Server Chipsets

This chapter is from the book 

Intel Pentium Pro/II/III Chipsets for Servers

Intel was the leading vendor of chipsets for its P6 processor families, which included the Pentium Pro, Pentium II, and Pentium III. Table 3.6 shows the Intel chipsets used on Pentium Pro motherboards. All the chipsets shown in Table 3.6 were designed to be suitable for use in server applications. However, most systems using these chipsets have been retired.

Table 3.6. Intel Pentium Pro Motherboard Chipsets (North Bridge) ^[1]

Feature	450KX	450GX	440FX
Codename	Mars	Orion	Natoma [2]
Date introduced	Nov. 1995	Nov. 1995	May 1996
Bus speed	66MHz	66MHz	66MHz
SMP (dual CPUs)	Yes	Yes (up to 4) [3]	Yes
Memory types	FPM	FPM	FPM/EDO/BEDO
Parity/ECC	Both	Both	Both
Maximum memory	1GB	4GB	1GB
L2 cache type	In CPU	In CPU	In CPU
Maximum cacheable	1GB	1GB	1GB
PCI support	2.0	2.0	2.1
AGP support	No	No	No
AGP speed	n/a	n/a	n/a
South Bridge	PIIX3	PIIX3	PIIX3

Intel Pentium II/III chipsets that were suitable for use in servers are shown in Tables 3.7 and 3.8. 4xx series chipsets incorporate a North Bridge/South Bridge architecture (Table 3.7), whereas 8xx series chipsets support the newer and faster hub architecture. P6/P7 (Pentium III, Pentium 4, and Xeon) processor chipsets using hub architecture are shown in Table 3.8.

Table 3.7. P6 Processor Chipsets Using North Bridge/South Bridge Architecture

View Table

Table 3.8. P6 (Pentium III) Server Processor Chipsets Using Hub Architecture ^[1]

View Table

Most recent Intel chipsets for single-processor or dual-processor servers are designed as two-part systems, using a North Bridge (MCH or GMCH in hub-based designs) and a South Bridge (ICH in hub-based designs) component. Often the same South Bridge or ICH component can be used with several different North Bridge (MCH or GMCH) chipsets. Table 3.9 shows a list of all the Intel South Bridge components used with P6-class processors and their capabilities. The ICH2 is also used as part of some of the first seventh-generation (Pentium 4) Intel chipsets.

Table 3.9. Intel South Bridge/ICH Chips for P6 Class CPUs ^[1]

View Table

The following sections examine the server-class chipsets for P6 processors up through the Pentium III.

The Intel 450KX/GX (Mars/Orion) Chipsets

The first chipsets to support the Pentium Pro were the 450KX and GX. Although both are commonly known as Orion, the 450KX was originally known as Mars. The 450KX was designed for networked or standalone workstations and is also suitable for low-end servers; the more powerful 450GX was designed for servers. The GX server chipset was particularly suited to the server role because it supports up to four Pentium Pro processors for SMP servers, up to 8GB of four-way interleaved memory with ECC or parity, and two bridged PCI buses. Some vendors, such as ALR, with its Revolution 6x6, designed systems that could use up to six processors using the GX chipset. The 450KX is the low-end server or workstation (standalone user) version of Orion and, as such, it supports fewer processors (one or two) and less memory (1GB) than the GX. The 450GX and 450KX both have full support for ECC memory—a requirement for server and workstation use.

	See "Pentium Pro Processors," p. 87.

The 450GX and 450KX North Bridge comprises four individual chip components: an 82454KX/GX PCI bridge, an 82452KX/GX data path (DP), an 82453KX/GX data controller (DC), and an 82451KX/GX memory interface controller (MIC). Options for QFP or BGA packaging were available on the PCI Bridge and the DP. BGA uses less space on a board.

The 450's high reliability was obtained through ECC from the Pentium Pro processor data bus to memory. Reliability is also enhanced through parity protection on the processor bus, control bus, and all PCI signals. In addition, single-bit error correction is provided, thereby avoiding server downtime because of spurious memory errors caused by cosmic rays.

	See "Parity and ECC," p. 389.

Until the introduction of the following 440FX chipset, these chipsets were used almost exclusively in fileservers. After the debut of the 440FX, the expensive Mars/Orion chipsets all but disappeared due to their complexity and high cost.

