Memory Chip Architecture Development - Harnessing Multiple Cores for Parallel Execution

"My design borrows extensively from today's modern multicore CPUs," said Joseph Ashwood, an independent security cryptanalyst and design consultant residing in Gilroy, Calif. Ashwood was lead cryptanalyst for Arcot Systems in Santa Clara, Calif., before going independent in 2001. "As far as concurrency goes, my memory architecture shares some features with Fibre Channel."

According to Ashwood, his architecture provides parallel access to bit cells on memory chips, breaking the serial bottleneck that is strangling nonvolatile storage media like flash, with an architecture that can be applied to any memory chip bit cell. The Ashwood memory architecture works by integrating smart controller circuitry next to the memory array on a single chip, providing parallel access to the array for hundreds of concurrent processes, thereby increasing throughput and lowering average access time.

"We have a new way of assembling the memory, with a few new elements I was led to by my experience with cryptography. I am basically applying very deep cryptographic techniques to memory architecture, resulting in a unique new design that is very fast and compact. Bringing in these new elements enables a lot of good things, especially concurrency, permitting hundreds of simultaneous memory operations," said Ashwood.

"Compared to DDR, for instance, my architecture goes inside the chip and reorganizes how the bit cells are accessed, thereby utilizing them much more efficiently," he added. "Transfer rate is faster, too—for instance, right now, DDR-II for DRAM only goes up to 12 Gbytes per second, but our architecture can deliver 16 Gbytes per second when using flash memory and is compatible with PRAM or any other nonvolatile semiconductor memory cells."

Sound too good to be true? JoAnne Leff, founder of J.L. Associates (New York), thought the same thing when she was first approached by Ashwood to represent him in licensing the technology. So she sent the design over to Carnegie Mellon University for confirmation.

"We were skeptical, of course, but Carnegie Mellon confirmed for us that the Ashwood memory architecture really is a breakthrough in memory design," said Leff. "Now we want to license it to all major players involved in the applications of this technology, not only to improve the performance of individual memory chips, but also to give users fast, parallel access to solid-state drives."

Carnegie Mellon's evaluation for J.L. Associates claims that solid-state drives using nonvolatile memory chips is an especially good application, and that the Ashwood memory architecture could rejuvenate the nonvolatile memory markets, such as flash, by improving their performance today and tomorrow, since scaling to larger capacities also increases concurrency.

"This new technology enables parallel data storage and access on a single nonvolatile memory chip. The scalability created by this technology provides superior speed in accessing and storing data with higher storage capacity at a single chip level. With on-chip power management, the technology is truly enabling for applications that require on-chip high-speed data transfer for various high-capacity, nonvolatile memory devices," said the Carnegie Mellon evaluation. "It has been predicted by many people, such as Gordon Bell, that by 2015 personal devices such as PDAs and cell phones will need to have a nonvolatile storage capacity of at least 1 terabyte. This memory technology invention provides a solution to meet that need."

Ashwood, however, does admit to two downsides to his memory architecture. First, it is still just a paper design. He plans to work with licensees to implement his design on their memory arrays, but so far only a software simulation has been completed.

"I have fully developed the memory chip architecture, and I have run a software simulation to verify that it works, but so far I have not done a design at the electrical signal level—that kind of detail is dependent on who ends up licensing it," he said.

The second downside is that the parallel access overhead of the Ashwood memory architecture slightly slows down memory access times to individual memory cells—a disadvantage that is offset by its many parallel access channels, Ashwood said.

"For instance, if a NAND flash chip has an access time of 20 to 50 nanoseconds today, adding my architecture would increase that access time to 50 to 70 nanoseconds," he said. "But remember, during that time, 100 or more other memory retrieval operations could be in progress concurrently, yielding an effective access time of just a few nanoseconds per retrieval."

Late last year Ashwood filed a patent on his memory architecture, but because chip makers could implement it before the patent is grated, he is choosing to keep most of the architecture secret until the patent is granted next year

"This architecture is so easy to implement—a chip maker could roll out a working prototype in as little as three months," said Ashwood.

Ashwood, however, has disclosed the main outlines of its functions—reorganizing the memory hierarchy for parallel access to a chip's data—as well as described its features and made performance comparisons with both DRAM and hard disks.

"Unlike traditional memory architectures that degrade in performance as more bit cells are added to an array, our memory architecture's performance increases each time you add cells," said Ashwood. "For instance, because of the way our architecture scales, if you double the capacity of our memory chip, it also becomes twice as fast as before."

Overhead is low, too, adding only about 3 percent to the die area of a memory chip, according to Ashwood.

"Current flash cells are already extremely dense—they should be able to fit a terabyte in less than a cubic inch," said Ashwood. "The problem is that their yields would fall to 30 percent and the speed to just 32 million bytes per second. Utilizing my technology with the same flash cells allows a yield in the upper 90 percent [range] and a speed of 16 billion bytes per second."

When using the Ashwood memory architecture with multiple flash chips configured as a solid-state disk (SSD), instead of adding 3 percent, each memory chip's die area could actually be smaller than it is today, because the access circuitry to the SSD would be common to all the memory chips on the drive, he said.

Using his memory architecture will also increase the lifetime of a drive, according to Ashwood. Flash cells can only endure about 100,000 cycles before they burn out. By offering greater flexibility in the page reassignment, he said, the Ashwood memory architecture can increase the lifetime of a drive by about 500 times.