6 ms·

Texas Instruments’ Biggest Blunder: The TMS9900 Microprocessor (2017)

127 points by YakBizzarro 4y ago

NonNefarious 4y ago
NoNoSong 4y ago
Too bad about the Motorola, but good note at the end about looking to the future.
dang 4y ago
Discussed at the time (of the article): Texas Instruments’ Biggest Blunder: The TMS9900 Microprocessor - https://news.ycombinator.com/item?id=14619360 https://news.ycombinator.com/item?id=14619360 - June 2017 (125 comments)
McGlockenshire 4y ago
A while back I decided to build a homebrew system based on the TMS99105A CPU, thanks both to this article and to Ben Eater's terrific videos. It's a pretty cool platform. It's been my first real exposure to working in assembly language and I've been having a blast. Due to the way the it provides bus status signals, you can effectively build a system with completely separate memory address spaces for executable code and data ... and a another separate address space for both of those when you use the built-in memory mapper support. It gets more fun if you can take advantage of the way it handles a subset of unimplemented opcodes: it shelves the current workspace and treats the opcode like a branch call. The 990/12 minicomputer and the 99110A CPU use this technique, branded "Macrostore," to add floating point instructions to the platform. That's three separate memory address spaces each for instructions and data... and another entirely separate address space for I/O devices! I really like this thing and one fine day I'll actually have something concrete enough to publish an article about it. If you want to learn more about the clever and insane things that 9900-series fans are doing, your best bet is heading over to the Atari Age forums where there are a bunch of homebrew software and hardware projects in progress: https://forums.atariage.com/forum/119-ti-994a-development/ https://forums.atariage.com/forum/119-ti-994a-development/ Somewhere buried in there you'll also find commentary from TI employees and other insiders about this article.
PaulHoule 4y ago
The TI-99/4A was unique among 1980s computers for having a distinctly different memory architecture. The 2022 Commander x16 similarly uses IO ports to access video RAM and eliminates some bottlenecks that kind of machine had but it has generous RAM, not a pittance.
- chasil 4y ago
  I actually used the DNOS operating system on the mini line where the 9900 originated. I flew around the country ripping these out of industrial equipment dealerships (replacing them with SCO), so I never saw any of the DNOS networking features. https://en.wikipedia.org/wiki/TI-DNOS https://en.wikipedia.org/wiki/TI-DNOS
- renewedrebecca 4y ago
  There were actually a ton of computers that used the same video chip (or its successors) that the TI-99/4A used. The entire MSX/MSX-2 series for one. Also, to be slightly pedantic, the X-16 doesn't have IO ports. It's using a 65C02 processor, where all I/O is memory mapped.
  - bogantech 4y ago
    > Also, to be slightly pedantic, the X-16 doesn't have IO ports. It's using a 65C02 processor, where all I/O is memory mapped. The video ram is not on the 65C02's bus apparently and is accessed through a set of registers which must slow down things quite a bit, you get plenty of video memory this way though I guess. Must be what they mean by "IO Ports" https://github.com/commanderx16/x16-docs/blob/master/VERA%20Programmer's%20Reference.md#video-ram-access https://github.com/commanderx16/x16-docs/blob/master/VERA%20...
    - PaulHoule 4y ago
      The X16 has a clever trick in that it has a memory mapped register that points to the next address to write to, and a register that causes anything you write to it to be written to that pointer in VRAM… and the pointer auto increments. So if you are doing a block copy it is as fast to write to VRAM. One advantage of this scheme is that the CPU is not sharing access to RAM with the video processor so ordinary RAM access is fast.
dtagames 4y ago
The chip, along with the rest of the TI-99/4A has been fully emulated in JS in this fantastic project you can run in the browser. It even has a debugger and all the original cartridge software, too. https://js99er.net/#/ https://js99er.net/#/
- nsxwolf 4y ago
  There's no better or faster way to experiment with a TI-99/4A today. Besides the great hardware emulation, a huge collection of the software released for the platform is available right there on the site.
