Linuxdoc Linux Questions
Click here to ask our community of linux experts!
Custom Search
Next Previous Contents

2. Why PCI?

2.1 General overview

The PC-architecture has several BUS-Systems to choose from:

ISA

16 or 8bit, cheap, slow (usually 8Mhz), standard, many cards available>

EISA

32bit, expensive, fast, few cards available, fading>

MCA

32 or 16bit ex-IBM-proprietary, fast, becoming rare>

VESA-Local-Bus

32bit, based on 486 architecture, cheap, fast, many cards available>

PCI-Local-Bus

32bit (64 bit coming), cheap, fast, many cards available, nowadays standard>

MCA worked fine, but never achieved much market, being used on only some early IBM PS/2 machines. There were very few cards.

EISA was reliable, but rather expensive, and intended more for servers, than for the average user. It has the next fewest cards available.

VESA-Local-Bus (VLB) had some problems with high bus-speeds, and was not very reliable, but mainly due to its low price and better-than-ISA performance, sold very well. Technically, it's almost a direct map of the 486 processor bus. Most VESA boards should be stable by now. At the beginning of 1996, many 486 motherboards still support VESA, but PCI is growing. VESA busses are tied directly to the speed of the memory bus for 486's, or half the speed for Pentiums.

PCI now has the advantage. Like EISA it is not proprietary. It is as faster than EISA or MCA, and cheaper. Most current Pentium motherboards use the PCI bus; VESA is fading. Virtualy all PCI motherboards and cards sold at the beginning of 1996 are 32 bit, and run at 0-33 MHz.

Currently, most Pentium motherboards run the PCI bus at 1/2 the memory speed (ie: 33 MHz for the 66 MHz memory bus on the P66,P100,P133,P166; 30 MHz for the 60 MHz memory bus on the P60,P90,P120,P150; and 25 Mhz on the 50 MHz memory bus of the P75). This is probably true of Cyrix 6x86 motherboards too. NexGen 5x86 implemention isn't known. The PCI spec does allow the PCI bus to be run asynchronously from the processor, (eg: 33 Mhz bus on P75), but this is not common yet.

PCI 2.1 has been defined, allowing 64 bit PCI, and/or 0-66 MHz operations, but no x86 chipsets yet support these options. 64 bit PCI will probably appear first, in 32/64 bit dual compatible versions. That is, you will be able to mix 32 and 64 bit cards. 66 MHz PCI will take longer, as it's technically demanding, can only support one or maybe two slots per bridge, and may not work well with 33 MHz cards.

PCI is not processor dependent like the VESA Local-Bus. This means you can use the winner-1000-PCI in an Alpha-driven-PCI computer as well as in a i486/Pentium-driven PCI computer, with the appropriate BIOS and software. Beside Intel and DEC Alpha platforms, PCI is used on some PowerPC's.

Some PCI variations to be aware of: some implementations support "Bus Master" cards in all PCI slots, some in only one slot, and some not at all; some implementations support "bridging" on cards and some do not.

2.2 Performance

taken from Craig Sutphin's early Pro-PCI-Propaganda

Unlike some local buses, which are aimed at speeding up graphics alone, the PCI Local Bus is a total system solution, providing increased performance for networks, disk drives, full-motion video, graphics and the full range of high-speed peripherals. At 33 MHz, the synchronous PCI Local Bus transfers 32 bits of data at up to 132 Mbytes/sec. A transparent 64-bit extension of the 32-bit data and address buses can double the bus bandwidth (264 Mbytes/sec) and offer forward and backwards compatibility for 32 and 64-bit PCI Local Bus peripherals. Because it is processor-independent, the PCI Local Bus is optimized for I/O functions, enabling the local bus to operate concurrent with the processor/memory subsystem. For users of high-end desktop PC's, PCI makes high reliability, high performance and ease of use more affordable than ever before; no trivial task at 33 MHz bus-clock rates. Variable length linear or toggle mode bursting for both reads and writes improves write dependent graphics performance. By comprehending the loading and frequency requirements of the local bus at the component level, buffers and glue logic are eliminated.

See the chapter about Benchmarks for some crude (and perhaps meaningless) benchmarks on ASUS PCI Boards with 486 and 586.

2.3 The onboard-SCSI-II-chip NCR53c810

One very nice feature of some PCI mother boards is the NCR onboard-SCSI-II-chip, which is said to be as fast as the EISA-Adaptec-1742, but much cheaper. Drivers for DOS/OS2 are available. Drew Eckard has released his version of his NCR53c810-driver, which is in the standard kernel since v1.2.

