While DRAM memory cells have continually shrunk, scaling has, to some extent, reached its limit. Also, the mixed analog and digital technologies used to implement DRAM storage cells, sense amplifiers, and control logic have forced designers to compromise on some performance issues when co-integrating analog and digital technologies.
But now, a radical reorganization of DRAM circuitry that leverages wafer-stacking technology promises higher-density and lower-cost memories that deliver top-notch performance. This use of wafer-stacking technology and the revamping of the memory architecture and circuitry were devised by Tachyon Semiconductor Corp., a startup based in Naperville, Ill.
"To achieve the best digital and analog performance," explains Robert Patti, Tachyon's CTO and one of its founders, "we drastically revised the way a DRAM is architected, separating the digital circuitry from the analog, placing digital circuitry on one wafer and analog circuitry on another."
Separating them optimizes each circuit for the best performance possible. SDRAM access times of 8/4/4/4 ns (initial and three subsequent accesses) and densities of 1 Gbit/package will be possible in the first generation of memories built using the stacking technology.
In contrast, today's SDRAMs offer access times of 45/6/6/6 ns and top capacities of 256 Mbits/chip. Stacked SDRAMs can deliver high-speed accesses competitive with the best SRAMs while consuming little power, occupying less board space than standard SDRAMs or SRAMs, and costing less (estimated at about half the per-megabyte cost of SDRAMs).
Getting high capacity at low cost meant including built-in self-test (BIST) and repair circuitry to find bad cells, map them out of the array, and map in replacement good cells. Tachyon's proprietary technology, BiStar, was used to accomplish this task. BIST and repair circuitry employ an internal controller that performs a very fine-grained mapping of all the good bits at power-up. The circuitry can also work in the background to continually test the bits and replace defective ones. Separating logic and the highly interconnected memory cells makes cell-by-cell mapping possible.
BiStar technology includes error-checking and correction logic to catch transient errors due to alpha-particle upsets or marginally good cells. A memory stack can tolerate well over 10,000 bit errors per device. BiStar stacks can provide a manufacturing yield of 97%, keeping per-unit costs well below those of standard SDRAMs.
The wafer-stacking technology was developed by Tachyon but leverages several patents the company licensed from other industry sources. The process starts by creating two base wafers (Fig. 1). One holds optimized digital control logic and sense amplifiers, which are covered with an oxide insulator layer. The insulator layer is patterned to create contact openings. Copper is deposited, filling the contact holes and covering the oxide. That copper layer is then patterned before bonding to the second wafer.
Arrays of memory storage cells are formed on the surface of the second wafer. Intermixed with the memory arrays is the company's proprietary redundancy interconnect circuitry and self-repair structures, which greatly improve circuit yield and allow for in-system repair. This ensures that systems can remain online even if some memory bits fail.
Next, an oxide layer is deposited. Contact areas are opened. Copper is deposited to fill the via holes and then patterned. The two wafers are then attached, first by inverting the memory-array wafer, aligning it to the control wafer, and then by using thermal diffusion to bond the two together (Fig. 2).
The back side of the inverted memory array wafer is re-moved, leaving only a 10-µm thick layer by back-grinding and then chemically mechanically polishing to a flatness of ±1 µm. Deep vias are etched from the back side of the inverted wafer all the way through to the front side, exposing the copper deposited on the original top of the memory-array wafer. An oxide layer is then deposited and patterned, re-exposing the deep thru-wafer vias.
Copper is deposited on the back side of the inverted wafer, filling the vias and making contact with the exposed bottom of the now sandwiched original copper layer. The new copper layer is patterned and the surface is planarized, removing all the copper except for the metal now inside the via holes. A slight oxide etch lets the copper rise above the surface of the oxide, forming an array of microscopic posts.
Once the posts have been formed, another memory array wafer can be thermally diffusion bonded to the back of the first memory array wafer. This second memory wafer goes through the same back-grinding and deposition/patterning steps as the previous one. Another memory array wafer can then be attached, and so on. After all the layers are attached, the wafer stack is flipped over, and the bottom of the control logic wafer is patterned for bond pads or flip-chip mounting in the package.
Initially, Tachyon designers will use four memory array layers stacked on top of the single controller/sense-amplifier wafer. However, the controller/sense-amplifier circuitry has enough drive to handle up to 32 memory layers.
The initial stacked-wafer memory product will be a 1-Gbit double-data-rate (DDR) SDRAM (16 Mwords by 64 bits), operating at clock speeds up to 333 MHz. It will be internally configured as a quad-bank DRAM that can transfer two 64-bit words per clock cycle. A bidirectional data-strobe signal is available for each byte in the 64-bit word.
The DDR SDRAM will be able to operate from a 1.8-V supply and employ SSTL_2-compatible I/O lines. Engineering samples will be available in December. Customer sampling is expected in the first quarter of 2002, and a 4-Gbit version is expected in the fourth quarter of 2002.
For more information, contact Tachyon Semiconductor by phone at (408) 437-2300, or point your browser to www.tachyonsemi.com.