Guest Blog: Can We Afford ASICs?

In the medical realm and other areas (e.g., automotive, networking), architects often ask themselves at the time a new product has to be designed if they can afford application-specific integrated circuits (ASICs). In past years, a lot has been published about the pros and con of using ASICs. These reports always speak about power consumption, size, and capabilities, but in the end there is a prohibitive factor: the price.

In the past year, the price of masks has increased considerably, and it is expected to increase even more. Prices above $5 million for masks are not so uncommon anymore. Who can actually afford such costs!?

This problem can be more acute for some high-performance, demanding applications such as medical imaging. In such applications we need to have significant processing capabilities and good power consumption performances… Sure, a well-designed ASIC would work wonders here. For an OEM, a dream come true would be to convince a semiconductor company to take an ASIC, pack it with some ARM and/or other cores and add some peripheries (e.g., LVDS, USB) and produce a chip.

That sounds good in theory, but in practice it is not so easy. Semiconductor companies are reluctant to produce chips that can be sold only to one customer, except in cases where the money is right. In the consumer market the sales volume is significant, but in a medical environment the number of sold parts is small so there has to be a special deal between the OEM and the semiconductor company. One can imagine that only a big enough OEM can afford to consider such a direction.

For example, if we would produce at this moment a VScan- (GE) or P10- (Siemens) like device (small, handheld ultrasound scanners) with a mask cost of $5 million, we would have to sell around 600 such devices for the amortization of only the mask cost. One has to sell 5000–6000 such devices to obtain a decent margin.

These cost problems will be reflected ultimately in the cost of care, in the amount of money each of us is contributing to the system. So, what is the solution to this cost problem?

One solution would be to use off-the-shelf hardware that is widely used in other applications. Sure, nothing can be more power efficient than a well-designed ASIC, but the price tag attached is a little prohibitive.

Since I am more or less active in the medical imaging area, let's see what off-the-shelf hardware we can use there.

For the signal processing part:

  • FPGAs
  • DSPs
  • GPUs
  • Off-the-shelf ASICs

For the display and user interface part:

  • ARM, Atom-based platforms

GPUs nowadays are very sophisticated and powerful, and it is definitely worth it to consider them for some bigger sized devices. FPGAs and DSPs are common solutions for signal processing applications, and the Figure 1 summarizes the capabilities and costs of each (in a medical ultrasound application).

Figure 1. The capabilities and costs of FPGAs and DSPs.

Atom and ARM cores are proven architectures in the consumer area, so it is only natural to be present in the selection of display and interface section. Power architecture–based devices are used more in the networking domain, and as such, they might be a little off limits due to possible on-chip accelerators, lack of connectivity, etc.

Regarding the off-the-shelf ASICs, Samplify Systems, for instance, designs frontends for the medical ultrasound. The user has some degree of freedom to bring some of his IP inside, but not much. Because the frontend characteristics are practically the parts that will determine in the end the image quality, this is a sensitive area. It is very difficult to sell such a product to many OEMs because there will be no real differentiation between products and because many of the OEMs already have patented and tested solutions of their own.

In my next posts I will speak about some of the off the shelf hardware available nowadays, and we can debate the suitablily of these architecutres.

Robert Krutsch is a DSP software engineer at Freescale Semiconductor. He received a bachelors degree specializing in automatics and computer science from the Polytechnic University of Timisoara, a bachelors degree from the Automation Institute at the University of Bremen, and a doctorate degree in automatic control from Polytechnic University of Bucharest.