  - retrocryptid 4y ago
    And you can map emulated disks to Google Drive folders. It's disturbing how much I enjoy writing TI-BASIC programs with half-way realistic data files I stuffed up on the GDrive.
- retrocryptid 4y ago
  The hours I have wasted playing Hunt the Wumpus on js99er! Oh!
jes 4y ago
I never knew that TI's DS-990[1] series of minicomputers was created for Ramada Inns. It was one of the first computers I used intensively and I thought it was great. I think I wrote code in almost every language they supported. Good times, and the TI documentation was quite good, in my view. Hit me with your WP (Workspace Pointer) register, baby! [1] https://en.wikipedia.org/wiki/TI-990 https://en.wikipedia.org/wiki/TI-990
- blep_ 4y ago
  As a person from The Future, making a whole new computer model specifically for one specific hotel chain is really weird to me.
  - AdamH12113 4y ago
    This was only a few years past the era where every computer was a full custom installation that took up a large part of a room.
  - foobarian 4y ago
    Many other things happened that way. My favorite were all these systems research efforts where entire operating systems, network file systems, directory services, etc were developed within individual companies or CS departments and unleashed upon the staff :). MIT Athena, anyone?
  - tomcam 4y ago
    Just wait until you learn about the Hilton offset:segment addressing mode and Marriott memory mapping
  - shrubble 4y ago
    Look up the Singer System 10...
    - retrocryptid 4y ago
      Slightly related... I heard someone at Tandy say that when they were pitching the TRS-80, they said "hey, if the computers aren't popular, we can give them to each store to use for inventory tracking." -- That was probably around '75 / '76 when they were pitching the idea. Same time frame.
jhallenworld 4y ago
I remember looking at this book at my local library when I was a kid: https://archive.org/details/tibook_how-to-build-your-own-working-16-bit-microcomputer https://archive.org/details/tibook_how-to-build-your-own-wor... So I knew about the CPU before I knew about the home computer.. Oh, one other thing I remembered: there were integrated injection logic (I2L) versions of the 9900: the SBP9900 (this was bugging me because the I thought the TMS9900 was I2L, but it's NMOS): https://www.cpushack.com/tag/sbp9900/ https://www.cpushack.com/tag/sbp9900/ Here is a datasheet, the I/O structures are interesting: https://source.z2data.com/2020/11/29/8/20/42/604/RSELSA00012168821.pdf https://source.z2data.com/2020/11/29/8/20/42/604/RSELSA00012... Runs off of 500 mA at 1V.
retrocryptid 4y ago
So... a couple of things: 1. Rhines implies TI never made an 8-bit CPU, but tried to "leapfrog" the industry and go direct to 16-bit. This isn't exactly true. TI competed with Intel on the Datapoint contract, but didn't win. Intel went on to evolve the 8008 into the 8080 and turn it into a commercial product. But TI shelved it's TMX1795 project. My guess is they already had a nice chip business and didn't need to risk anything on building a "pie in the sky" project like a single-chip microprocessor. I mean, imagine, a CPU on a single chip! How unrealistic! https://www.righto.com/2015/05/the-texas-instruments-tmx-1795-first.html https://www.righto.com/2015/05/the-texas-instruments-tmx-179... 2. The 9900 series was't a loser by the standards of the day. Sure, it didn't get picked up by IBM, which made it a loser by 1981 standards. But by 1978, it was picking up some design wins because for less than $100k, you could buy a 990 mini-computer with a full-fledged development system (Assembler, Linker, Pascal Compiler, etc.) The 8080 was a lot cheaper in volume, but development was slightly more difficult. Intel's 8080 development systems weren't really that great, though by early 1980 they were light years ahead of TI (you could get an official Intel development system for around $35k IIRC.) As an interesting aside, Marinchip Systems shipped a S-100 board with a 9900 (9940? 9995?) CPU. But it didn't take off, so they started selling software and changed their name to AutoCAD. http://www.s100computers.com/Hardware%20Folder/Marinchip/9900%20CPU%20board/9900%20CPU.htm http://www.s100computers.com/Hardware%20Folder/Marinchip/990... 