We are running out of DSP board bandwidth.
We need more as DSPs take on more tasks and we pack more DSP power
into boards and modules. One solution is a faster system bus.
Another is to add adjunct buses for alternate fast datapaths for
data. A new alternate bus is also emergingserialized PCI. And
the DSP board community might well have a look.
Serialized PCI is a no-brainer. Yet it took until now for
designers to take serializing PCI as a connection option. It's
simple conceptually, but perhaps a bit harder implementation-wise.
Basically, serialized PCI does the following:
- Uses standard, ubiquitous PCI connections at both ends
- Uses a high-speed serial bus between the PCI connections to
match PCI bus bandwidth
- Takes in parallel data as PCI transactions, converts to serial,
passes to other end, de-serializes, and then presents as the
original PCI transaction
- Treats as a limited PCI bridging problem, addressing the PCI
- Is software and hardware compatible with PCI.
What you get is a PCI connection, with a peak bandwidth of up to
133-MB/sec. It is PCI compatible, so you don't have to change or
reinvent any software. It's really just a plug-and-go kind of
situation for users.
Serialized PCI is not going to give FPDP a challenge. FPDP is
already on its second generation, pushing 400-MB/sec. And it's a
simpler bus than PCI, basically a multi-drop ribbon cable-based bus
that can move data, but lacks (by design) more complex bus
protections. But serialized PCI provides an easy way to add one or
more PCI-speed alternative buses with virtually minimal software
changes. And it's not a foreign bus technologyit integrates
well with existing PCI-bus infrastructure on SBCs or
PCI Goes To Level 4
PCI is the only bus that exists at three different
implementation levels: as a systems bus (cPCI, PCI/ISA), as an I/O
bus (PCI), and as a mezzanine bus (PC-MIP, PMC, PC/104 Plus). PCI
follows in the footsteps of the IBM 360 and the DEC VAX. These were
the first computers to hold their Instruction Set Architectures
(ISAs) fixed, and to implement different price/performance points
with a common ISA. Before that, every time the silicon technology
took a jump, the computer designers would design a new computer and
ISA. But the 360 and VAX changed all that with fixed ISAs, which
were tremendously successful.
PCI ended up doing the same thing. The basic PCI bus was held
fixed, but it was implemented with different electromechanical
formfactors and drive capabilities. The same basic PCI bus runs in
PCI/ISA, CompactPCI, and PMC bus implementations. The only thing
that is really different is the card/board and connector. And the
drive and terminations may be tuned for different environments, but
is the same PCI bus and can use the same low-cost PCI silicon.
Today's serialized PCI adds another layer to that PCI
implementation heap, that of intersystem and inter-board
connection. Given that most 32-bit and larger systems have a PCI
bus of some form (the only exceptions are handhelds, games, and
special embedded systems), there is a PCI there to plug into. Now
with serialized PCI, two PCI-based systems can be connected via
their PCI bus implementations. Thus, PCI provides a fourth layer, a
link layer between systems.
Somehow this seems to be the right time for Serialized PCI;
implementations seem to be coming out of the woodwork. Here are
some variations, and more are on the way:
- PLX GigaBridgePLX adaptive Switch Fabric supports up to
224 PCI bus segments, one per bridge chip. Based on Seebring Ring
technology (announcement later this year).
- National Instruments MAXI-3Card set to link PCI, cPCI,
and PXI bus systems for test and operations (available).
- SBS Data BlizzardCard set links 64-bit, 66-MHz
PCI/cPCI/PMC connection. Includes DMA engine that delivers with
80-MB/sec sustained data rate (available).
- Texas Instruments PCI-to-PCI Bridge Transceiver Chip
SetSupports 132-MB/sec connection over copper TWP
Today there is a shift going on from parallel buses to serial
buses and crossbars, the switch fabric technology. High-speed
serial buses and switch fabrics are taking up the intersystem
connection load for next generation systems (these may be pseudo
serial buses with up to 16-bit datapaths). This is happening now.
Current mainstream serialized switch fabric efforts include the PC
industry's InfiniBand and Motorola/Mercury Computer's RapidIO. In
that arena, serialized PCI may well provide a low-end floor to
those switch fabricsswitching serialized PCI (or PCI-X) for
intersystem connections. In fact, PLX's GigaBridge chip provides a
PCI-based switch fabric based on an underlying ring technology.
So today, you can get an adjunct bus or an intersystem
connection on the cheap with a serialized PCI. And tomorrow, you
may well be able to upgrade to a switched fabric, which runs
serialized PCI or PCI-X. But the combination of PCI interfaces and
a serial connection between them seem to be here to stay.