The Intel 440FX (Natoma) Chipset

The first popular mainstream P6 (Pentium Pro or Pentium II) motherboard chipset was the 440FX, which was codenamed Natoma. Intel designed the 440FX to be a lower-cost and somewhat higher-performance replacement for the 450KX workstation chipset. Although the 440FX was designed for use in workstation applications, it was also used as a low-end server chipset by numerous vendors. It offered better memory performance through support of EDO memory, which the prior 450KX lacked.

	See "Early Server RAM Types: DRAM, EDO DRAM, and SDRAM," p. 361.

The 440FX uses half the number of components that the previous Intel chipset used. It offers additional features, such as support for the PCI 2.1 (concurrent PCI) standard, support for USB 1.1 ports, and reliability through ECC.

The concurrent PCI processing architecture maximizes system performance with simultaneous activity on the CPU, PCI, and ISA buses. Concurrent PCI provides increased bandwidth to better support 2D/3D graphics, video and audio, and processing for host-based applications. ECC memory support delivers improved reliability to business system users.

The main features of this chipset included the following:

• Support for up to 1GB of EDO memory
• Full 1GB cacheability (based on the processor because the L2 cache and tag are in the CPU)
• Support for USB 1.1
• Support for bus master IDE
• Support for full parity/ECC memory

The 440FX consists of a two-chip North Bridge. The main component is the 82441FX PCI bridge and memory controller, along with the 82442FX data bus accelerator for the PCI bus. This chipset uses the PIIX3 82371SB South Bridge chip that supports high-speed bus master DMA IDE interfaces and USB, and it acts as the bridge between the PCI and ISA buses. Figure 3.7 illustrates the design of the 440FX.

Note that the 440FX was the first P6 chipset to support EDO memory, but it lacked support for the faster SDRAM memory. Also, the PIIX3 South Bridge used with this chipset does not support the faster Ultra DMA IDE hard drives.

The 440FX was the chipset used on the first Pentium II motherboards, which have the same basic architecture as the Pentium Pro. The Pentium II was released several months before the chipset that was supposedly designed for it was ready, so early PII motherboards used the older 440FX chipset. However, this chipset was never designed with the Pentium II in mind, whereas the newer 440LX was optimized specifically to take advantage of the Pentium II architecture. When the 440LX was introduced, the 440FX was quickly superseded.

The Intel 440LX Chipset

The 440LX quickly took over in the marketplace after it was introduced in August 1997. This was the first chipset to really take full advantage of the Pentium II processor. The 440LX chipset was the first Intel Pentium II chipset to use a single-chip North Bridge design, setting a design standard that would be followed by subsequent designs. The 82443LX North Bridge chip incorporated the features that required two chips in its immediate predecessor, the 440BX, and added support for two then-new technologies, AGP video and 66MHz synchronous DRAM (SDRAM). The 440LX's South Bridge, the PIIX4, was also a new design, adding support for Ultra DMA 33 ATA/IDE drives.

	See "Pentium II Processors," p. 90.

The 440LX chipset's major features included the following:

• Single-chip North Bridge design (82443LX chip)
• Support for the (then-new) AGP video card bus
• Support for 66MHz SDRAM memory
• Support for the Ultra DMA ATA/IDE interface (UDMA/33)
• Support for USB 1.1 ports

The 440LX's design was flexible enough to support all types of Pentium II systems, from two-way servers to desktop computers. It was the most popular chipset for Pentium II systems from late 1997 through spring 1998.

The Intel 440BX Chipset

The Intel 440BX chipset, introduced in April 1998, was the first chipset to run the processor host bus (often called the FSB) at 100MHz. The 440BX was designed specifically to support the faster Pentium II/III processors at 350MHz and higher. The main change from the previous 440LX to the BX is that the 440BX chipset improves performance by increasing the bandwidth of the system bus from 66MHz to 100MHz. Because the chipset can run at either 66MHz or 100MHz, it allows one basic motherboard design to support all Pentium II/III processor speeds based on either the 66MHz or 100MHz processor bus.