3. The 99105 / 99110 (with a much larger address space) was definitely on the drawing board in '78. IBM probably asked "hey. what chips do you have right now?" 4. Everyone likes to rag on the 9900 for being a memory-to-memory architecture. But in the day, the plan was to put your register file in bipolar memory. This was before we all got the RISC religion, and it wasn't so obviously bad. 5. This article only scratches the surface of the dysfunction that was the TI-99/4. But... the industry learned a lot of what NOT to do by watching TI try to deliver a game console. But to recap. Sure. The 9900 had some problems. The 9980 kept having more and more problems. But the 9995 wasn't half bad. And the 99105 & 99110 weren't bad at all.
- tomcam 4y ago
  Analysis: true Username checks out bigly
- retrocryptid 4y ago
Stratoscope 4y ago
The article doesn't mention the TI 770 and later 771 Intelligent Terminal which were based on the TMS9900. The 771 with two 8" floppy drives is shown in this brochure: http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal/brochures/771_Intelligent_Terminal_Brochure_Jun78.pdf http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal... The 770 was the terminal part only, with tape cartridge drives in the two bays above the keyboard: http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal/silent_700/993023-9701_Operating_Instructions_for_Model_770_Intelligent_Data_Terminal.pdf http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal... They are the same machine other than the storage. Here's a look at the inside: https://forums.atariage.com/topic/312144-ti-771-intelligent-terminal/ https://forums.atariage.com/topic/312144-ti-771-intelligent-... The 770 was oriented toward business forms and interaction built with TI's TPL 700 (Terminal Programming Language). I put the 770 and TPL to good use at Tymshare in the 1970s. I'd been called down to our Houston office to design and build yet another prompt-and-response Teletype UI for one of our business customers. Being in Texas, the office just happened to have a TI 770 in a back room. When I saw it, I stayed up all night and used TPL to implement our data input forms onscreen instead of the Teletype interaction. Our sales guy loved it, and more importantly, so did our customers. The app went on to make the company a good bit of money for those days. One thing I didn't find in my web search: a TPL manual! If anyone finds one, please post it in a reply. Thanks!
TheOtherHobbes 4y ago
I interviewed with TI in the early 80s. They asked me about the 9900 and - thinking about those slow off-chip registers - I used the words "weird architecture". For some reason they seemed very unhappy about this. I'm very entertained that TI's own 9900 manager had the same opinion of it.
musicale 4y ago
> IBM’s Boca Raton group understood the momentum behind CP/M as an open standard, and so they commissioned Digital Research to develop a version, called CP/M-86. Later in the process, however, Microsoft came through with the MS-DOS operating system, about which much has already been written. As I understand it, 86-DOS/SDOS was basically an x86 clone of CP/M, and CP/M-80 source programs can even be automatically translated and then reassembled to run on DOS[1] or CP/M-86[2]. [1] https://github.com/ibara/8088ify https://github.com/ibara/8088ify [2] XLT86, Digital Research, Inc.
dreamcompiler 4y ago
I wrote assembler code for the 9900 when I worked at TI. I freaking loved that architecture: There was only one real register and it was a pointer to memory where the actual registers for the current thread were stored. This meant that a context switch only required the single master register to change. It was an absolutely glorious idea for writing multithreaded code. And also slow as frozen dog crap because every register access was a RAM access. Sigh.
cpswan 4y ago
The TMS9900 architecture lives on in TI's MSP430 line of MCUs, and they're everywhere. You probably won't get through today without interacting with a bunch of things that have MSP430s inside. https://www.cpushack.com/2015/02/05/ti-tms9900sbp9900-accidental-success/ https://www.cpushack.com/2015/02/05/ti-tms9900sbp9900-accide... "Like many processor designs, the 9900 didn’t die, it provided the basis of TI’s current 16-bit MCU line, the MSP430 series. The MSP430s instruction set is heavily influenced by the 9900 and largely retains its memory architecture and orthogonal instructions."

