Internet Data Sheet  
					HYS72T[64/128/256]4[00/20]HFA–[2.5/3S/3.7]–B  
					Fully-Buffered DDR2 SDRAM Modules  
					3.4  
					High-Speed Differential Point-to-Point Link (at 1.5 V)  
					Interfaces  
					The Advanced Memory Buffer supports one FB-DIMM  
					Channel consisting of two bidirectional link interfaces using  
					highspeed differential point-to-point electrical signaling. The  
					southbound input link is 10 lanes wide and carries commands  
					and write data from the host memory controller or the  
					adjacent DIMM in the host direction. The southbound output  
					link forwards this same data to the next FB-DIMM. The  
					northbound input link is 14 lanes wide and carries read return  
					data or status information from the next FB-DIMM in the chain  
					back towards the host. The northbound output link forwards  
					this information back towards the host and multiplexes in any  
					read return data or status information that is generated  
					internally. Data and commands sent to the DRAMs travel  
					southbound on 10 primary differential signal line pairs. Data  
					received from the DRAMs and status information travel  
					northbound on 14 primary differential pairs. Data and  
					commands sent to the adjacent DIMM upstream are repeated  
					and travel further southbound on 10 secondary differential  
					pairs. Data and status information received from the adjacent  
					DIMM upstream travel further northbound on 14 secondary  
					differential pairs.  
					3.4.1  
					DDR2 Channel  
					The DDR2 channel on the Advanced Memory Buffer supports  
					direct connection to DDR2 SDRAMs. The DDR2 channel  
					supports two ranks of eight banks with 16 row/column  
					request, 64 data, and eight check-bit signals. There are two  
					copies of address and command signals to support DIMM  
					routing and electrical requirements. Four transfer bursts are  
					driven on the data and check-bit lines at 800 MHz.  
					Propagation delays between read data/check-bit strobe lanes  
					on a given channel can differ. Each strobe can be calibrated  
					by hardware state machines using write/read trial and error.  
					Hardware aligns the read data and check-bits to a single core  
					clock. The Advanced Memory Buffer provides four copies of  
					the command clock phase references (CLK[3:0]) and write  
					data/check-bit strobes (DQSs) for each DRAM nibble.  
					3.4.2  
					SMBus Slave Interface  
					The Advanced Memory Buffer supports an SMBus interface  
					to allow system access to configuration registers independent  
					of the FB-DIMM link. The Advanced Memory Buffer will never  
					be a master on the SMBus, only a slave. Serial SMBus data  
					transfer is supported at 100 kHz. SMBus access to the  
					Advanced Memory Buffer may be a requirement to boot and  
					to set link strength, frequency and other parameters needed  
					to insure robust configurations. It is also required for  
					diagnostic support when the link is down. The SMBus  
					address straps located on the DIMM connector are used by  
					the unique ID.  
					3.4.3  
					Channel Latency  
					FB-DIMM channel latency is measured from the time a read  
					request is driven on the FB-DIMM channel pins to the time  
					when the first 16 bytes (2nd chunk) of read completion data is  
					sampled by the memory controller. When not using the  
					Variable Read Latency capability, the latency for a specific  
					DIMM on a channel is always equal to the latency for any  
					other DIMM on that channel. However, the latency for each  
					DIMM in a specific configuration with some number of DIMMs  
					installed may not be equal to the latency for each FB-DIMM  
					in a configuration with some different number of DIMMs  
					installed. As more DIMMs are added to the channel,  
					additional latency is required to read from each DIMM on the  
					channel. Because the channel is based on the point-to-point  
					interconnection of buffer components between DIMMs,  
					memory requests are required to travel through N-1 buffers  
					before reaching the Nth buffer. The result is that a 4 DIMM  
					channel configuration will have greater idle read latency  
					compared to a 1 DIMM channel configuration. The Variable  
					Read Latency capability can be used to reduce latency for  
					DIMMs closer to the host. The idle latencies listed in this  
					section are representative of what might be achieved in  
					typical AMB designs. Actual implementations with latencies  
					less than the values listed will have higher application  
					performance and vice versa.  
					Rev.1.01, 2007-06-20  
					10062006-RQWY-GI6S  
					14