AlphaServer GS60 and GS140

The DEC TurboLaser 8200/8400 was a series of high end Windows NT compatible servers/workstations introduced in 1995.  After DEC was sold to Compaq (in 1998) the 8200/8400 were upgraded from the EV5/EV56 (21164/21164A) to the 21264/21264A (EV6/EV67).  Compaq wasn’t as bold with code names it seems so instead of being referred to as the TurboLaser they were simply called the TL6.  The machines themselves were also renamed from the 8200 to the GS60 and the higher end 8400 to the GS140.  GS referring to ‘Global Solution’ to reflect Compaq’s international marketing of the computers.  The GS60 was the lower end rackmount model supporting up to 6 CPUs and 12GB of RAM and the GS140 full cabinet model supporting up to 14 CPU and 28GB of RAM.  Both could be configured with either 21264 525MHz CPUs with 4MB of B-cache each or 700MHz 21264A CPUs with 8MB of B-cache each.  The 21264A added support for writeback cache, as well as its faster speeds and some new instruction set extensions.  Initially availability of these systems was in late November of 1999, coinciding with the release of the 21264A CPUs.  By the time of their release Alpha support for Windows NT was lagging, so most if not all systems were sold with Tru64 UNIX or OpenVMS OS.

The GS60/140 were large cases similar to a rackmount system but self contained.  The processor modules for them contained a pair of CPUs, the cache for the CPUs and the entire chipset.  They connected to the main computer with a very large connector that provided power (48VDC) as well as all the Memory/IO and clock signaling.  This was referred to as the TLSB (TurboLaser System Bus).   The fastest of these was the KN7CH (also known as the E2067-DA) which had dual 700MHz 21264A processors with 8MB of Cache each.

DEC KN7CH 6/700 Processor Board

This processor board is quite interesting, its a rather early board (PLDs are dated March of 2000) and the pair of Samsung 21264A processors are dated 9944, these are some of the very first production 21264As.  Also of interest is that these Samsung CPUs are 733MHz models (KP21264A-733UCN).  The 21264A was to be made in 600, 650, 667, 700, 733, and 750MHz versions, though I have only actually seen 667 and 733MHz versions.  Making only 2 speed grades of the processor would greatly simplify testing and logistics, and with a rather limited customer base, there wasn’t a clear marketing need to make so many different speeds, these were not CPUs that were generally available outside of OEM use.  These servers were also designed to be high reliability systems, running a 733MHz rated CPU at 700MHz would increase reliability by decreasing heat related wear and tear.

Build Sheet for a 8-Node GS140 with Eight 6 CPU GS140 6/700 Systems. Each with 12GB of RAM. A nice $9 million system

The entry price for the AlphaServer GS60 with 4 GB of memory was $199,990 ($340,000 in 2022). The AlphaServer GS140 system price started at $399,400 ($680,000 in 2022). These were very expensive systems.  One look at the processor board shows what that kind of expense gets you, a whole lot of gold.  Its hard to find another computer system built in 2000 that has 9 gold/ceramic chips on each processor board.  A single dual processor board was $45,000 ($76,000 in 2022USD), and each 4GB of RAM was another $49,000.  One can easily see how such a system could quickly cost several million dollars.  Each of these boards cost as much as a really nice car!  Lets look at what that $45,000 gets you

Top Row (L->R) SWI, Alpha 21264A, SWI – Bottom Row: TDI, TDI, TCC, TDI, TDI

2x KP21264A-733UCN. Each 21264A chip has a separate address and data bus for the B-cache and system operations. The 21264A chip has a 64-Kbyte instruction cache and a 64-Kbyte data cache.  These are made by Samsung on a 0.25u process and dissipate 85Watts at 2.0V.

20x IBM SRAM Cache Memory: 8-Mbyte ECC L2 cache per CPU made using 16x IBM 0418A81QLAA-4 512Kx18 8Mb ECC SRAM chips and 2x 128Kx36 / 2x 256×18 for the TAG RAM

2x DEC 21-47306-01 SWI: Two swizzle (SWI) chips receive data from the 256-bit wide DLSB (the DEC Local Bus) and pass it to one of the CPU chips over the 64-bit wide data interface bus.  These are located on either side of the pair of CPUs.

4x DEC 21-47307-01 TDI: Four TurboLaser Data Interface (TDI) chips receive data from the TLSB (the main system bus that connects all cards in the system) and pass the data over the DLSB to the two SWI chips.  These are the outer 4 chips on each end of the row of 5 gold chips on the bottom.  Each one handles 64-bits of the 256-bit TLSB.

1x DEC 21-47315-01 TCC: The TurboLaser control chip (TCC) takes commands from both CPUs and issues them to the TLSB. It also controls all data movements through the TDI and SWI chips. This is the center chip between the pairs of TDI chips.

2x AMD AM29F080DB-90EC: 5V 8Mbit Flash for the system firmware

4x Galaxy Power DC-DC Converters.  These regulate down the 48VDC supplied by the systems redundant power supplies to the voltages needed for the board.  There is a pair of 2.2V 7A converters for the CPUs, and a 7A 3.3V converter for all the I/O.  There also is a smaller 2A converter of unknown voltage (likely 5V).

Pair of Samsung KP21264A-733 Processors surrounded by cache chips

The TurboLaser line was replaced in 2002 by the WildFire servers (GS80, GS160 and GS320) which upgraded the CPU support to 32 21264Cs with 256GB of RAM.  Unfortunately by this time Compaq had merged with HP and the combined server line was a bit cluttered, having Alpha, PA-RISC, Itanium and Xeon based systems.  The Wildfire and its Marvel follow on were the end of the road for the Alpha.  Unfortunately the same thing happened with the PA-RISC and Itanium (ok maybe not so unfortunately with Itanic) as well.  The days of boards full of golden RISC are past, replaced by BGAs with enormous heatsinks.