127 points by YakBizzarro 4y ago

NonNefarious 4y ago

NoNoSong 4y ago

Too bad about the Motorola, but good note at the end about looking to the future.

dang 4y ago

Discussed at the time (of the article): Texas Instruments’ Biggest Blunder: The TMS9900 Microprocessor - https://news.ycombinator.com/item?id=14619360 https://news.ycombinator.com/item?id=14619360 - June 2017 (125 comments)

McGlockenshire 4y ago

A while back I decided to build a homebrew system based on the TMS99105A CPU, thanks both to this article and to Ben Eater's terrific videos. It's a pretty cool platform. It's been my first real exposure to working in assembly language and I've been having a blast. Due to the way the it provides bus status signals, you can effectively build a system with completely separate memory address spaces for executable code and data ... and a another separate address space for both of those when you use the built-in memory mapper support. It gets more fun if you can take advantage of the way it handles a subset of unimplemented opcodes: it shelves the current workspace and treats the opcode like a branch call. The 990/12 minicomputer and the 99110A CPU use this technique, branded "Macrostore," to add floating point instructions to the platform. That's three separate memory address spaces each for instructions and data... and another entirely separate address space for I/O devices! I really like this thing and one fine day I'll actually have something concrete enough to publish an article about it. If you want to learn more about the clever and insane things that 9900-series fans are doing, your best bet is heading over to the Atari Age forums where there are a bunch of homebrew software and hardware projects in progress: https://forums.atariage.com/forum/119-ti-994a-development/ https://forums.atariage.com/forum/119-ti-994a-development/ Somewhere buried in there you'll also find commentary from TI employees and other insiders about this article.

PaulHoule 4y ago

The TI-99/4A was unique among 1980s computers for having a distinctly different memory architecture. The 2022 Commander x16 similarly uses IO ports to access video RAM and eliminates some bottlenecks that kind of machine had but it has generous RAM, not a pittance.

chasil 4y ago
I actually used the DNOS operating system on the mini line where the 9900 originated. I flew around the country ripping these out of industrial equipment dealerships (replacing them with SCO), so I never saw any of the DNOS networking features. https://en.wikipedia.org/wiki/TI-DNOS https://en.wikipedia.org/wiki/TI-DNOS
renewedrebecca 4y ago
There were actually a ton of computers that used the same video chip (or its successors) that the TI-99/4A used. The entire MSX/MSX-2 series for one. Also, to be slightly pedantic, the X-16 doesn't have IO ports. It's using a 65C02 processor, where all I/O is memory mapped.
- bogantech 4y ago
  > Also, to be slightly pedantic, the X-16 doesn't have IO ports. It's using a 65C02 processor, where all I/O is memory mapped. The video ram is not on the 65C02's bus apparently and is accessed through a set of registers which must slow down things quite a bit, you get plenty of video memory this way though I guess. Must be what they mean by "IO Ports" https://github.com/commanderx16/x16-docs/blob/master/VERA%20Programmer's%20Reference.md#video-ram-access https://github.com/commanderx16/x16-docs/blob/master/VERA%20...
  - PaulHoule 4y ago
    The X16 has a clever trick in that it has a memory mapped register that points to the next address to write to, and a register that causes anything you write to it to be written to that pointer in VRAM… and the pointer auto increments. So if you are doing a block copy it is as fast to write to VRAM. One advantage of this scheme is that the CPU is not sharing access to RAM with the video processor so ordinary RAM access is fast.

dtagames 4y ago

The chip, along with the rest of the TI-99/4A has been fully emulated in JS in this fantastic project you can run in the browser. It even has a debugger and all the original cartridge software, too. https://js99er.net/#/ https://js99er.net/#/

nsxwolf 4y ago
There's no better or faster way to experiment with a TI-99/4A today. Besides the great hardware emulation, a huge collection of the software released for the platform is available right there on the site.
- retrocryptid 4y ago
  And you can map emulated disks to Google Drive folders. It's disturbing how much I enjoy writing TI-BASIC programs with half-way realistic data files I stuffed up on the GDrive.
retrocryptid 4y ago
The hours I have wasted playing Hunt the Wumpus on js99er! Oh!

