Volume 1, Issue 2

September, 2002

• AGP 8x
  AGP Specs
  Slot Diagrams
• Gigabit Ethernet
  Copper
  Fiber
• Optical Storage Technology Guide
  Feature Matrix

• PCI to PCI-X
  Advantages
  Specs
• Server Technologies
  PCI Subsystem
  Memory Subsystem
  Storage Subsystem
• New Memory Technology
  DDR / Dual Channel DDR
  RIMM4200/ RIMM3200 RDRAM

PCI Subsystem:

• PCI-X: The PCI-X specification, an extension of the PCI 2.2 specification, was driven by the growing need for I/O bandwidth. The PCI 2.2 specification provides a maximum throughput of 528 MB/s of shared bandwidth on a given bus. With new I/O technologies, such as Gigabit Ethernet, Gigabit Fibre Channel and multi-channel Ultra320 SCSI, the 528 MB/s bandwidth of PCI 2.2 is fast becoming a constraint in servers.

The PCI-X specification supports a 64-bit bus operating at speeds of 66 MHz, 100 MHz and 133 MHz, providing throughput speeds of 528 MB/s, 800 MB/s and 1064 MB/s respectively. PCI-X also supports hot-swap and is fully backward compatible to PCI 2.2 devices. Be aware that, as is the case with PCI 2.2, the PCI-X bus operates at the frequency of the lowest frequency card connected to the bus.

        - A PCI-X bus operating at  66 MHz can support up to four devices.

        - A PCI-X bus operating at 100 MHz can support a maximum of two devices

        - A PCI-X bus operating at 133 MHz can support only a single device.

• PCI Hot-plug (PHP): is the ability to add or replace a PCI adapter card while the system is running. This provides increased availability because the system doesn’t need to be powered down to replace a failed card. Hot-plug types include:

       - Hot replace: The process of removing an adapter card and then inserting an identical adapter into the same slot. The replacement adapter card will use the same PCI resources that were assigned to the previous card, and its driver does not need to be updated. Sometimes referred to as "Like-for-Like Replacement."

       - Hot add: The process of inserting an adapter card into a previously unoccupied slot. This operation requires a driver for the added adapter and a reserved PCI resource by the system BIOS. Sometimes referred to as hot expansion.

       - Hot upgrade: The process of removing an adapter card and inserting an upgraded adapter that requires different PCI resources than the original card. The adapter’s driver may or may not use the same driver as the previous adapter.

Notes:

1. The Intel Quad Xeon MP Server system SPSH4 (ASI SKU: 17382)  and SRSH4 (ASI SKU: 17249) support the above hot-plug PCI types, but not all operating systems do.  

2. To support PCI hot-plug, systems require hot-plug hardware, a hot-plug compatible operating system, and hot-plug capable adapter drivers. To ensure backward compatibility, a combination of hot-plug and conventional versions of each of these components is permitted, including mixing both hot-plug and conventional adapter drivers. If a conventional driver is loaded under a hot-plug capable operating system, or a hot-plug driver is loaded under a conventional operating system, the driver has the capability it always had in the conventional environment.

Memory Subsystem:

• Single-bit error correction: If a single-bit error is detected, the ECC logic generates a new "recovered" 64-bit QWord with a pattern that corresponds to the originally received 8-bit ECC parity code. The corrected data is returned to the requestor, most likely the processor or a PCI master.

• Multi-bit error detection: Additional errors within the same QWord constitute a multi-bit error, which maybe unrecoverable. In the case of a multi-bit memory error, a non-maskable interrupt (NMI) is issued that instructs the system to shut down to avoid data corruption. Multi-bit errors are very rare.

• Memory scrubbing: Error correction is performed on data being read from memory. The correction is then passed to the requestor and at the same time the error is "scrubbed" or corrected in main memory. Memory scrubbing prevents the accumulation of single-bit errors in main memory that would then become unrecoverable multi-bit errors. 