GammaChrome XM18 – Engineering Sample

Recently I got in some very nice S3 GammaChrome GPUs.  The GammaChrome was S3 (owned by VIAs) follow on to the DeltaChrome and included support for such things at PCI-E.  The S18 (Code name Brooklyn) supported speeds of up to 500MHz and was made on a 130nm process by TSMC.  S3 also made a mobile version of the S18 called the XM18 (Code name Metro MPM) in 64MB and 32MB versions.  Clock speed on these was around 350MHz (memory on the samples I have is 350 so core should be similar).  The XM18 was packaged on a MPM (Multi Package Module) with 2 RAM chips and the GPU mounted on a small chip size BGA with around 800 balls.  This is very similar to how ATI packaged some of their mobile GPUs (like the Mobility Radeon 7500 and 9600).

HP 93000 (from HP Brochure)

So how do you test one of the XM18 Engineering Samples? Or any large scale chip for

86C813 ES Gamma Chrome XM18 ULP MPM64

that matter?  With Automated Test Equipment.  ATE systems are designed to rapidly test various chips to verify their design/performance before they go into full production (or to test samples of production ones).  The HP/Agilent 93000 (spunoff as Verigy in 2007 and acquired by Advantest Corporation in 2011) was introduced in 1999 to handle such testing, and at the time was rather revolutionary.  Previously most test systems used a simple test head that would mount the chip to be tested, with all the processing and customizations being contained in the main test machine.  This worked fine for a single design, but to test multiple chips got pretty expensive.  HP moved the testing to the test head directly, interfacing to the target chip via a large PCB.  This way changing chips only required updating the test program, and changing out the PCB.  Design changes required reworking a single PCB, rather then the entire test machine.

HP 93000 Test Head – Notice the 16 groups of pins (some covers and some mangled in this old sale photo)

The 93000 was the first ATE that achieved (on its low end (200Mbps) a cost of $1000/pin tested, and on the high end, test speeds of up to 1250Mbps (for the P1000 version, at a cost of $6-7000 per pin).  The XM18 has around 800 pins, half are probably power/ground so 400 some odd testable pins, in a mid range HP 93000 and you see these systems were not inexpensive. Well over a million dollars for a midrange system.

GammaChrome XM18 – Metro MPM Test Board

To use such a system the chip to be tested would be mounted on the test board, usually with a BGA socket.  This board breaks out all the various connections of the chip to 16 sets of contacts, which the probe head of the HP 93000 made contact with using spring loaded contacts.  The board is then clamped down and tests are run.

Connection List

These boards are very very large, each one is 17x23inches (43x58cm) and 5mm thick.  They weigh about 7lbs (3.1kg) as well.  They got used a lot and need to be rather robust and durable.  You can see the boards are marked with tables of all the connections, and where they are brought out to.  Useful information about what supporting equipment is need (sockets and stiffeners etc) is marked on the board as well.

Back of board. Notice all the capacitors, a crystal, and a series of 5VDC reed relays (the red devices)

These boards appear to be a ‘static’ type item, but they do require adjustment, notice the markings that say not to use this board, it needs recalibrated.  Looking closely at the board you can see capacitors have been removed/replaced, and many of the capacitors have felt tip marker markings on them.  Keeping the capacitance and inductances at their proper values 9and matched, considering the long trace lengths) would be a very important thing.

S3/VIA Matrix Test Board. The Matrix was the code name for the GammaChrome S14/S19

These test boards are from 2006, the 93000 systems are still being used today in upgraded form (now called the V93000) to test SoCs and other chips.  As chips have gotten more and more complex, faster, and with larger pin outs, test equipment continues to grow ins peed, and cost as well, but is an essential part to the process of designing, producing and supporting a successful GPU or CPU.

C5M – Ezra-T Prototype Pathfinder – PGA370

Most are familiar with the history of VIA so we won’t dive extensively into that but a quick summary is in order.  VIA was founded in California in 1987 before moving to Taiwan, and previous to 1999 was well known for making chipsets and other support chips for computers.  In 1999 VIA bought both Cyrix (from National Semiconductor) and Centaur Technologies (from IDT, who made the Winchip series of processors.

These purchases did two main things for VIA, it first gave them access to the x86 architecture, and it gave them legal leverage to continue down the x86 road.  Cyrix possessed a license to the P6 processor bus (through a cross licensing with Intel) that was good until 2006.  This allowed VIA to make what became the Centaur based CyrixIII/C3 processor on the P6 based Socket 370 platform. These are the processors and socket we are most familiar with for VIA CPUs.  With clock speeds of 466-1.2GHz and eventual support for the Tualatin based boards these chips were the most ‘public’ facing CPUs.  VIA also of course made many BGA versions, used in ITX form factor, and other mini type systems.

CNA – Isaiah – Interestingly using the old Pentium III-M pin out

The VIA designs, despite originally being called ‘CyrixIII’ were all based on the Centaur designed core.  Intel, as was its custom, sued VIA in 2001 asserting patent infringement, which it is likely VIA was expecting.  As with the case of Intel and Cyrix, VIA countersued, asserting Intel was infringing on patents VIA had acquired with the Centaur deal.  In 2003 a settlement was reached that included a 10 year patent cross license between Intel and VIA and allowed VIA to continue to make x86 compatible processors (extended in 2013 by 5 years until 2018(.  The deal also granted VIA a 4 year (with an extra optional year) license to continue to make chipsets compatible with Intel processors (they had originally signed a deal in 1998 to allow VIA to do so. This is how we continued to get VIA chipset based motherboards for Intel processors.  The deal also added a small detail that leads to todays discussion, it granted VIA a 3 year grace period to continue making bus and pin compatible processors up through 2006.

C5J (Left) and C5R (Right) – Banias Compatible Pentium M pin out

This last part is interesting, the fact that it was a grace period means it reflected what VIA was currently doing, not what they were planning to do in the future.  The obvious example here is the C3 line on Socket 370 using the P6/Tualatin bus, but that was pretty old news in 2003 so what was VIA working on?  CPU’s on more modern sockets of course, namely Socket 479 (mPGA479M) used by the Pentium-III-M (Tualatin) and Pentium M (Banias/Dothan).  These use the same physical socket on a motherboard, but the keying pins are different on the CPUs themselves.  These are all mobile designs which lend themselves well to VIAs low power designs.  VIA did also make several reference boards for these CPU’s so its clear that there was plans for releasing them to the broader market, and likely with additional motherboard support.