Here are the Intel 440BX highlights:

• Support for 100MHz SDRAM (PC100); the now-common PC133 RAM can also be installed, but it will still run at just 100MHz
• Support for both 100MHz and 66MHz system and memory bus designs
• Support for up to 1GB of memory in up to four banks (four DIMMs)
• Support for ECC memory
• Support for ACPI power management

	See "Pentium III Processors," p. 93.
	See "ACPI," p. 290.

The Intel 440BX consists of a single North Bridge chip called the 82443BX host bridge/controller, paired with a new 82371EB PCI-ISA/IDE Xcelerator (PIIX4E) South Bridge chip. This South Bridge adds support for the ACPI specification version 1.0 to the features of its predecessor, the PIIX4. Figure 3.8 shows a typical system block diagram using the 440BX.

The 440BX was a popular one-way and two-way server chipset during 1998 and into 1999. It offered superior performance and high reliability through the use of ECC, SDRAM, and DIMMs.

The Intel 440GX Chipset

The Intel 440GX AGP set was the first chipset optimized for high-volume midrange workstations and lower-cost servers, and it was Intel's first chipset for the server/workstation version of the Pentium II, the Pentium II Xeon. The 440GX also supports the Pentium III Xeon processor. The 440GX is essentially a version of the 440BX that has been upgraded to support the Slot 2 (also called SC330) processor slot for the Pentium II/III Xeon processor. The 440GX can still be used in Slot 1 designs, as well. It also supports up to 2GB of memory, twice that of the 440BX. Other than these items, the 440GX is essentially the same as the 440BX. Because the 440GX is core compatible with the 440BX, motherboard manufacturers could quickly and easily modify their existing Slot 1 440BX board designs into Slot 1 or 2 440GX designs.

The main features of the 440GX include the following:

• Support for Slot 1 and Slot 2
• Support for 100MHz system bus
• Support for up to 2GB of SDRAM memory

This chipset allows for lower-cost, high-performance workstations and servers using the Slot 2–based Xeon processors.

	See "Pentium II Xeon," p. 91.
	See "Pentium III Xeon," p. 98.

The Intel 450NX Chipset

The 450NX chipset (originally known as the 440NX) was designed for multiprocessor systems and standard high-volume servers based on the Pentium II/III Xeon processor. The Intel 450NX chipset consists of four components: the 82454NX PCI expander bridge (PXB), 82451NX memory and I/O bridge controller (MIOC), 82452NX RAS/CAS generator (RCG), and 82453NX data path multiplexer (MUX). As Table 3.10 shows, a full implementation of the 450NX uses two or more of the PXB, RCG, and MUX chips.

Table 3.10. Details of the 450NX Chipset

View Table

The 450NX supports up to four Pentium II/III Xeon processors at 100MHz. Two dedicated PCI expander bridges can be connected via the expander bus. Each PXB provides two independent 32-bit, 33MHz PCI buses, with an option to link the two buses into a single 64-bit, 33MHz bus.

Figure 3.9 shows a typical high-end server block diagram using the 450NX chipset.

The 450NX supports one or two memory cards. Each card incorporates an RCG chip and two MUX chips, in addition to the memory DIMMs. Up to 8GB of memory is supported in total.

The primary features of the 450NX include the following:

• Slot 2 (SC330) processor bus interface at 100MHz
• Support for up to four-way processing
• Support for two dedicated PCI expander bridges
• Support for up to four 32-bit PCI buses or two 64-bit PCI buses

The 450NX chipset does not support AGP because high-end video is not an issue in network fileservers.

Fujitsu also used the 450NX chipset, to develop its TeamServer M800i series. This eight-way server used two 450NX chipsets linked by Fujitsu's Synfinity interconnect technology and non-uniform memory access (NUMA) memory architecture to create what is, in effect, a tightly coupled computer cluster in a single unit.

	See "NUMA Multiprocessing," p. 38.

The Intel (Corollary) Profusion Chipset

The Profusion chipset, which supports eight-way Pentium III Xeon servers, is extremely different from most other Intel chipsets for servers. As we have seen, most Intel chipsets for Pentium II/III Xeon servers use some form of the North Bridge/South Bridge chipset architecture originally developed for use with the Intel 486 processor. However, the greater complexity of four-way and larger multiprocessor server architectures using Intel processors requires the use of several specialized chips.

