Predator PCI/AGP board for the A1200 tower and A4000 Desktop
Overview
The Predator busboard is a design commissioned by Eyetech to allow full advantage
to be taken of the high speed local bus built into every Cyberstorm PPC and Blizzard
PPC accelerator to enable PC-type PCI expansion and graphics cards to be used on
a PPC-equipped Amiga.
The response to our Predator announcement in August 2000 has been overwhelming and
entirely positive. However, many potential purchasers have asked us if it would be
possible to incorporate a G3/G4 processor slot for future expansion, along with
the AGP and SDRAM socket of the original specification. Our engineers have
confirmed that this is possible, but that it would require the reworking of some
of the complex logic associated with the AGP and SDRAM interface implementations
of the Predator. Nevertheless we believe that this step is very worthwhile and
will be enormously beneficial to potential Predator purchasers wishing to achieve
the very best levels of performance from their machines. We have therefore decided
to produce two levels of Predator board for both the A1200 and A4000, to suit
different budgets and performance requirements, the Predator-SE and the
Predator-Plus.
All Predator models therefore represent a high performance solution for owners
of phase5/DCE PPC accelerators only. This contrasts with the superficially
similar A1200-only Mediator design from Elbox. The Mediator is designed to
communicate over the 8MB wide Zorro2 address space, and at A1200 expansion
connector bus speeds, and as a consequence is therefore hardware compatible
with 68000 series accelerators. Subject to satisfactory performance testing
Eyetech will also stock the Mediator board for less demanding configurations
and/or users who do not intend to purchase a Predator-compatible PPC
accelerator.
The Predator will be available in two different form factor designs, one for
the A4000 desktop and one for the A1200. The A1200 form factor version will
come with 3xPCI and 1 x AGP simultaneously usable slots and will fit in any
tower capable of taking a Z4 busboard. The A4000 desktop form factor version
will come with 2xPCI and one AGP slots plus one shared PCI and regular, Video
Toaster-compatible A4000 video/Zorro3 slot (making 4 simultaneously usable
slots altogether). Both form factor versions will also have provision for a
G3/G4 PPC & SDRAM slot.
Each form factor design will be available in two versions, the Predator-SE
and the Predator-Plus. The Predator-SE will be a low cost version supporting
PCI cards only, albeit over the high speed PPC local bus connector (see
below). The Predator-Plus will be the fully featured version with AGP, G3/G4
and SDRAM slots activated. A trade-in/upgrade option will be available for
purchasers of the Predator-SE who subsequently wish to add Predator-Plus
facilities.
Development Timeframes
The PCB design for both the A4000 form factor versions (ie Predator-SE and
Predator-Plus) is complete, with the A1200 PCB versions at about the 70%
level and expected to be complete by the end of September 2000.
The logic design and simulation is complete for both A1200 and A4000 Predator-SE
versions (which are logically identical) . Testing of the A4000 Predator-SE
hardware is expected to be underway at the end of September/early October with
retail shipments expected to start about 5 weeks later. The A1200 version will
follow about 4 weeks behind the A4000 product.
We are currently planning to ship the A1200/A4000 Predator-Plus versions
(with AGP/G3/G4/SDRAM slot) in early 2001.
Technical design philosophy
The following notes help to illustrate the technical approach we have adopted
in designing the Predator expansion board and how it differs from the
superficially similar Mediator design from Elbox.
Accelerator to PCI interface
In designing the (PPC) accelerator to PCI/AGP bus interface we had two
possible alternatives. We could use the existing Blizzard PPC/CyberstormPPC
local bus interface normally used for connecting Permedia-type graphics cards
(which is very fast but restricts the use to PPC-equipped Amigas only).
Alternatively we could transfer data over the existing A1200 expansion
interface (edge connector) or Zorro3 bus (A4000), an approach that would
allow compatibility with 68000 series accelerators. This latter approach has
(in our opinion) several major performance disadvantages over the PPC local
bus method. The relative merits of both approaches are as follows:
- PPC local bus interface
The specifications of the local bus controller chip used in the
Blizzard/CyberstormPPC show that it is capable of transferring data across
the local bus connector at 50MB/s in single byte access mode or up to 80MB/s
in burst access mode. In addition, as the full 4GB address space is linearly
decoded there is no register-setting overhead (as with a paged memory data
transfer system) and DMA facilities can be implemented fairly easily. Finally,
as the interface for both BlizzardPPC and CyberstormPPC local bus connectors
are identical, one Predator logic design can be used for both A1200 and A4000
implementations.
- Amiga expansion connector interface
The alternative method of transferring data - that is over the A1200 expansion
bus - must use a paging method to map the (maximum) 8MB of Zorro2 address
space to the 4GB of PCI address space. Using paged memory means that a
paging register has to be set at the start of each data transfer and whenever
a data transfer block crosses a page boundary. For small block transfers -
such as when the CPU is updating a graphics card - this overhead can be
significant. For single byte transfers it reduces the actual transfer speed
to exactly half the nominal speed.