C5J (Left) and C5R (Right) – C5R is a 110nm part with a slightly larger die the the C5J

Another socket was just being developed at the time of this agreement, and that is perhaps the most interesting.  Intel LGA775 chips began sampling in late 2003, which is after the grace period of 3 years had begun so it would make sense for VIA to not develop CPU’s using a socket they were going to lose access to in a few years.  The package likely was in development for a couple years prior which is likely why VIA made a few (likely VERY few) samples for it.  The samples are marked C5R which is a C7 Esther core, if VIA’s naming is consistent, this would be the TSMC 110nm version of the 90nm C5J.

C5R With heatspreader and with heatspreader removed.

The Esther core code names are a bit confusing because of how some CPUID programs identify them. X-86-guide.net has a quite nice ID guide that goes into some great detail on them.  In summary there was a 90nm Rev A C5J made by IBM, and later a 90nm C5J (called Rev D) made by Fujitsu with some additional features.  This Rev D part often gets identified as a C5R, or a C5J shrink, neither of which is correct.  The actual C5R (and related C5Q) were what appear to be backup plans for the IBM produced parts, using a larger 130/110nm process at TSMC. Looking at the mPGA479 unfinished packages (labeled C5J and C5R) the die attach area on the C5R is actually slightly LARGER then the C5J (~35mm2 compared to 28mm2 of the C5J)

C5R Esther – 110nm TSMC in LGA775

Most VIA samples are labeled with the code name in Cxx format and not the marketing code name (Esther Isaiah etc) as each of the Marketing code names (for lack of a better term) consisted of many actual sub-cores.

Looking at the table above we can see VIA took many roads in the development of their CPUs, with many that went nowhere.  Some may see this as a lack of direction or focus, but in a lot of ways VIA seemed to be trying to figure out the best CPU for the market at the same time they were trying to make the best CPU from an engineering standpoint.  Where these two paths converged you had a marketable CPU that made it into mass production, and where they didn’t, or where legal road blocks arose, the design was canceled.  VIA’s CPU development is even more obscure now, though they have made a few other designs we will cover in a later article, as well as the return of Intel to the VIA party.

This is turning into a bit of a series on Soviet processors.  Continuing from our article earlier on the 1801VM2 LSI-11.  The 1801VM3 is the further development of 1801VM1/VM2 and is the highest performance microprocessor in 1801 series. It’s a 16-bit single-chip microprocessor that includes an operating unit, a firmware control unit, an interrupt unit, a memory controller and Q-BUS control unit. A distinctive feature of 1801VM3 is a large amount of addressable memory (4MB vs 64K for the 1801VM1 and 64k+64K for the VM2), high performance and ability to connect a floating-point coprocessor 1801VM4.

1801VM2 die

1801VM3 Die

1801VM3 Specifications

• Number of processor Instruction: 72 Fixed Point and 46 Floating Point (with 1801VM4 FPU)
• Address Space: 4MB
• General Purpose Registers: 8
• Manufacturing process: 4 micron N-channel silicon gate MOS technology (later migrated to 3 micron)
• Die size 6.65 × 8 mm
• Transistor count: 28,900 active transistors, 200,000 integral elements
• Clock rate: 4MHz  (1801VM3V) 5MHz (1801VM3B) 6MHz (1801VM3A, upgraded to 8 in 1991)
• Performance: For register based operations (like addition) up to 1,500,000 instruction/s (1.5 MIPS)
• IRQ Lines: 4
• Supply voltage + 4.75V-5.52V
• Power consumption: 1.7-2 W
• Packages: CDIP64 (KM1801VM3) LQFP64 (KA1801VM3) CQFP64 (KN1801VM3/N1801VM3)

Like the VM2 before it the speeds were denoted by a series of dots on the package (or lack thereof)

KM1801VM3A – 6MHz (no extra dot) CDIP64 package from 9008

KM1801VM3B – 5MHz (one extra dot) CDIP64 package from 9003

KM1801VM3V – 4MHz (two extra dots) CDIP64 package from 9202

KA1801VM3 – 8MHz (no extra dot – post 1991) PQFP64 package from 9108

N1801VM3 – 8MHz (no extra dot – post 1991) CQFP64 package from 9324 – Remarked from a military part (rhombus marking marked over)

The KM1801VM3 appeared as part of the DVK line of computers, starting with the DVK-3M model (PCB ”Electronics МС 1201.03” and “Electronics МС 1201.04”).  Using the same ISA (Instruction Set Architecture) allowed DVK (and others) to rapidly update their computer line when new processors were available, and allow for a wider software base.  This is very much like the original IBM PC using the x86 architecture.  The transition from 8086 to 80286 was relatively easy to design, and nearly seamless for the end user.

DVK PCB Electronics МС 1201.03 board on the top.

Many devices built on the basis of the 1801 series CPU contain other microcircuits of the same series (support circuits).
In addition to microprocessors, this series includes:
– ULA 1801VP1-xxx
– masked ROM 1801REх-xxx
– EEPROM 1801RR1

ULA and EEPROM

The 1801VP1-xxx is a ULA- (Uncommitted Logic Arrays). It’s made using a 3 micron N-channel silicon gate MOS technology with one metal layer. First, base silicon wafers are made that contain transistors. These are doped regions of silicon and a separate oxide-insulated layer of polysilicon gates. Then all this is covered with an oxide layer. Base wafers are ready.

In this form, the wafers can be stored for a long time or transferred to another fab. All 1801VP1-xxx chips, regardless of number, have the same structure and arrangement of transistors. And they are made on the same base wafers.

