networkZONE Products for the week of March 7, 2005
Cypress Semiconductor Says…
A Trie Grows In San Jose: Cypress Delivers Industry's First Single-Chip Algorithmic Search Engine
Innovative Algorithmic Search Engine Enables Lowest-Cost Packet Processing Subsystems
Cypress Semiconductor Corporation today announced samples availability of the Sahasra 50000 Network Search Engine (NSE), the industry's first single-chip algorithmic search engine. This new device offers superior price/performance for longest prefix match (LPM) and exact match packet classification applications, and is especially effective for high-capacity forwarding applications (one-million entries and above) where traditional solutions may not be cost effective.
The adoption of IPv6 and other new protocols, combined with increasing line rates and the need to support advanced network services such as virtual private networks (VPNs) and virtual router forwarding (VRF), puts tremendous strain on packet classification subsystems. By combining advanced search algorithms and high-performance memory on a single chip, the Sahasra 50000 NSE provides the optimal solution for forwarding applications. The deterministic, high search performance of 250 million searches per second (MSPS), and the ability to support as many as 1.5-million IPv4 routes in a single chip, allow the Sahasra device to offer the lowest cost-per-route of any network search engine solution in the market today. The integrated SRAM at the core of the Sahasra device enables full speed performance while consuming as little as 10% of the power of a comparable TCAM-based solution. This, coupled with the TCAM-like interface of the Sahasra device will greatly reduce system level design complexity.
"We are very excited to be able to offer customers an algorithmic device to complement our portfolio of TCAM-based search solutions," said Cathal Phelan, Vice President of Cypress's Data Communication Division. "Our customers have repeatedly told us that to address large forwarding tables, they need a solution with the performance of a TCAM but at lower cost and power. Coupled with the ease of migration from an existing TCAM-based solution, the Sahasra 50000 family delivers on our commitment to provide easy-to-use, leading edge search solutions."
"Sampling of the Sahasra 50000 NSE is a major milestone for Cypress, and the search engine market as a whole," said Jag Bolaria, Senior Analyst with The Linley Group, the leading independent analyst in the networking silicon industry. "This further establishes Cypress as the technology leader in the search engine market, with the broadest portfolio of high performance search solutions."
The Sahasra 50000 NSE is supported by a complete software solution, the Cynapse Software Platform, which includes the production-ready table management code as well as models and documentation to support the design-in process. The Cynapse architecture allows migration between the various Cypress NSE families with minimal application level code changes. This architecture, coupled with the interface characteristics of the Sahasra device, allows easy software and hardware migration from existing solutions.
EN-Genius Says . . .
I don't know Sanskrit, so I'll take Cypress Semi's claims that "Sahasra," the name of their newly-released search engine architecture is the word the ancients used for "infinite." And if the technology lives up to its claims, the name would be an apt appellation because of the many possibilities it opens up. This new SRAM-based search engine represents a bold move to re-think the TCAM market which could significantly affect the economics of this relatively small, but important niche market. In fact, if their claims hold up, the sharp cost reductions made possible by this new family of devices could change the way search engines are used.
Until now TCAMs have been a costly, but indispensable element of high end enterprise and core equipment for solving next-hop lookup, forwarding, and, ACL and other security functions at wire speed. In recent years, we've seen some lookup functions being handled by embedded logic that performs DRAM-based lookup to provide limited levels of CAM-like functionality, but they have not been a substitute for the speed, flexibility, and ability to handle large tables that TCAMs provide in large applications.
Cypress's Sahasra architecture combines SRAM technology and its own algorithmic search interface to deliver TCAM-like performance without the cost or speed of a TCAM device -- with only a couple of differences that won't matter in MOST (not all) applications. They do this by using a mesh consisting of clusters of SRAM linked by embedded logic which does most of the actual lookup work. The engine's entries are mapped into a Trie which is used to sequentially parse incoming route addresses using a succession of matches.
