PB INTERCONNECTIONS


Each Processing Board has three different kinds of interconnections plus power supply and the global free-running clock:

 3D data connections

There are two bi-directional channels (one for each possible direction) in each of the three dimensions (X Y Z).

Each channel is 80 bits wide and supports a transfer rate of 33-66 MHz (it is driven by the slow clock).

Data transfers are point-to-point and EDAC corrected.





 Global signals

The PB gives the following signals to the external world:

The PB receives the following signals from the Root Board and from the host:





 THE ROOT BOARD

The Root Board, the board which assures the global synchronisation of the Processing Boards - a critical feature of a SIMD machine, will receive local signals from the PBs and send back global signals to them as described in the previous section.

The signals glb_ifstatus, gbl_halt, and glb_CNB are set active if at least one of the corresponding local signals is active.The glb_clock_enable signal is always high unless one of the clock_stretch signals is active.The refresh signal is delivered to all the PBs according to the clock cycle and to the memory constraints.



 THE APE CHANNEL

Each PB is connected to the host network through the APE Channel. The APE channel is a 32 bits wide (+ EDAC), 33MHz, synchronous interface. All the PB's are slaves of the APE Channel. The APE Channel is in principle capable of a 133Mbyte/s performance (33MWords/s, 4 bytes each).Tmille interfaces the PB to the APE Channel. It waits for a valid control word on the APE channel. The control word contains the BoardId number to which the operation is requested, the operation code, the burst size and the direction of the transfer. When Tmille decodes an APE Channel control word it checks the BoardId in the control word against its own hard-wired SlotId or it checks if the control word is a broadcast request (directed to all the PBs). Shouldn't any of these situations occur, it ignores the control word and follows the transaction protocol without performing any PB operation. The second word in the protocol contains the address of the board device involved in the data transfer, while the data stream follows in the proper direction.




 THE HOST PCI INTERFACE BOARD

The APEmille Processing Boards will be interfaced with a network of computers equipped with Peripheral Computer Interconnect (PCI) I/O bus, which seems the emerging bus standard for workstations (133Mbyte/s). These computers will be referred to as APEmille Host Processors (AHP).

A Synchronous Interface will be implemented to connect the APE Channel to the PCI bus and a single AHP will be able to drive a set of Processing Boards (the ones connected to the same APE Channel; we call this set of boards an APEmille Unit: AU). The physical connection with the AHP is provided by a PCI interface board which will be plugged in one PCI slot in the AHP. Depending on the application requirements the number of PB driven by each AHP (i.e. the number of board in an AU) will be chosen to satisfy the needed I/O data rates. Up to 16 PBs can be connected to the same APE channel.The AHP will be able to access all the AU memories and control registers though the APE channel.



 THE HIPPI CHANNEL

The Host PCI Interface Board will host a driver for a HIPPI (High Performance Parallel Interface) channel.

Data coming from the APE Channel can be delivered either to the PCI host or to the HIPPI channel which can be connected to a high performance device such as a disk pack. The host drives the protocol initialisation with the PB and then switches the APE Channel path towards the HIPPI interface.Vice versa, the host can drive the PCI board so that after the protocol initialisation the boards can read data from the HIPPI channel.

This will allow very high data transfer performances. HIPPI is a mature communication technology with a rich I/O infra-structure developed for mainframes.




 THE HOST NETWORK

A host network will be implemented for APEmille configurations having more than 16 PB. It will be a TCP/IP network based on any commercial interconnection e.g. Ethernet (10Mbit/s), Fast Ethernet (100Mbit/s), FDDI (100Mbit/s) or ATM (155 or 622Mbit/s). Each AHP will be able to access its own disks and peripherals as well as other AHP and network resources.