KR1801VP1-22 die

Differences between the chips appear only at the last stage of manufacturing. In the upper oxide layer, the die is etched by photolithography to access the required transistors. And then form a metallic pattern from aluminum. This pattern defines the electrical circuit. The number in the marking identifies the purpose of the chip. For example, 1801VP1-033 is an external device controller.  This is similar to how a MaskROM is made but instead of only memory elements, it contains logic elements allowing for a custom IC to be made (like a mask programmable PAL/GAL)

KR1801VP1-119

The 1801VP1-119 is a companion chip for 1801VM3. It can be said to be the “north bridge“.
The 1801VP1-119 performs the following functions:
-forms control signals for DRAM;
-forms control signals for system SRAM;
-generates signals to select system ROM;
-generates control signals for detection and correction of memory errors (EDC) using Hamming code (555VGH1). Error correction circuits reduced performance by 10-15%. Therefore in some computers, there were jumpers to enable/disable the EDC
-buffer data register control;
-generate other signals

This was the beginning of what would be come chipsets, replacing loads of TTL with custom circuits.  The exact same evolution was occurring in the west with the PC environment, until nearly all the support circuits were integrated into just a couple large ASICs.   Its interesting to see the development paths of the Soviet computers and the West.  While they were entirely different instruction sets, they evolved in very much the same way.  East or West, LSI-11 or x86, at the end of the day, a computer is a computer and will evolve in similar fashion.

The Soviet-made 1801VM2 CPU (a binary-compatible implementation of the PDP11 instruction set and QBUS interface) was developed in 1982. The 1801VM2 is a further development of the earlier 1801VM1 doubling the original 5MHz clock speed. From a constructive standpoint this CPU is a completely independent development.

1801VM2 die – 1983 dated

1801VM2 Specifications

• Number of processor Instruction: 72
• Manufacturing process: 4 micron N-channel silicon gate MOS technology
• Die size 5.3 × 5.35 mm
• Transistor count: 18,500 active transistors, 120,000 integral elements
• Clock rate: Up to 10 MHz
• Performance: For register based operations (like addition) up to 1,000,000 instruction/s (1 MIPS) – for operations like multiplication, up to 100,000 instructions/s
• Supply voltage + 5V
• Power consumption: up to 1.7 W
• The case is 40-lead, ceramic DIP (KM1801VM2) or plastic DIP (KR1801VM2). (a surface mount version was also made)

To increase noise immunity in comparison with 1801VM1, additional ground contacts were made for the address / data bus.
The 1801VM2 was manufactured at two factories: Angstrem and Solnechnogorsk Electromechanical Plant (SEMZ).  As was typical of the time speed grading was done by adding extra marking to the chips post-testing.  Its very easy to miss these, if a chip was tested at 10MHz and passed it received no extra marking and was considered an 1801VM’A.’  If the device failed at 10MHz but ran at 8MHz a small dot was added to the package (and was considered a grade ‘B’ device).  This dot was not to be confused with the dot for the pin one marker, though often placed…next to it.

Ceramic DIP 1801VM2A Angstrem – 1989 No extra dot

Ceramic DIP 1801VM2B Angstrem – 1987 – Note the extra dot in this case by the date code

Plastic DIP 1801VM2A Angstrem – 1990

KN1801VM2- Angstrem 1985 CQFP Surface mount version (image Baator)

Ceramic DIP 1801VM2 Solnechnogorsk Electromechanical Plant – 1990 – Extra dot by pin 1 marker

In comparison with 1801VM1, expanded arithmetic instructions (MUL, DIV, ASH, ASHC – part of a the set of PDP-11 EIS), and also operations from the floating point instruction set (FIS) were added. The FIS instructions (FADD, FSUB, FMUL, FDIV) are realized through subroutines – when performing these instructions there is a special type of interrupt and the program handler in memory (“shadow” system ROM K1801RE2) of the console mode is executed, a ‘firmware’ style of FIS implementation, as its not truly hardware (the ROMs break down the FIS instructions into something the 1801VM2 can execute)
During the design of the microprocessor, a microcode error was made, leading to a malfunction of the processor when reading with addressing method 17 ( MOV (PC), R0).

DVK-1 Computer

The 1801VM2 was the heart of a number models of DVK computer. DVK was developed at the Research Institute of Precision Technology , Zelenograd (just outside of Moscow). The first model DVK-1 was developed in 1981, and released in 1983. Architecturally DVK copies mini-computers from DEC PDC-11 and PDP-11. By 1990, 200,000 DVK computers of the nine different models were produced.

Romashka Word Processor

Use of the processor continued well into the 1990’s. The “Romashka” belonged to the latest generation of electronic typewriters, which in their functionality were close to computer text editors. This typewriter made it possible to automatically format text (set alignment, change the spacing between characters and between lines, use bold and underlined fonts, etc.) and had an electronic memory of at least one page (3800 bytes).  In the West these half typewriter half computer were called Word Processors, and were quite popular through the 1980’s.   The machine’s control unit was a microcomputer based on the KM1801VM2 processor.
“Romashka” was produced by the Kursk PO “Schetmash” in the first half of the 1990s.

“Electronics IM-05 “- Soviet chess computer, contains 1801VM2 inside. It was a continuation of the line of chess computers “Electronics”. Produced by the Svetlana Association, Leningrad.

In 1984, the military-grade microprocessor 1806VM2 was released.
This microprocessor functionally corresponds to the 1801VM2, but is made using CMOS technology.

• Clock rate: up to 5 MHz
• Number of Instructions: 77
• Contains 134,636 integral elements
• Power consumption: up to 0.025W

The 1806VM2 developers fixed the microcode bug present in 1801VM2 (much to the relief, or annoyance of programmers). The 1806VM2 was supplied in a 42-lead dual in-line ceramic package with flat leads, N1806VM2 in a 64-lead CQFP. The rhombus marking on the chips denotes a military-grade device.

1806VM2 – Angstrem 1991 in the nice pink flat pack

N1806VM2 – Angstrem 1999 in a Ceramic quad flat pack

CQFP N1806BM2 on a ceramic substrate forming a military Single Board Computer – circa 1987 (image Baator)

These 1806VM2 are still being made by Angstrem, if you need to build a PDP-11 computer to run Tetris on, or repair a Buran shuttle you may have laying around.

In 1990, a radiation-hardened microprocessor was introduced, compatible with the 1806VM2, known as the 1836VM2/N1836VM2.  Just like in other countries, existing code base and known reliability are more of a driver of what the military/industry uses than having the latest and greatest.  There are still MIL-STD-1750A processors being made and used, rad-hard 8051s and 80186s, and Soviet PDP-11 processors right there with them.

