For two US dollars, you can buy a bottle of water from the vending machine in a New York hotel lobby, or you could buy a single subway token. You could get into a New York taxicab, but you’d have to get right back out again. If you’re driving your own car, you could buy one gallon of unleaded gasoline. At most coffee houses, you could get a cup of plain drip coffee, but not an espresso drink. You could probably talk to your attorney for about 10 seconds. When it comes right down to it, two dollars won’t buy you very much. Now, however, it will get you a Xilinx Spartan-3E 100,000-gate FPGA with 72KB of block RAM, 4 18X18 hard multipliers, and 108 user I/Os.
Xilinx reckons that, within the last seven years, they’ve dropped the price of a single gate of programmable logic by a factor of thirty. This more than doubles the not-so-leisurely pace of Moore’s Law. During the same period, cell-based ASICs have probably fallen behind Mr. Moore due to skyrocketing NREs (and possibly Mr. Moore’s high-tech running shoes.) What this means is that the old equation for deciding which device should be at the center of your new high-volume application has undergone some major changes.
Spartan-3E is Xilinx’s fourth FPGA family on 90nm (assuming you count Spartan 3L and 3E as new families, but don’t count all the variants of Virtex4 as separate ones. It seems strange that Xilinx passed on the chance to claim that 3E is their sixth 90nm family starting with Spartan-3 over a year ago, and continuing with Virtex-4LX, 4SX, and 4FX, and more recently Spartan-3L, and 3E.) However you keep score, Xilinx now has a tremendous amount of experience on this tricky process node, while most of the semiconductor industry is just now coming to grips with the issues.
Sometimes the best magic tricks are the simplest. Xilinx used common sense and straightforward thinking to create this new family. With no technological miracles, Spartan-3E achieves most of its 30% cost savings over Spartan 3 by simply eliminating a ring of I/O pads. This shrinks the die enough to provide most of the area (and therefore cost) reduction. If your design is gate-heavy and I/O light, you won’t pay for the silicon occupied by all those extra I/Os. Unlike ASICs, FPGA families are designed with enough I/O pads for the highest pincount package they’ll fit inside. All of the smaller packages, therefore, leave I/O pads idle on the die. Spartan-3E takes advantage of that by giving you the opportunity to effectively turn those pads back in for silicon-store credit.
Spartan-3 will ship in 5 sizes, ranging from 2160 logic cells (100K system gates), to 33,192 logic cells (1.6M system gates. If you’re still confused by what a “system gate” is, we offer some insight [here].) The devices have extra goodies, including from 4 to 36 hard multipliers, from 72K to 648K of block RAM, and from 2 to 8 DCMs. The maximum I/O ranges from 108 to 376 pins. While the 100K 3S100E rings in at a paltry two bucks a chip, the 1.2M gate 3S1200E will go for $9, dropping our benchmark lowest price for a million-gate FPGA by about 25%.
In addition to lowering the cost of their silicon, Xilinx is apparently taking notice of total system cost arguments made by competitors such as Actel. In this family, they’ve attacked the total system cost problem by adding SPI/Parallel commodity flash configuration memory support. Xilinx points out that, if your system has existing flash memory (and you can swipe a bit of it for FPGA configuration use), you can get essentially the zero-cost configuration touted by non-volatile FPGA vendors.
Xilinx is also taking this family straight to the consumer electronics battlefield, fortifying it with consumer I/O standards such as Mini-LVDS, PCI-64/66, PCI-X, and DDR 333. In the highly competitive consumer market, features such as interface standards can be more of a deciding factor in chip selection than the traditional price/density and price/speed benchmarks.
Spartan-3E is not restricted by a lack of advanced DSP and embedded processing features either. Xilinx has included enough hard multipliers (now boosted to 325MHz operation) to make Spartan-3E into a very competitive cost/performance DSP solution, claiming that their MicroBlaze RISC processor can be implemented in only about 48 cents worth of logic. So, for applications that have moderate to serious math crunching and some requirement for embedded software and/or microcontroller functions, Spartan-3E gives a pile of capability, integration, and flexibility at a price highly competitive with most any type of solution on the market.
Xilinx is clearly going after the consumer market with this product, citing examples such as LCD TV where volume is high and time to market is precious, and also where flexibility and programmability offer decisive advantages in handling diverse standards, balancing feature sets, and extending product life in the field. With the pace of recent announcements in low-cost high-volume FPGAs (three in the past month including Actel, Lattice Semiconductor, and now Xilinx), it will be fascinating to see which devices get traction in which application areas. This battle for the cheap parts is clearly the hardest fought game in programmable logic right now. However, since FPGA prices won’t have to drop much more before we start finding them free inside boxes of breakfast cereal, cost isn’t likely to be the final deciding factor.