism compression and port concentration algorithms on an ultra-low power, 16-channel, 12-bit, 65Msps ADC device optimized for ultrasound and wireless.
And, there is much benefit to this integrated approach. In medical ultrasound, for example, the ADCs in the front-end of a 256-channel console machine will generate upwards of 150 Gbps. All this data has to be transported from the ADCs to the receive beamforming FPGAs. This translates to 576 FPGA pins for data and clocks alone, resulting in complex backplanes, large connectors, and high I/O power consumption. Using compressing and port concentrating ADCs can reduce the I/O count of this interface by up to 75 percent, which translates to a savings of hundreds of FPGA pins. In portable, hand-carried ultrasound machines, reducing power consumption is paramount for operating off a battery. The power savings of employing compression and port concentration in hand-carried ultrasound enables support of advanced imaging modes such as color Doppler and harmonic imaging without degradation of image quality or battery life.
Integration offers other benefits as well. For example, an ADC with integrated compression can also facilitate new imaging modes that require software beamforming. In these applications, sampled ultrasound echo data is buffered in memory in the front-end of the system prior to being sent across PCIe to the back-end for processing. A compressing ADC will reduce the data transfer rates, memory size, and costs and may allow use of fewer PCIe channels. These savings will in turn result in reduced I/O and power consumption.
In wireless systems, compressing ADCs can reduce transport infrastructure costs by reducing data rates to allow advanced systems generating higher amounts of data to continue using existing infrastructures. Applying this compression technology in 4G base stations, and wireless repeaters enables the continued use of lower-cost fiber optic transceivers or low-cost CAT-5 cabling, as well as low-cost FPGA transceivers.
With the availability of technologies such as Samplify Systems’ compression and port concentrating ADCs, medical imaging and wireless system designers can finally attain the benefits of data compression that have been commonplace for decades in most audio and video applications. Applying this technology significantly reduces their I/O count and I/O power, removes bandwidth bottlenecks, and delivers less complex cheaper systems.
Daniel Kreindler is Director of Marketing for Medical and Imaging at Samplify Systems, a fabless mixed-signal semiconductor company based in Santa Clara, California. Daniel holds a BSEE from The Technion, Israel Institute of Technology and is currently completing his MBA at Santa Clara University. He can be reached at firstname.lastname@example.org.