Photos of microprocessors from the collection of Perfiliev Andrey (Andreycpu).
Article written originally by Contributing Author Vladimir Yakovlev (edited by cpushack)

IBM 4020 Q-Pacs – 1960’s

Back in the late 1950’s two things were happening (ok more then 2 but 2 relevant to todays discussion) the military was looking to replace the new but now already out of date tube based SAGE and AN/FSQ-7 Strategic Air Command (SAC) computers, and multiple bits of data were beginning to be called bytes.  The SAC was in charge of all of the US’s Strategic bombers, ICBMs, and detecting/tracking the threats of bombers/ICBMs from the USSR.  The older tube based SAGE computer was designed for relaying, consolidating, and displaying data from Early Warning RADARs across North America to paint a situation picture of what was going on.  It worked fine, for bombers, but the late 1950’s also brought about ICBMs, and ICBMs are much much faster then mere bombers.  The SAGE, and the AN/FSQ-7 lacked the processing speed to keep up with the changing data from a RADAR track of an ICBM so something faster was needed.

Each module weighs around 90 grams

IBM developed and proposed the AN/FSQ-31 (and the FSQ-7A which got renamed the FSQ-32) which were based on the newly developed IBM 4020 military computer.  The IBM 4020 was completely transistor based and designed for reliability and speed.  Marketing materials of the time refer to its ‘resistance to the effects of nuclear blast,’ clearly this was the 1950’s.  At the heart of the 4020 designs was the Q-Pac. These were pluggable, ceramic encapsulated circuit packages. The majority of all logic requirements can be met
by seven basic types of Q-Pacs, each containing from one to four circuits. The use of transistors, diodes, and resistors/caps on each Q-Pac served as what TTL/RTL of the 1960’s/1970’s formed, discrete logic elements, albeit simple ones. In the 4020 the computer was divided into modules (racks) which each contained 16 drawers. Each drawer could hold 96 individual Q-Pac (or 48 double Q-pacs).  That’s 1536 logic elements per module, and the 4020 had 8 modules, resulting in around 12,288 Q-Pacs.  It appears each Q-pac could support 6 discrete transistors, so the 4020’s basic data path (not counting memory, I/O or storage subsystems would max out at 73k transistors.  Obviously there would not be a system that was ALL transistors but this gives us an idea of the scale of the computer. This is around what the Motorola 68000 CPU had or a Intel 80186.  The typical 4020 (again not counting the peripherals) was water cooled, used 13kw of power and took a good 85 sq ft of floor space.

Five simple transistors in the one on the left, and a pair of diodes on the right.

The 4020 was a 48-bit word length (pus 2 parity bits) computer and was capable of around 400,000 Instructions per second with a 2.5microsecond cycle time (6.25MHz).  It supported 128kwords of drum storage (remember 48 bit words, so this is about 6Mbit.  The 4020 also supported byte processing, using the 48-bit word as 8 6-bit sections which IBM called bytes.  This is one of the first official commercial usages of the term ‘byte’ for a chunk of data.  We think of bytes as 8-bits but thats only a standard thats been around the last 30 years or so.  Back in the 1950’s it was the wild west of data naming.  It was common to use 6-bits for BCD (Binary coded Decimal) and 6-bits to represent characters, so a 6-bit byte was only natural for IBM to use.  This eventually gave way to the 8-bit bytes we all know and love by the late 1960’s, though some processors even in the 1970’s used 12-bit words (Intersil 6100 and some PICs) and other oddities (14 bits from the PIC16).

AN/FSQ-31

The process of integrating the 4020’s into SAC facilities took longer then expected, not being completed until 1968, by which time they were of course outdated again.  By 1975 most of them had been replaced by newer Honeywell systems.  Interestingly, the 4020’s tube driven predecessor lasted in some bases until the early 1980’s.

It wouldn’t surprise me if, even after 60+ years, these Q-Pac modules still worked, after all, that was their intended design, to be rugged and reliable.

The Q-Pacs are in a lot of ways an early predecessor the IC’s of today, a single module containing various logic elements, while not on a silicon die, they were ‘built’ by hand, on a ceramic substrate.

TCS008 Adder – TCS017 FPU Control and TCS060 Shift Register – 1974-1975

Today most chips we use are made in CMOS (Complimentary Symmetrical Metal Oxide Semiconductor), which is a process using both p-type and n-type MOSFET transistors.  It was invented back in 1963 by Fairchild, but was commercialized by RCA in 1968 with the introduction of the CMOS based 4000 series of MSI logic devices.  These were basic IC’s with such things as NOR gates, Adders, Flip flops and the like.  A CMOS equivalent to TI’s popular TTL based 7400 series.

RCA also made a series of computers in the 1960’s (to compete with IBM) as well as other electronic products. including many for the US Air Force, NASA and US Army.  In 1970 RCA created the SSTC (Solid State Technology Division) in Somerville, New Jersey to develop CMOS processes (and Silicon on Sapphire versions) into more commercial products. At the time most IC’s (outside the 4000 series) were made in PMOS or NMOS, CMOS was considered too slow, despite is lower static power usage and high noise immunity.  SSTC was to develop processes, standard, and eventually devices, that RCA could then commercialize and/or use in their other products (such as their computer line, radios, and military products).  It was out of this project that the famous COSMAC processors (CDP1801 and CDP1802 line) came from.

TCS002 16×16 Multiplier 200ns – Note the hand written characterization markings – 670uA @ 5V

SSTC also made a series of essentially standard test devices.  These were based on a common cell architecture (more common in ASICs today) with a series of chips made to demo what was possible with the CMOS-SOS (CMOS on Sapphire) process.  These ‘standard’ IC’s would then be used in various demo products for potential customers.  The most interested customers at the time were the US Air Force and NASA.  The RCA CMOS process allowed for a great power savings, and especially when built on a sapphire substrate, exhibited a high tolerance to radiation, useful for the then rapidly expanding satellite/space market.

AN/GVS-5 Laser Range Finder – 1970’s. They were huge, but very impressive for their day

The first of these chips were made in 1974-1975 and were made with a 7 mil (178micron) standard cell height, on a 20 micron process.  Versions were also made with a 5 mil (127 micron) size, specifically for the military market.   These were not typically commercially available devices, but used internally for test, evaluation, and to build specific products, though the technology used for them was often turned into generic products.