Intel's first chipset to use multiple specialized chips was the 450NX for four-way Pentium II/III Xeon servers (see the section "The Intel 450NX Chipset," earlier in this chapter). Profusion's design, originally developed by Corollary in cooperation with Compaq (now a Hewlett-Packard brand), is even more sophisticated than the 450NX's because of the added challenge of supporting eight processors. Although Corollary began developing the Profusion chipset in 1996, the first systems that used Profusion were not released until late 1999, after Intel purchased Corollary in late 1997, making it a wholly owned subsidiary, and released the chipset to server developers in June 1998.

The Profusion chipset creates a five-port (dual memory banks, dual processor buses, and I/O bus) nonblocking crossbar switch using two components:

• One memory access controller (MAC) chip, which also provides a three-way processor bus bridge, support for up to 32GB of SDRAM, and TAG SDRAM management
• One data interface buffer (DIB) chip, which also provides three processor bus data ports with ECC support, two SDRAM data ports with ECC support, and concurrent data transfer on all ports and 64 cache line buffers

These two chips form a five-way crossbar switch that handles data flow between the processors, memory, and I/O bus. The combination of MAC and DIB chips in the Profusion chipset replaces the physical switch often used in other eight-way SMP architectures to interface the processors with memory and I/O bridge connections to system RAM and I/O devices such as PCI slots and PCI/ISA ports. Figure 3.10 illustrates how the MAC and DIB chips work together.

Other components of the Profusion chipset include the following:

• A 64-bit PCI bus bridge chip, the PB64 (also co-developed by Corollary and Compaq), which supports up to 8 66MHz, 64-bit PCI bus masters and up to 16 33MHz, 64-bit PCI bus masters. Up to 4 PB64 chips can be used to provide redundancy and to provide support for mixed PCI speeds. Typically, 1 of the PB64 chips is devoted to 33MHz PCI slots, and the others provide support for 66MHz slots.
• Cache coherency filters (also known as cache accelerators), which are used to improve cache memory performance on eight-way systems; one is required for each four-processor section of an eight-way configuration. A four-way system does not need a cache coherency filter.
• A South Bridge chip (Intel's PXII4E) to provide support for keyboard, mouse, USB, and serial and parallel ports.

The crossbar switch and support for multiple PCI bus bridge chips enable servers based on Profusion to handle memory, processor, cache coherency, and PCI bridge failures while continuing to operate.

Figure 3.11 shows a block diagram of a typical eight-way system based on the Profusion chipset.

From late 1999 through 2001, the Profusion chipset was the leading eight-way server chipset used with Intel server-class processors. Although the Profusion chipset itself has been discontinued, an improved version of the chipset was developed by Compaq (now owned by Hewlett-Packard) for use in servers using the Xeon MP processor. This chipset, the Hewlett-Packard F8, is covered later in this chapter, in the section "The F8 Chipset for Xeon MP."

The Intel 820 and 820E Chipsets

The Intel 8xx series chipsets introduced in 1999 represent a major departure from the 4xx-series chipsets previously used by Intel for its one-way and two-way servers. The 4xx-series chipsets used the North Bridge/South Bridge design, which use the 133MBps PCI bus to carry signals between the components. The 8xx-series chipsets replace the North Bridge with an MCH, and the South Bridge with an ICH. Intel 8xx systems replace the PCI bus interconnection between chipset components with a 266MBps dedicated bus known variously as the Accelerated Hub Architecture (AHA) or, in later systems, Hub Architecture 1.5.

	See "Intel Hub Architecture," p. 151.

The Intel 820 chipsets were designed to support Slot 1 or Socket 370 processors, such as the Pentium III and Celeron. The 820 chipset was the first to support RDRAM memory technology, a 133MHz system bus, and 4x AGP.

The 82820 MCH provides the processor, memory, and AGP interfaces. Two versions are available: One supports a single processor (82820), and the other supports two processors (82820DP), making the 820E suitable for use in low-end two-way servers. Either is designed to work with the same 82801 ICH as used with the other 800-series chipsets, such as the 810 and 840. The 820 chipset also uses the 82802 firmware hub (FWH) for BIOS storage and for the Intel random number generator (RNG).

