Third-Time Lucky for Fast-Chip?
The tactic wasn't very successful, however, because it was tough for the company to counter criticisms from competitors and analysts that its first characterization was more accurate -- and that the chip was too limited in the applications it could address.
Now the company is out to prove its critics wrong. Yesterday it unveiled a whole bunch of improvements to the PolicyEdge, which should open up new markets for the chip (see Fast-Chip Delivers Fast Chips).
This time, Fast-Chip wants to be seen to compete with the network processor elite, which currently includes four companies actually shipping 10-Gbit/s packet processing silicon -- Applied Micro Circuits Corp. (AMCC) (Nasdaq: AMCC), Bay Microsystems Inc., EZchip Technologies, and Terago Communications Inc. (see Bay Joins the Big Leagues, EZchip Sallies Fourth, and Terago Springs a Surprise).
"People can throw stones at our architecture," says Charlie Jenkins, Fast-Chip's VP of marketing and business development. "But we've added features based on customer-driven needs. This is the value of being a second-generation part, whereas everyone else still has first-generation parts."
So, has Fast-Chip now got a real network processor? That's a tough question to address, because network processors are complex chips, making true apples-to-apples comparisons with competing products tricky. The answer also depends on to whom you talk. All the same, it's worth looking at the improvements Fast-Chip has made in the second-generation PolicyEdge and, where possible, positioning this against other leading network processor vendors.
In a nutshell, the new chip is twice as fast as the first; it has multifunction interfaces that allow it to connect directly to chips from almost any other vendor; it has additional packet editing features; and it also sports a companion chip called the RouteExpand, that enables it to handle lookup tables with millions of entries. Probably the most significant change is the addition of RouteExpand, so let's look at that in more detail.
If the old PolicyEdge had a downside, says Jenkins, it was the number of lookup table entries it could support, which was limited by the size of the on-board memory to around 16,000. This precluded the chip from being used in core applications, which typically require table sizes in the region of a million entries. To fix this, Fast-Chip has added RouteExpand, which is an algorithmic search engine that also gives access to off-chip memory.
RouteExpand is designed to work with RLDRAM (reduced latency dynamic RAM) chips -- a new kind of memory chip being developed by Infineon Technologies AG (NYSE/Frankfurt: IFX) and Micron Technology Inc., aimed at combining the density of DRAM with the faster access time of SRAM. Right now, RouteExpand can support lookup tables with up two million entries. When the next generation of RLDRAM becomes available in 2003, RouteExpand will support up to four million table entries.
In addition, RouteExpand holds accumulators and timers that store information and statistics for billing purposes, making it similar in function to the Accountant chip from Silicon Access Networks Inc. (see Silicon Access Launches Billing Chip).
But there is a potential downside. As a result of adding RouteExpand, Fast-Chip now offers a two-chip chipset, whereas other vendors have more integrated solutions -- a fact that Jenkins acknowledges. EZchip, for example, claims it can support lookup tables with millions of entries using a single chip (see EZchip Sallies Fourth).
Jenkins defends Fast-Chip's approach. "If you put all the large table support on-chip, you're actually forcing people to buy something that they don't need for many applications," he says, giving the following example. One of the key applications for PolicyEdge is in Ethernet LAN devices, where it is used to aggregate multiple Fast Ethernet or Gigabit Ethernet ports onto a 10-Gbit/s link for backhaul across the network. This application doesn't require large lookup tables, says Jenkins; the PolicyEdge, sans RouteExpand interfaces directly with an Ethernet LAN switch chip.
The cost advantage becomes apparent when considering that Fast-Chip's 10-Gbit/s second-generation PolicyEdge (FC6402) is priced at under $350 in quantities of 5,000, whereas EZchip's NP-1 network processor costs $1,150 in quantity.
In fact, whether to integrate packet processing functions more tightly, with the power and space advantages that brings, or to break them out for flexibility is one of the hottest debates in the network processor space right now.
Let's allow Fast-Chip to have the last word on the subject of being a "proper" network processor. "[PolicyEdge] is more capable in terms of packet editing than any competing network processor, especially with respect to AMCC, Bay, and EZchip -- and I'm quite familiar with what Bay can do," Jenkins contends, referring to the fact that he used to work at Bay.
PolicyEdge can perform up to eight edit operations on a single packet, he claims. In other words, it can perform recursive searches -- where the next lookup depends on the results of the previous one -- and then modify the packet after each search. That's important, he says, because it allows customers to do complex protocol translations, which need a lot of edits.
Vendors like Bay and Xelerated AB, on the other hand, only have one pass at a packet, because they use a pipelined architecture (see Swedes Claim Processor Advance. At each clock tick, the packet is pushed onto the next stage in the pipeline, so information cannot be passed backwards. As a result, there will be applications that they cannot handle, Jenkins contends. However, Bay's chip includes traffic management functions, whereas Fast-Chip's does not -- yet another factor to take into account in the complex calculation of tradeoffs that systems vendors must make when deciding whether to use a network processor.
Fast-Chip plans to sample its second-generation PolicyEdge in August.
— Pauline Rigby, Senior Editor, Light Reading