jes 4y ago

I never knew that TI's DS-990[1] series of minicomputers was created for Ramada Inns. It was one of the first computers I used intensively and I thought it was great. I think I wrote code in almost every language they supported. Good times, and the TI documentation was quite good, in my view. Hit me with your WP (Workspace Pointer) register, baby! [1] https://en.wikipedia.org/wiki/TI-990 https://en.wikipedia.org/wiki/TI-990

blep_ 4y ago
As a person from The Future, making a whole new computer model specifically for one specific hotel chain is really weird to me.
- AdamH12113 4y ago
  This was only a few years past the era where every computer was a full custom installation that took up a large part of a room.
- foobarian 4y ago
  Many other things happened that way. My favorite were all these systems research efforts where entire operating systems, network file systems, directory services, etc were developed within individual companies or CS departments and unleashed upon the staff :). MIT Athena, anyone?
- tomcam 4y ago
  Just wait until you learn about the Hilton offset:segment addressing mode and Marriott memory mapping
- shrubble 4y ago
  Look up the Singer System 10...
  - retrocryptid 4y ago
    Slightly related... I heard someone at Tandy say that when they were pitching the TRS-80, they said "hey, if the computers aren't popular, we can give them to each store to use for inventory tracking." -- That was probably around '75 / '76 when they were pitching the idea. Same time frame.

jhallenworld 4y ago

I remember looking at this book at my local library when I was a kid: https://archive.org/details/tibook_how-to-build-your-own-working-16-bit-microcomputer https://archive.org/details/tibook_how-to-build-your-own-wor... So I knew about the CPU before I knew about the home computer.. Oh, one other thing I remembered: there were integrated injection logic (I2L) versions of the 9900: the SBP9900 (this was bugging me because the I thought the TMS9900 was I2L, but it's NMOS): https://www.cpushack.com/tag/sbp9900/ https://www.cpushack.com/tag/sbp9900/ Here is a datasheet, the I/O structures are interesting: https://source.z2data.com/2020/11/29/8/20/42/604/RSELSA00012168821.pdf https://source.z2data.com/2020/11/29/8/20/42/604/RSELSA00012... Runs off of 500 mA at 1V.

retrocryptid 4y ago

So... a couple of things: 1. Rhines implies TI never made an 8-bit CPU, but tried to "leapfrog" the industry and go direct to 16-bit. This isn't exactly true. TI competed with Intel on the Datapoint contract, but didn't win. Intel went on to evolve the 8008 into the 8080 and turn it into a commercial product. But TI shelved it's TMX1795 project. My guess is they already had a nice chip business and didn't need to risk anything on building a "pie in the sky" project like a single-chip microprocessor. I mean, imagine, a CPU on a single chip! How unrealistic! https://www.righto.com/2015/05/the-texas-instruments-tmx-1795-first.html https://www.righto.com/2015/05/the-texas-instruments-tmx-179... 2. The 9900 series was't a loser by the standards of the day. Sure, it didn't get picked up by IBM, which made it a loser by 1981 standards. But by 1978, it was picking up some design wins because for less than $100k, you could buy a 990 mini-computer with a full-fledged development system (Assembler, Linker, Pascal Compiler, etc.) The 8080 was a lot cheaper in volume, but development was slightly more difficult. Intel's 8080 development systems weren't really that great, though by early 1980 they were light years ahead of TI (you could get an official Intel development system for around $35k IIRC.) As an interesting aside, Marinchip Systems shipped a S-100 board with a 9900 (9940? 9995?) CPU. But it didn't take off, so they started selling software and changed their name to AutoCAD. http://www.s100computers.com/Hardware%20Folder/Marinchip/9900%20CPU%20board/9900%20CPU.htm http://www.s100computers.com/Hardware%20Folder/Marinchip/990... 3. The 99105 / 99110 (with a much larger address space) was definitely on the drawing board in '78. IBM probably asked "hey. what chips do you have right now?" 4. Everyone likes to rag on the 9900 for being a memory-to-memory architecture. But in the day, the plan was to put your register file in bipolar memory. This was before we all got the RISC religion, and it wasn't so obviously bad. 5. This article only scratches the surface of the dysfunction that was the TI-99/4. But... the industry learned a lot of what NOT to do by watching TI try to deliver a game console. But to recap. Sure. The 9900 had some problems. The 9980 kept having more and more problems. But the 9995 wasn't half bad. And the 99105 & 99110 weren't bad at all.