The 820 chipset was designed to use RDRAM memory, which has a maximum throughput of up to 1.6GBps. The 820 supports PC600, PC700, and PC800 RDRAM, delivering up to 1.6GBps of theoretical memory bandwidth in the PC800 version. PC800 RDRAM is a 400MHz bus running double-clocked and transferring 16 bits (2 bytes) at a time (2x400MHzx2 bytes = 1.6GBps). Two RIMM sockets are available to support up to 1GB of total system memory.

The AGP interface in the 820 enables graphics controllers to access main memory at AGP 4x speed, which is about 1GB per second—twice that of previous AGP 2x platforms.

820 chipset features include the following:

• Single- or dual- (820E) processor support
• 100/133MHz processor bus
• Intel 266MBps hub interface
• PC800 RDRAM RIMM memory support
• AGP 4x support
• ATA-100 (820E) or ATA-66 interface
• Intel RNG
• LPC interface
• AC '97 controller
• One (820) or two (820E) USB 1.1 buses with either two or four ports, respectively

The 820 chipset consists of three main components, with a few optional extras. The main component is the 82820 (single-processor) or 82820DP (dual-processor) MCH, which is a 324 BGA chip. That is paired with an 82801 ICH, which is a 241 BGA chip. Finally, it has the 82802 firmware hub (FWH), which is really just a fancy flash ROM BIOS chip. Optionally, there can be an 82380AB PCI-ISA bridge that is used only if the board is equipped with ISA slots.

The newer 820E version uses an updated 82801BA ICH2, which supports ATA-100 and incorporates dual USB controllers with two ports each, for a total of four USB ports. Although the 820E supports dual processors, few, if any, dual-processor 820E motherboards were ever built.

The Intel 840 Chipset

The Intel 840 was a high-end chipset designed for use in high-performance multiprocessor systems using Slot 1, Slot 2 (Xeon processor), or Socket 370 processors. The 840 chipset uses the same hub architecture and modular design as the rest of the 800 family of chipsets, with some additional components that enable more performance. Figure 3.12 shows a photo of the Intel 840 chipset.

As with the other 800 series chipsets, the 840 has three main components:

• 82840 MCH— This provides graphics support for AGP 2x/4x, dual RDRAM memory channels, and multiple PCI bus segments for high-performance I/O. Equivalent to the North Bridge in older chipset designs.
• 82801 ICH— This is the equivalent to the South Bridge in older chipset designs, except that it connects directly to the MCH component via the high-speed Intel Hub Architecture bus. The ICH supports 32-bit PCI, IDE controllers, and dual USB 1.1 ports.
• 82802 FWH— This is basically an enhanced flash ROM chip that stores system BIOS and video BIOS, as well as an Intel RNG. The RNG provides truly random numbers to enable stronger encryption, digital signing, and security protocols.

In addition to the core components, parts are available for scaling up to a more powerful design. Three additional components shown in Figure 3.12 can be added:

• 82806 64-bit PCI controller hub (P64H)— Supports 64-bit PCI slots at speeds of either 33MHz or 66MHz. The P64H connects directly to the MCH using Intel Hub Architecture, providing a dedicated path for high-performance I/O. This is the first implementation of the 66MHz 66-bit PCI on a PC motherboard chipset, allowing for a PCI bus four times faster than the standard 32-bit 33MHz version.
• 82803 RDRAM-based memory repeater hub (MRH-R)— Converts each memory channel into two memory channels for expanded memory capacity.
• 82804 SDRAM-based memory repeater hub (MRH-S)— Translates the RDRAM protocol into SDRAM-based signals for system memory flexibility. This is used only in 840 systems that support SDRAM.

Figure 3.13 shows the 840 chipset architecture.

840 chipset features include the following:

• Support for one or two Pentium III or Pentium III Xeon processors
• 100/133MHz processor bus
• Dual RDRAM memory channels, operating simultaneously and providing up to 3.2GBps memory bandwidth (requires 82803 chip)
• 16-bit wide implementation of Intel Hub Architecture (HI16), which enables high-performance concurrent PCI I/O with the optional P64H component
• AGP 4x
• Prefetch cache, unique to the 840 chipset, which enables highly efficient data flow and helps maximize system concurrency
• Intel RNG
• USB 1.1 support

Optionally, network interface and RAID controller interface chips could be added as well. Both Intel and third-party vendors have used the 840 chipset to build dual-processor server motherboards.