Below is a list of some of these devices SSTC made. The TCS prefix was used to denote these being made by SSC on a CMOS-SOS process.  A TCC prefix is a standard CMOS process.

**Used to build the NASA 32-bit SUMC (Space Ultrareliable Modular Computer)

These were used in many military products such as the AN/GVS-5 handheld laser rangefinder, a Programmable waveform generator used in FM RADARs, and for the imaging system (digitization and compression of video to be sent) in the remotely piloted Lockheed MQM-105 Aquila drone (yah drones, back in 1975).  The Aquila project was particularly challenging, as the circuitry had to be small enough, and low power enough to fit on a small airframe, yet still handle video compression fast enough that a ground station could receive and decode useful imagery.   This was done with several large hybrid circuit modules consisting of many TCS057 Multipliers and TCS065 Adders.  This was capable of 200-1600Kbps data rates, not bad in 1975.

Aquila Artillery Spotting Drone (Lockheed Martin)

Most of the TCS line of components was capable of 10MHz operation while running at 5V, and voltage and clock rate scaled with each other, so they could be clocked lower for less voltage and power usage, or clocked higher at the expense of more power.

It is a bit unfortunately that RCA lost its way in the 1970’s, attempting to became a conglomerate, they became known as Rugs, Chickens and Automobiles (having bought parts of Hertz Rental Cars, a frozen TV dinner company, a carpet company and others).  They were bought by GE in the 1980’s and in 1988 the Solid State Division, with what remained of the SSTC was purchased by Harris Corporation, which continued to make the 180x line of CMOS processors for over 20 years.  If RCA had stayed focused on making CMOS a commercial success, we may have had more and faster CMOS processors nearly a decade sooner.

NEC uPD70108C – V20 CPU – Late 1984

NEC had a cross license agreement with Intel dating back to April of 1976 that allowed each company to make/sell products based on each others patents.  This was particularly important in the 1970’s as having a viable ‘second source’ for your designs was considered critical for it to be viable in the market.  This was especially true for Intel, who wanted to get into the Japanese market. In 1979 NEC began to produce and sell the 8086 and 8088 processors.  NEC wasn’t going to succeed by just being a second source to Intel though, designing their own processors was of great importance.  While producing the 8086/8088 they also began working on their own version, which would be an enhanced 8086/8088 processor.

NEC V30 Die (courtesy Birdman) – 8086 with many enhancements

The result was the rather well known V20/V30 processors of 1984.  These were not just clones of the Intel MCS-86 (though determining this took several court cases and resulted in the Chip Act of 1984).  The V30 had some pretty big differences, notably, internally it had dual 16-bit busses, allowed data to be moved much more efficiently, as data could be moved into and out of a register at the same time (nearly).  It also increased the microinstruction word from 21 bits to 29 bits, added a hardware effective address generator, additional instruction pointers, and a hardware shift/loop counter.  Taking advantage of these features added some new instructions as well, 156 compared to the 8086’s base 133.  The V30/V20 were the beginning of a line of V-series processors.  NEC went on to make  ‘186/188 style processor (the V40/V50) as well as a series of microcontroller versions  (V25/V35 and others).  The V20/V30 were to be supported by a math coprocessor like the 8087 called the upd72091.  Very little info is available on the 72091 as it was cancelled very early on in its design, as by 1984-1985 it was already out of date.  Its replacement was to be a bit more powerful.

Design of the the upd72191 started likely at the same time the V30 was released, around 1984-85, with specifications released in 1986, and plans for chips by 1987.  This chip was in an advanced state of planning, such that many products, including motherboards (such as the Ampro Little Board PC) and industrial controllers designed with sockets for it.  Preliminary datasheets exist, but alas, no chips seem to be found.

LittleBoard PC (Ampro) with support for canceled upD72191 (V40 based)

The upd72191 was made in CMOS and is a bit like an enhanced 80C187 but with support for the V20/V30.  It is fully IEEE-754 compatible (the 8087 wasn’t as the standard wasn’t finished yet) and supports a similar instruction set as the 80C187 (and thus the 80387).  Unlike the 8087 it supports the full set of Exponential, Trig, Logarithmic, and Hyperbolic instructions.  The 8087 was somewhat limited in this, as it was already pushing the limits of what was possible on a single chip at thee time of its release.  The 72191 supports FSIN/FCOS which the 8087 doesn’t and many other functions (its full instruction set could not be found).  The 72191 has a mode pin that selects between interfacing between the V20/V30 and the V40/V50, (as these talked to coprocessors differently) so it was compatible with 4 distinct processors.  The 80C187 could only be used with the 80186 and the 8087 could only be used with the 8086/8088.

upD72191 FPU Block Diagram – 1986ish

Looking at the block diagram of the ‘191 we notice something else, its a dual bus design, much like the V30 processor.  Internally there are a pair of 74-bit busses for the mantissa (fraction) side and a pair of 16-bit busses for the exponent side.  This is a striking difference from that of the 8087 and the ‘187.  The 8087 has a single 16-bit bus for the exponent, and a 64-bit (68-bits into the shifter and ALU) for the mantissa.  There are 3 extra bits for enhanced accuracy, and a extra leading bit that is always 1 for floating point math, giving 64 bits of ‘data’.

The dual bus design makes sense as NEC did the same for the V-series.  Coupled with the right microcode, it can greatly enhance the speed of the FPU.   So why then is the bus expanded to 74-bits for the mantissa?   In the 80187 and 80387 this bus is still only 68-bits.  We look to the design of NECs follow on FPU for the answer.  The upd72291 (and its 32-bit bus 72691 version) are rather different beasts, made for the the V33/V53 x86 CPUs and V60/V70/V80 non x86-CPUs.  We’ll talk about them in more detail later, but they share the same 74-bit mantissa as the 72191, and in this case, the designers wrote a paper on its design.

The FPP [72691] is the only floating point processor that provides the power function x^y.  This function (called FPOWER in the instruction set) is difficult to implement not only for its complex definition but also for sufficient accuracy. The equation X^y = e^(y*log_eX)
does not give good accuracy because the accuracy error of the log function is augmented by the exponential function.  The FPP solves this problem by providing a 74-bit data width for the mantissa data bus.