This works so well I sold my adaptec-1542B-ISA soon after I bought the ASUS SP3-saturn-chipset II PCI board, and found the onboard NCR-SCSI controller to be much faster.

The NCR53c810-chip is onboard on some PCI-motherboards. There are add-on-boards available too, for about US$ 70.00.

There is only one thing I noticed did not work with the NCR-drivers when I tried them. Disconnect/Reconnect did not work, so using a SCSI-tape could be a pain, especially when using "mt erase" or the like blocks the whole SCSI-bus until it has finished. Since this was very unsatisfying for me, I bought one of these nice but expensive DPT PCI SCSI controller and had no such problems anymore.

People have reported this problem has been solved by Drew by now.

FreeBSD does support the NCR53c810 for quite a long time already, including Tagged Command Queues, FAST, WIDE and Disconnect for NCR 53c810, 815, 825. Drew said, it would be possible to adapt the FreeBSD driver to Linux. I somewhere saw some patches to do exactly this, any pointer to the location?

I personaly have the impression there are some important wheels invented more than once because of the differently evolving of FreeBSD and Linux. Some more cooperation could do both systems very well...

2.4 Drew Eckhardt on PCI-SCSI:

Drew said on end of March 95 about the SCSI on PCI: (slightly edited for clarity in context)

The Adaptec 2940, Buslogic BT946, BT946W, DPT PCI boards, Future Domain 3260, NCR53c810, NCR53c815, NCR53c820, and NCR53c825 all work for some definition of the word works.

2.5 New Alpha Version of the NCR driver

Well, this is not exactly *that* new anymore, please try to he versions which are in the kernel by version 2.0.x before going for this entry.

Alpha versions of the NCR driver which do neat things like disconnect/reconnect and synchronous transfers are now publically available. Any one interested in playing with them should

2.6 The EATA-DMA driver and the PCI SCSI controllers from DPT

The EATA-DMA scsi driver has undergone extensive changes and now also supports PCI SCSI controllers, multiple controllers and all SCSI channels on the multichannel SmartCache/Raid boards in all combinations of WIDE, FAST-20 (ULTRA) and DIFFERENTIAL.

The driver supports all EATA-DMA Protocol (CAM document CAM/89-004 rev. 2.0c) compliant SCSI controllers and has been tested with many of those controllers in mixed combinations.

Those are:             (ISA)   (EISA) (PCI)
      DPT Smartcache: PM2011  PM2012B 
      Smartcache III: PM2021  PM2022  PM2024 
                              PM2122  PM2124
                              PM2322 
      Smartcache IV:  PM2041  PM2042  PM2044
                              PM2142  PM2144
                              PM2322
      SmartRAID     : PM3021  PM3122  
                              PM3222  PM3224
                                      PM3334
      and some controllers from NEC, AT&T, SNI, AST, Olivetti and Alphatronix.

On a "base" DPT card (no caching or RAID module), a MC680x0 controls the bus-mastering DMA chip(s) and the SCSI controller chip. The DPT SCSI card almost works like a SCSI coprocessor.

The DPT card also will emulate an IDE controller/drive (ST506 interface), which enables you to use it with all operating systems even if they don't have an EATA driver.

On a card with the caching module, the 680x0 maintains and manages the on-board cacheing. The DPT card supports up to 64 MB RAM for disk-cacheing.

On a card with the RAID module, the 680x0 also performs the management of the RAID, doing the mirroring on RAID-1, doing the striping and ECC info generation on RAID-5, etc.

The entry level boards utilize a Motorola 68000, the high-end, more raid specific DPT cards use a 68020, 68030 or 68040/40MHz processor.