tomcam 4y ago
Analysis: true Username checks out bigly
retrocryptid 4y ago

Stratoscope 4y ago

The article doesn't mention the TI 770 and later 771 Intelligent Terminal which were based on the TMS9900. The 771 with two 8" floppy drives is shown in this brochure: http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal/brochures/771_Intelligent_Terminal_Brochure_Jun78.pdf http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal... The 770 was the terminal part only, with tape cartridge drives in the two bays above the keyboard: http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal/silent_700/993023-9701_Operating_Instructions_for_Model_770_Intelligent_Data_Terminal.pdf http://bitsavers.informatik.uni-stuttgart.de/pdf/ti/terminal... They are the same machine other than the storage. Here's a look at the inside: https://forums.atariage.com/topic/312144-ti-771-intelligent-terminal/ https://forums.atariage.com/topic/312144-ti-771-intelligent-... The 770 was oriented toward business forms and interaction built with TI's TPL 700 (Terminal Programming Language). I put the 770 and TPL to good use at Tymshare in the 1970s. I'd been called down to our Houston office to design and build yet another prompt-and-response Teletype UI for one of our business customers. Being in Texas, the office just happened to have a TI 770 in a back room. When I saw it, I stayed up all night and used TPL to implement our data input forms onscreen instead of the Teletype interaction. Our sales guy loved it, and more importantly, so did our customers. The app went on to make the company a good bit of money for those days. One thing I didn't find in my web search: a TPL manual! If anyone finds one, please post it in a reply. Thanks!

TheOtherHobbes 4y ago

I interviewed with TI in the early 80s. They asked me about the 9900 and - thinking about those slow off-chip registers - I used the words "weird architecture". For some reason they seemed very unhappy about this. I'm very entertained that TI's own 9900 manager had the same opinion of it.

musicale 4y ago

> IBM’s Boca Raton group understood the momentum behind CP/M as an open standard, and so they commissioned Digital Research to develop a version, called CP/M-86. Later in the process, however, Microsoft came through with the MS-DOS operating system, about which much has already been written. As I understand it, 86-DOS/SDOS was basically an x86 clone of CP/M, and CP/M-80 source programs can even be automatically translated and then reassembled to run on DOS[1] or CP/M-86[2]. [1] https://github.com/ibara/8088ify https://github.com/ibara/8088ify [2] XLT86, Digital Research, Inc.

dreamcompiler 4y ago

I wrote assembler code for the 9900 when I worked at TI. I freaking loved that architecture: There was only one real register and it was a pointer to memory where the actual registers for the current thread were stored. This meant that a context switch only required the single master register to change. It was an absolutely glorious idea for writing multithreaded code. And also slow as frozen dog crap because every register access was a RAM access. Sigh.

cpswan 4y ago

The TMS9900 architecture lives on in TI's MSP430 line of MCUs, and they're everywhere. You probably won't get through today without interacting with a bunch of things that have MSP430s inside. https://www.cpushack.com/2015/02/05/ti-tms9900sbp9900-accidental-success/ https://www.cpushack.com/2015/02/05/ti-tms9900sbp9900-accide... "Like many processor designs, the 9900 didn’t die, it provided the basis of TI’s current 16-bit MCU line, the MSP430 series. The MSP430s instruction set is heavily influenced by the 9900 and largely retains its memory architecture and orthogonal instructions."