Being as the 72191 was canceled, the ‘291/691 would in fact have been the only FPU to support this in hardware, but it seems it was first implemented on the ‘191.  The solution only works well for larger (greater then 32) values of y, otherwise iterative multiplication is used, but where it can be used it greatly speeds up the calculation.

When the 72191 was canceled NEC thoughtfully provided a single chip solution called the upd9335C for allowing an 8087 to be interfaced to the V40/V50 processors which, like a 186, used a HOLD/HOLDACK bus release protocol instead of the 8086/8088s (and V20/V30s) REQUEST/GRANT.  For applications using a V20/V30, an 8087 could be used directly.

NEC upD70632R-20 20MHz V70 Processor

In 1989 NEC released the next of the V-series, the V60, V70 and later the V80 processors.  These were a departure from the previous in that they were no longer based on the x86 architecture, but rather a completely new ISA (though the V60 and V70 had a V20/V30 emulation mode).  These were full 32-bit designs, and were Japan’s first widely available 32-bit processors.  Of course with a new processor comes the need for a new FPU and NEC had not one, but 2 FPU options for these.  The upd72291 and upd72691 are based on the same design, but with some major feature differences.  The 72291 is designed to work with processors that have a 16-bit data bus such as the V60.  It also could be used with the older V33/V53 x86 designs.  Internally it has eight floating point registers and supports all your typical floating point functions as well as vector math functions.  The upd72691 is designed for 32-bit data paths, but adds a bit more…

NEC updD72291R-16 FPU

In addition to expanding the register set to 32 FP registers, the ‘691 also added a complete suite of matrix  math functions. The ‘691 was made on a 1.2u CMOS process and contained 433,000 transistors. (nearly 50,000 MORE then the V60 processor) Running at 20MHz it was capable of around 6.7MFLOP and supported 24 vector/matric instructions as well as 22 mathematical functions.  Like the 72191 it had a 74-bit mantissa datapath, but expanded the exponent path to 17-bits to support double extended precision number formats. It is a highly microcoded design using a 3072 word (43 bit word) microcode ROM, 20% for vector/matrix, 37% for arithmetic, and the rest for exceptions handling and other house keeping instructions. Interestingly, these microps themselves encode additional instructions that NEC call nano-ops, these controlled just the ALU operations of the instruction (the rest being bus control and sequencing).  These nano-ops were stored in a 256 word x 74-bit Nano ROM (only 120 words were used, the rest for potential expansion). This was the last of the line of NECs dedicated FPUs (excluding the few MIPS FPUs they made).  Its a bit ironic that it seems they canceled as many designs as they made.

…but perhaps they didn’t?

Read More »

Introduction

On this site you can read about thousands of processors models. And every year it is more and more difficult to write about some new (old) processors, since everything has been known for a long time. But there are also exceptions to the rule which we love to find. In 2021, I learned about one unusual processor, the information about which I want to share with you. The roots of this processor’s history go back to Italy, in the distant year or 1998. This time just falls on the confrontation between Intel and its second generation Pentium and AMD K6-2 and K6-3 processors. The Cyrix MII processors from Cyrix Corporation, IDT WinChip 2s and Rise mP6s were still going strong as well.

But before we talk about the Genesys B52 MMX processor, we should take a closer look at Intel Pentium II processors in general, as the Italian processor primarily owes its appearance to them.

Intel Pentium II

From 1993 to 1997, the Pentium dominated all market segments. Over time, the name of the “Pentium” trademark even grew into a household name (Its all about the Pentiums baby), but with the release of the Pentium II, everything changed. If earlier Intel did not deeply segment the market, there were Pentium Pros for workstations and servers, and for everything else there were various models of Intel Pentium processors, in which, at the end of their domination, Intel added MMX instructions, depriving and thereby putting an end to its server segment. The new slot form factor of the processor, the abandonment of the usual pins and ceramics and further segmentation of the market (using Intel Celeron processors and the new Xeon line) radically changed the further course of development of the history of microprocessors.

May 7, 1997 saw the light of the first models of Intel Pentium II processors, manufactured on a 350nm process with a core voltage of 2.8 volts. The first models were based on the Klamath core (named after the river by which The CPU Shack is located) core, operating at 233 and 266 MHz. The main differences from the Pentium Pro predecessor it was based on were the L1 cache increased from 16 to 32 Kb, and the presence of a block of SIMD instructions called MMX first introduced on the last P55C processors. Like the Pentium Pro it featured its own L2 cache on the module, but in this case it was 512KB fixed on the same PCB as the processor core, a much cheaper solution then the dual ceramic cavity package of the Pentium Pro.

Before the Pentium II, only the Pentium Pro could boast of its own cache, running at the frequency of the CPU core. But, placing the CPU core and L2 cache on the same substrate was an expensive pleasure even for Intel, and the processors had to be cheaper for better competition, which was getting more and more intense. Intel then made a “wise” decision, as a result of which the Pentium II got a its own L2 cache next to the CPU core This engineering solution significantly reduced the cost of manufacturing processors. BSRAM L2 cache chips were manufactured by Toshiba, SEC and NEC at that time, rather then being made in house by Intel, further easing the cost burdens.

Pentium II Klamath SECC1 PBGA Core 2 x Cache on front 2x + TAG on back

For all models of Pentium II processors, the cache size remained unchanged and equaled 512 KB, while different Pentium Pro models had a cache from 256 to 1024 KB. The L2 cache of the first Pentium II processors consisted of four microcircuits located on both sides of the cartridge processor board and operated at half the core frequency. In addition to the processor core and 4 L2 cache chips, there was also a tag-RAM chip on the cartridge PCB, a total of 6 IC’s.