In addition to this, the speed of the A1200 expansion bus data transfer rates
are much lower than those for the PPC local bus interface. The A1200 expansion
bus is clocked at 14.2 MHz and requires a minimum of 3 clock cycles to transfer
each word of up to 4 bytes. In the worst case - that of a single byte transfer
to a random PCI address - the data transfer rate (including setting the paging
register) is therefore just 2.36MB/s. This rises to a maximum of 18.9MB/s for
block transfers within the same memory page. These rates assume an optimal bus
transfer mechanism with no wait states. In practice an asychronous interface is
usually necessary to match the CPU and expansion connector bus speeds resulting
in the addition of wait states during data transfer thus further slowing actual
data transfer speeds. For technical reasons it is not practical to implement a
‘burst’ data transfer mechanism (within the PCI definition of ‘burst’ mode)
over the A1200 edge connector, with existing accelerator designs.
A further compromise is in the size of the the Zorro2 address ‘window’ used
for memory paging. An 8MB window (ie as large as is possible) will tend to
minimise the amount of paging overhead - but has the disadvantage of disabling
non-relocatable Zorro2 addresses - such as the A1200’s PCMCIA slot. A smaller -
say 4MB - paging window imposes more overhead but allows (for example) full
PCMCIA functionality.
Based on the above we believe that there are overwhelming performance merits in
using the PPC local bus to interface to the Predator board.
Slot-to-Slot data transfer speeds
Provided that the busboard is properly designed - which at high speeds means
separate ground and power planes and trace length equalisation - there is no
limitation on the card-to-card transfer speeds other than that imposed by the
PCI cards themselves - irrespective of how the PCI to CPU bus is implemented.
Having said that, a high card-to-card transfer rate is only of limited
functionality, and is used mainly (in PC’s) for overlaying TV card etc output
directly on a graphics screen.
DMA
DMA (Direct Memory Access) is the mechanism whereby a peripheral card (eg a
network card) can transfer its data directly into a processor-accessible
memory area without cpu intervention. All PCI/AGP cards with ‘bus mastering’
facilities have the potential to carry out DMA both between cards (as above)
and - more importantly - between card and CPU memory. However PCI cards are
designed to carry out DMA in a linear (ie non-paged) cpu address space, making
it extremely difficult - if not impossible - to implement on an paged memory
interface using Zorro2 address space.
AGP vs PCI
Why did we choose to implement AGP on the Predator-Plus? Well apart from AGP
graphics cards being the defacto standard in the PC market, and therefore much
more easily available than their PCI equivalents, an AGP graphics card has two
significant performance advantages over its PCI equivalent. It is therefore an
appropriate choice of technology for the highest performance expansion boards.
These advantages are:
- Protocol & bus mastering
The AGP protocol is designed to be much more efficient than PCI at moving the
sort of data used by graphics cards - particularly in bus mastering mode when
transfering data directly to the cpu’s memory area. Amongst other things it
eliminates PCI bus inefficiencies associated with non-deterministic data
transfer lengths (which can be significant in graphics card operations).
This results in much more usable data being transferred within a given time
slot than for a PCI card running at the same nominal bus speed.
- PCI card compatibility
Because of the AGP protocol enhancements (over PCI) an AGP card will continue
to operate at full bus speed even when there are slower PCI cards on the bus
(eg modem cards). The PCI protocol by contrast automatically drives the whole
bus at the speed of the slowest card. In addition, by using the direct PPC
local bus interface, the graphics card can effectively use CPU memory as its
own for increased performance under certain circumstances.
Finally it is worth pointing out that, in all versions of the Predator, all
the PCI and AGP protocols are implemented in hardware and thus put no load on
the (PPC) cpu during associated data transfers.
G3/G4/SDRAM
The CPU / SDRAM slot provided on the Predator-Plus is designed to take a
plug-in board carrying a G3 / G4 CPU and/or SDRAM memory with minimal additional
support circuitry. This approach allows us to keep the Predator CPU/memory
upgrade costs low, and is possible because the Cyberstorm PPC / Blizzard PPC
accelerator (which is still needed) is used to provide the relatively complex
local-bus-to-Amiga-motherboard interfacing logic.
The Predator for A4000 tower systems and A3000 desktop and tower systems.
Our current plans are to bring the A4000 desktop and A1200 tower versions of
the Predator-SE and Predator-Plus to market before undertaking any further
design work for additional Amiga platforms. We will however start conducting
feasibility studies for the introduction of Predator expansion boards for the
A4000 Tower (both A4000 desktop conversions and Amiga Technologies A4000T’s)
and for the Commodore manufactured A3000 desktops and towers once the Predator-SE
for A1200 and A4000 systems are shipping in volume. The priority with which
we tackle these new designs will depend on the technical/mechanical complexity
of the design and the perceived demand for the products.
So if you are interested in potentially purchasing a Predator for any of
these types of systems it would be enormously helpful if you could
email our sales
address indicating in the subject line the type of system you would like to use
the Predator with. It does not matter at this stage whether you are interested
in the Predator-SE or Predator-Plus models.
Your subject line should read "Predator for xxxxx"
where "xxxxx" is:
- "A4000 desktop tower conversion using (manufacturers name) tower"
or
- "A4000T tower Amiga International design"
or
- "A3000 desktop system by Commodore"
or
- "A3000 tower system by Commodore"
Thanks for your help.
|