• "Chipkill": Chipkill is the ability of the memory system to withstand a multi-bit failure within a DRAM device, including a failure that causes incorrect data on all data bits of the device. When "x4" memory is installed the ECC function can detect and correct a four-bit error caused by a single failed memory chip and the system continues to function, though system performance will be affected. When "x8" memory is installed the ECC function will detect an eight-bit error caused by a single failed memory chip but will not be able to correct the error. In this situation a fatal error will be issued.

• Memory Interleaving (2-Way & 4-Way): Memory Interleaving is an advanced technique used by high-end server motherboards and chipsets to improve memory performance. Memory interleaving increases bandwidth by allowing simultaneous access to more than one chunk of memory. Interleaving works by dividing the system memory into multiple blocks. This improves performance because the processor can transfer more information to/from memory in the same amount of time, and helps alleviate the processor-memory bottleneck that is a major limiting factor in overall performance. This design requires that DIMMs operate in pairs (2-way) or quads (4-way), where a pair consists of two specific DIMM sockets to provide the aggregate 144-bit wide memory data path, and where a quad consists of four specific DIMM sockets to provide the aggregate 288-bit wide memory data path. Each block of memory is accessed using different sets of control lines, which are merged together on the memory bus. When a read or write has begun to one block, a read or write to other blocks can be overlapped with the first one. The more blocks, the more that overlapping can be done.

Note: Current server chipsets that support all of the above memory features are the Intel E7500, & Serverworks GC-HE & GC-LE.

Storage Subsystem:

• Zero-channel RAID: To use this function, one PCI slot on the motherboard must be specially wired for RAID I/O Steering (RAIDIOS). The RAID adapter is dependent on the baseboard’s SCSI controller to transfer data to the system’s storage devices. RAIDIOS has two components. The first is IRQ steering, which re-routes the IRQs of the onboard SCSI controller. The second is the initialization device select (IDSEL) signal, which allows the ZCR controller to take control of the SCSI device. The ability to control the IDSEL signal is what allows this newer version of zero-channel RAID to be Windows Hardware Quality Labs (WHQL) certified.  ASI carries Adaptec and Intel based Zero-Channel RAID cards.
 

As you may have noticed, the memory names are based on the amount of bandwidth that they have. Dual Channel DDR gives you dual bandwidth. For example, PC2100 Dual Channel DDR gives you 4.2GB/s.  Current server chipsets like the Intel E7500 and Serverworks GC-LE utilize Dual Channel DDR 266/200 ECC Registered memory ONLY!!

RIMM4200, RIMM3200 RDRAM
     => Same Performance as PC1066/PC800 respectively but only a single module needed

Unlike DDR memory, the new RDRAM modules are changing in design technology as well as increasing in speed. The latest RDRAM developments, RIMM4200 and RIMM3200, work by  combining either two PC1066 or PC800 modules onto a single 32-bit RIMM module. Of course, the most obvious benefit to 32-bit RDRAM is the ability to upgrade with a single RIMM instead of two. Present RDRAM based PCs are found with 16-bit single channel modules (PC800 or PC1066) that has at least two modules working as a pair. In contrast, the RIMM4200 module provides the same performance with only a single module, greatly reducing required number of parts such as sockets and continuity RIMM on the motherboard.

As you can see below, in order to avoid mixing the 32-bit RIMMs with older, incompatible 16-bit slots, Rambus has keyed RIMM4200 and RIMM3200 with 232 pins instead of 184 like the earlier RDRAM modules such as the PC800.

PC800 RDRAM	PC3200 RDRAM

An example of a MB that can accept either RIMM4200 or RIMM3200 memory is the ASUS P4T533 with the Intel 850E chipset. This MB has 2 RIMM sockets and can utilize RIMM4200 or RIMM3200 memory. The included C-RIMM (Continuity RIMM) must be used in the unpopulated RIMM socket if applicable.

Note: RIMM4200 modules require the use of 533 MHz FSB processors 
 

See You Next Month!