 This rigid cascaded search method means that any Sahasra search takes 54 clock cycles -- one of the differences between it and an equivalent-sized traditional TCAM which can take between 12 and 20 cycles to complete its search. Once the search engine's pipeline is filled, however, it pops out lookups in a deterministic manner at its clock speed (up to 250 MHz). Cypress argues that the longer latency (around 225 ns) is not critical in most forwarding applications since they are usually mostly look-aside functions where the extra few ns of delay can be dealt with by a slightly-deeper packet buffer. They also make a good point that it is more than worth it because their engine offers completely deterministic timing, something that TCAMs cannot do.
The second difference between the Sahasra architecture and regular TCAMS is that the search mechanism can only have "Don't Care" arguments at the beginning or end of search field. This means that they cannot support ACL functions and other lookups which have "Don't Care" fields interleaved with exact values. But for the many applications which can live with this limitation, the Cypress device seems to offer many advantages in cost and power. Because it is fabricated using "vanilla" CMOS instead of the proprietary process used for CAMs, and the much smaller size of SRAM memory elements, the basic silicon cost is lower. The higher yields that the smaller chips and standard process should make possible (my speculation) should also help to bring the $/Mbit of fast search capabilities down. Even in its first incarnation, the Sahasra device offers 4x the search area (72 Mbit) of the largest commercial TCAM on the market today -- and at an equivalent price. Using SRAM technology has also cut power consumption radically. The 72-Mbit device burns 5 W, over a 10x reduction from the 60 W for four TCAMs.
About the only other difference you might notice is the internal ECC scheme on the internal data path. Adding 8 bits of ECC to the 36-bit bus provides 1-bit correct, 2-bit detect capability. While this is needed because of the slightly higher soft error rate in SRAMs, Cypress claims that the ECC gives Sahasra a much better error rate than normal TCAMs.
But as interesting as the differences between the Sahasra and TCAMs are, the similarities are even more fascinating. For one thing, the interface and programming is identical to Cypress' TCAMs. It also can be segmented the same way, allowing multiple tables to be built in various widths. It handles 1.5-M 36-bit entries, or just under 750-k 72-bit entries, and so forth, all the way up to 576 bits. You can also search multiple tables simultaneously for things like combined forwarding and policy lookups. It even supports multiple hits in the same way that their TCAM devices do.
Sahasra's lower cost and power, plus the humongous address space it offers, makes the enterprise backbone and metro/edge applications that Cypress has targeted for these devices look like easy pickings -- if they can convince the relatively-conservative design community to gamble on a "new-fangled" technology. Towards this end Cypress has invested a lot of effort to win their trust and to make the transition to their new product very straightforward. For example, they spent over 2 years developing application software, a table manager, and drivers that have user interfaces and functions that are identical to their TCAM line with only difference in the hardware-specific drivers. This lets developers work with TCAMs and Sahasra devices with the same tools, mixing and matching chips as needed for optimum cost and power.
If the devices work as-advertised (and I believe they do), they could significantly alter the economics of high-end equipment. Until now, CAM costs have been a significant part of the cost of a line card or switch. Typical TCAM content cost will drop from $1000 to around $250. Designers should also realize additional savings in board space, lower thermal loads, as well as reducing the numbers of decoupling caps, and other signal-integrity efforts.
Cypress is to be congratulated on having the courage to "eat its children" by offering products that directly undercut sales of its traditional TCAMs in some markets. I think that in the end their effort to lower the cost of lookup will be rewarded as designers now find uses for the devices in places that were not considered practical before. I might even expect to find Cypress offering smaller-sized Sahasra devices, (perhaps as low as 9 Mbit) for use in applications closer to the network edge where quick classification could assist in security, or QoS management. Cypress also gets extra "brownie points" in my book for delivering on their claims because they have had samples in hands of customers for a couple of months before the actual release.
As noted, Sahasra is sampling with production scheduled for the Q2 2005 priced between $300 and $325, depending on volume.
Editor's Note: Apart from being a female holy name in India, such as Visnu Sahasra-Nama, sahásra is 1000, not infinity, in Sanskrit numbers. ékaadhikasahásra, for example, would be 1001, ie. éka (1) + adhika (plus) + sahásra (1000).
|