Backside with 2x cache + TAG

The tag-RAM size/configuration determines which range of main memory can be cached. For example, if the L2 cache is 256 KB and the tag RAM is 8 bits wide, then this is enough to cache up to 64 MB of main RAM. However, if you add additional RAM in the process, it will not be cached unless you also expand the tag RAM. On Socket 1-3 486 systems, most motherboards allowed adding and modifying additional L2 cache and tag-RAM chips for this purpose. The Pentium Pro had built-in L2 cache and tags capable of caching up to 4GB of main memory, whereas the first Pentium IIs could cache up to 512MB of RAM.  This was in part to set them apart from the server oriented Pentium II Xeon which had full speed cache capable of caching 4GB (or 64GB with PSE-36),

In January 1998, Intel announced the Pentium II processor, built on a new core, codenamed Deschutes (Another river in Oregon). The processor core was manufactured using the smaller 250nm process, which lowered the operating voltage to 2.0 V, instead of 2.8 V for “Klamath”. The L2 cache of 512 KB still worked at half the core frequency, but it was made in the form of two BSRAM chips located to the side of the processor package. In later modifications of the Pentium II Deschutes core, Intel replaced the tag-RAM chip, thanks to which the processors could cache up to 4 GB of RAM (the 82459AD revision).

The first generation of Intel Celeron processors were based on the “Covington” core were essentially processors on the “Deschutes” core, but without ANY L2 cache. Thanks to this, they had very poor performance, but they overclocked very well, demonstrating the best overclocking figures up to double the nominal clock frequency.

Deschutes core with Organic BGA core and 2x cache chips on front. TAG on back

All overclocking of Pentium II, as a rule, rested on the characteristics of microcircuits used by BSRAM and tag-RAMs. The latter, like the cache, was much disliked voltage rises, and with inept handling, an expensive Pentium II could turn out to be a Celeron “Covington”, if such microcircuits failed.By the way, they warmed up decently on Pentium II processors based on the “Klamath” core so cooling was very important as well. The multiplier in 99% of Pentium II processors was locked (very early production ones were unlocked and Engineering Samples of course), so overclocking was performed by raising the FSB frequency, this being dependent always on the cache and TAG chips installed in that particular processor.

A simple example. In Costa Rica, where Intel has an advanced advanced processor assembly/test factory, which simultaneously assembled high-frequency models with 450 and 300 megahertz. The cartridge and core for these processors are identical (and the multiplier was the same 4.5x as well 66×4.5 for the 300 and 100×4.5 for the 450). The difference was only in the installed cache memory with different speed rating in nanoseconds. Sometimes on the assembly line there was only a fast cache memory capable of operating at a frequency of 225 MHz, intended for models of processors with 450 MHz. In this case, it was also installed on the model with a frequency of 300 MHz, as a result of which they overclocked perfectly.

Genesys B52 MMX CPU

The history of the Italian processor began in the city of Monopoli, in the province of Bari in Italy. In 1998, Italian Marcello Console founded Genesys, which initially employed 10 people. The main idea of the Genesys business was the production of modified Intel Pentium II processors based on the “Deschutes” core, at a much lower price than the Pentium II ones of similar clock speed. Plus a warranty period extended to 3 years and productivity increased by 5% or more. It turns out to be a solid Attraction of Generosity!

Genesys had registered its own domain www.b52mmx.com and is getting ready to implement their processors in ready-made system units. Unfortunately, nothing is known about the manufacturing process, it remains a mystery to this day. There is not so much information on these processors, but let’s try to figure out what these processors were.

Read More »

Semiconductor Complex LTD SCL6502 CPU

India in the 1970’s was often considered a third world county, supported by a largely agrarian economy and with a wide swath of the population still based off of subsistence living.  They also however, had a robust space program, had mastered nuclear technology and had a largely stable government that supported the advancement of technology development in the country.  All the pieces were there to begin making the shift to the robust high tech economy that they possess today.  In the 1970’s India had several govt entities working on semiconductors and electronics, all managed under the direction of the Dept of Electronics.  There was also a fair number of companies with plants in India doing electronics manufacturer and assembly.  This was largely small scale production of older technology.  TTL circuits  (starting with the 7420) were made in Bangalore by BEL back in 1971.  But TTL circuits won’t get you far, and at that time the best process India had was around 8 microns, so in 1972 an initiative was started to develop an indigenous semiconductor industry within India.

SCL Fab – Currently 0.18 Micron

Politics are the same everywhere, and so this process took some time, people with experience had to be recruited to run it, and a suitable (politically and geographically) location selected.  Eventually in the late 1970’s the Semiconductor Complex LTD was formed in the city of Mohali ( Chandigarh ) in the Punjab province of India.  SCL was to be the state supported enterprise to bring indigenous high end (LSI and above) semiconductor production to India.  Two things were needed to make this work: Technology, and People who were experts in that field.  SCL was tasked with going to Japan, America, and Western Europe in search of a company that would assist with the technology transfer, as well as finding some Non-Resident Indians who would be willing to come back to India to work on it.  Many Indians had high skill jobs in the industry outside of India, and it turned out convincing them to come back to help their country was a non-issue (though generous incentives were provided).  Getting the technology on the other hand was a bit more work.

The first trip of the technology transfer team of SCL was to Hitachi in Japan.  Negotiations with Hitachi were grueling, and while not unproductive, did not yield the results SCL wanted.  Hitachi was happy to license some designs to SCL, for a high fee and royalties, but did not want to immediately help create the 3-5 micron production fab that SCL envisioned.  Hitachi called thei ‘one step at a time’  whereas the Indians wanted to go all in from the start.  Hitachi agreed only to help (some) with a 5 micon process) and only to license products for digital clocks and watches.  The SCL team then turned to the United States, likely expecting similar results.

The chosen company in the USA was AMI (American Microsystems Inc), a company with 7-8 times the turnover of Hitachi.  AMI was at the time the largest maker of custom ICs in America, as well as a very large provider of second source ICs  (such as the 6800 and 9900 CPUs).  AMI’s CEO Roy Turner readily agreed to help SCL, much to the surprise of their negotiation team, and on the very first day offered SCL AMI’s 5 micro CMOS and NMOS processes, with the option to license their 3 micron CMOS and NMOS processes within 4 years of the agreement becoming effective.  AMI also offered SCL access to all of AMI’s standard products catalog, as well as the possibility of joint development of additional products, all at a simple 50/50 split.  AMI even offered to help with the technology export license that would be required by the US State Dept to transfer the fab tech to India.  The agreement was signed in April of 1981.

Read More »