Official list prices range from $ 265 to $1.645 (January 18, 1996)

Since I've been asked numerous times where you can buy those boards in Europe, I asked DPT to send me a list of their official European distributors. Here is a small excerpt:

 
Austria: Macrotron GmbH            Tel:+43 1 408 15430   Fax:+43 1 408 1545 
Denmark: Tallgrass Technologies A/S Tel:+45 86 14 7000   Fax:+45 86 14 7333
Finland: Computer 2000 Finnland OY Tel:+35 80 887 331    Fax:+35 80 887 333 43
France : Chip Technologies         Tel:+33 1 49 60 1011  Fax:+33 1 49 599350
Germany: Akro Datensysteme GmbH    Tel:+49 (0)89 3178701 Fax:+49 (0)89 31787299
Russia : Soft-tronik               Tel:+7 812 315 92 76  Fax:+7 812 311 01 08
U.K.   : Ambar Systems Ltd.        Tel:+44 1296 311 300  Fax:+44 296 479 461 

"IMHO, the DPT cards are the best-designed SCSI cards available for a PC. And I've written code for just about every type of SCSI card for the PC. (Although, in retrospect, I don't know why!) ;-)" Jon R. Taylor (jtaylor@magicnet.net) President, Visionix, Inc.

The latest version of the EATA-DMA driver and a Slackware bootdisk is available on: ftp.i-Connect.Net:/pub/Local/EATA

Since patchlevel 1.1.81 the driver is included in the standard kernel distribution.

The author can be reached under these addresses: neuffer@mail.uni-mainz.de or mike@i-Connect.Net

2.7 BT-946C fully supported with kernel 1.3.x and newer

There is a driver in the 1.3.x kernels (available as a patch for the 1.2.13 kernel) written by someone associated with buslogic that fully supports the 946C and ALL of it's features including strict round robin, tagged queueing, multiple scatter/gather, multiple mailboxes, IRQ sharing, and yes, 15 devices on Fast/Wide. It is no longer necessary to use any ISA emulation with the driver (no DMA channel, no ISA address), and the driver is /fast/ and /stable/ (it's out of BETA and into full release).

The driver is available on ftp.dandelion.com (the newest version can always be got by doing "get BusLogic*"). It supports ALL BusLogic controllers with the exception of the FlashPoint LT, which uses a different interface. The driver is included in the 1.3.x kernels as standard for BusLogic devices.

2.8 Future Domain TMC-3260 PCI SCSI

Rik Faith (faith@cs.unc.edu) informed me on Wed, 1 Feb 1995 about the Future Domain TMC-3260 PCI SCSI card being supported by the Future Domain 16x0 SCSI driver. Newer information might be contained in the SCSI-HOWTO.

2.9 other thoughts on scsi

James Soutter (J.K.Soutter1@lut.ac.uk) asked me to add the following information on Fast-Wide-SCSI-2:

Fast Wide SCSI-2 is sometimes incorrectly called SCSI-3. It differs from the normal Fast SCSI-2 (like the Adapted 1542B?) because it uses a 16 bit data bus rather than the more usual 8 bit bus. This improves the maximum transfer rate from 10 MB/s to 20 MB/s but requires the use of special Fast Wide SCSI-2 drives.

The added performance of Fast Wide SCSI-2 will not necessarily improve the speed of your system. Most hard disk drives have a maximum internal transfer rate of less than 10 MB/s and so one drive alone can not flood a FAST SCSI-2 bus.

In Seagate's Oct 1993 product overview, only one Fast Wide SCSI-2 drive has an internal transfer rate of more than 10 MB/s (the ST12450W). Most of the drives have a maximum internal transfer rate of 6 MB/s or less, although the ST12450W is not the only exception to the rule. In conclusion, Fast Wide SCSI is designed for the file server market and will not necessarily benefit a single user workstation style system.

Rather than buying a PCI system with a SCSI interface on the motherboard, or rather than waiting for the NCR driver, you could purchase a separate PCI based SCSI card. According to Drew, the only PCI SCSI option that stands a chance of working is the Buslogic 946. It purports to be Adaptec 1540 compatible, like the EISA/VESA/ISA boards in the series.

Drew commented that other PCI based SCSI controllers are unlikely to be supported under Linux or the BSD's because the NCR based controllers are cheaper and more prevalent.

I definitly recommend reading the SCSI HOWTO in regards to newer information about PCI SCSI drivers.

Ernst Kloecker (ernst@cs.tu-berlin.de) wrote: (edited)

Talus Corporation has finished a NS/FIP driver for PCI boards with NCR SCSI. It will be shipping very soon, might even be free because a third party might pay for the work and donate the driver to NeXT.

Not every PCI-Board has got the chip. The old ASUS do, and one of the J-Bond boards does, too. (Most of the boards nowadays (6/95) do expect you to buy the NCR53c810 seperately.) Some vendors provide an alternative as you can read in Drew's text...

The NCR-Chip is clever enough to work with drives formatted by other controllers, and should be no problem.


Next Previous Contents