Can't ever imagine FPGA tech ever ever going away. Every modern jet fighter has them, and will continue to have them for decades to come.

Keatah wrote on 2022-01-01, 08:18:

Can't ever imagine FPGA tech ever ever going away. Every modern jet fighter has them, and will continue to have them for decades to come.

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be found in a blob on kids Speak and Spells before we will have a good enough cost / value proposition to use them in place of late 90’s era cpus.

I just don’t see enough 32bit single core speed in a reasonably priced FPGA yet to start making fast CPU replacements for late 90’s Pc tech, early 90’s sure, late 90’s not even close.

Perhaps an FPGA could handle transputer duties for modern to vintage silicon duties but I don’t believe oscillation speeds are high enough to be a glorified bus latch and flip flop with some glue, maybe wrong?

Despite all evidence to the contrary Moors law broke in 2009, we have been coasting on more highly parallel units for a decade, that is extraordinarily inefficient and has impacted bloat to the point you can’t run a plain graphics version of Solitaire on a 3ghz p4.
This slow down expands out to everything, so I’m not sure how many years it will take for a robust enough FPGA to move downstream to us mere mortals to upgrade beyond a 600mhz k6-3 for example. Or make faster voodoo cards for that matter.

Would love to be proved wrong

rmay635703 wrote on 2022-01-01, 14:49:

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be f […]
Show full quote

Keatah wrote on 2022-01-01, 08:18:

Can't ever imagine FPGA tech ever ever going away. Every modern jet fighter has them, and will continue to have them for decades to come.

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be found in a blob on kids Speak and Spells before we will have a good enough cost / value proposition to use them in place of late 90’s era cpus.

I just don’t see enough 32bit single core speed in a reasonably priced FPGA yet to start making fast CPU replacements for late 90’s Pc tech, early 90’s sure, late 90’s not even close.

Perhaps an FPGA could handle transputer duties for modern to vintage silicon duties but I don’t believe oscillation speeds are high enough to be a glorified latch and flip flop with some glue, maybe wrong?

Despite all evidence to the contrary Moors law broke in 2009, we have been coasting on more highly parallel units for a decade, that is extraordinarily inefficient and has impacted bloat to the point you can’t run a plain graphics version of Solitaire on a 3ghz p4.
This slow down expands out to everything, so I’m not sure how many years it will take for a robust enough FPGA to move downstream to us mere mortals to upgrade beyond a 600mhz k6-3 for example. Or make faster voodoo cards for that matter.

Would love to be proved wrong

Using a fpga as a translator to plug in say a real atom into a socket 7 seems like it would be totally doable considering it was what? 2005 or something when gigabyte used a fpga to make the iRam a thing.
https://www.anandtech.com/show/1742/2

Sphere478 wrote on 2022-01-01, 14:54:

rmay635703 wrote on 2022-01-01, 14:49:

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be f […]
Show full quote

Keatah wrote on 2022-01-01, 08:18:

Can't ever imagine FPGA tech ever ever going away. Every modern jet fighter has them, and will continue to have them for decades to come.

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be found in a blob on kids Speak and Spells before we will have a good enough cost / value proposition to use them in place of late 90’s era cpus.

I just don’t see enough 32bit single core speed in a reasonably priced FPGA yet to start making fast CPU replacements for late 90’s Pc tech, early 90’s sure, late 90’s not even close.

Perhaps an FPGA could handle transputer duties for modern to vintage silicon duties but I don’t believe oscillation speeds are high enough to be a glorified latch and flip flop with some glue, maybe wrong?

Despite all evidence to the contrary Moors law broke in 2009, we have been coasting on more highly parallel units for a decade, that is extraordinarily inefficient and has impacted bloat to the point you can’t run a plain graphics version of Solitaire on a 3ghz p4.
This slow down expands out to everything, so I’m not sure how many years it will take for a robust enough FPGA to move downstream to us mere mortals to upgrade beyond a 600mhz k6-3 for example. Or make faster voodoo cards for that matter.

Would love to be proved wrong

Using a fpga as a translator to plug in say a real atom into a socket 7 seems like it would be totally doable considering it was what? 2005 or something when gigabyte used a fpga to make the iRam a thing.
https://www.anandtech.com/show/1742/2

Doable at what R&D cost and volume?

Maybe the 50mhz Atom thing linked to was just a case of bad design ?

rmay635703 wrote on 2022-01-01, 15:01:

Sphere478 wrote on 2022-01-01, 14:54:

rmay635703 wrote on 2022-01-01, 14:49:

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be f […]
Show full quote

I certainly don’t see them going away but now powerful versions like the VU19P FPGA would have to move downstream enough to be found in a blob on kids Speak and Spells before we will have a good enough cost / value proposition to use them in place of late 90’s era cpus.

I just don’t see enough 32bit single core speed in a reasonably priced FPGA yet to start making fast CPU replacements for late 90’s Pc tech, early 90’s sure, late 90’s not even close.

Perhaps an FPGA could handle transputer duties for modern to vintage silicon duties but I don’t believe oscillation speeds are high enough to be a glorified latch and flip flop with some glue, maybe wrong?

Despite all evidence to the contrary Moors law broke in 2009, we have been coasting on more highly parallel units for a decade, that is extraordinarily inefficient and has impacted bloat to the point you can’t run a plain graphics version of Solitaire on a 3ghz p4.
This slow down expands out to everything, so I’m not sure how many years it will take for a robust enough FPGA to move downstream to us mere mortals to upgrade beyond a 600mhz k6-3 for example. Or make faster voodoo cards for that matter.

Would love to be proved wrong

Using a fpga as a translator to plug in say a real atom into a socket 7 seems like it would be totally doable considering it was what? 2005 or something when gigabyte used a fpga to make the iRam a thing.
https://www.anandtech.com/show/1742/2

Doable at what cost and volume?

Maybe the 50mpz Atom thing linked to was just a case of bad design ?

In the link they were simulating a entire atom cpu. I’m talking about using a real atom and just using a fpga to translate the bus signals it wouldn’t have to exactly be a atom, could be any number of cpus.. pentium III or dothan comes to mind.

Sphere478 wrote on 2022-01-01, 15:02:

rmay635703 wrote on 2022-01-01, 15:01:

Sphere478 wrote on 2022-01-01, 14:54:

Using a fpga as a translator to plug in say a real atom into a socket 7 seems like it would be totally doable considering it was what? 2005 or something when gigabyte used a fpga to make the iRam a thing.
https://www.anandtech.com/show/1742/2

Doable at what cost and volume?

Maybe the 50mpz Atom thing linked to was just a case of bad design ?

In the link they were simulating a entire atom cpu. I’m talking about using a real atom and just using a fpga to translate the bus signals

That makes sense then, you would still need an FPGA that can handle the required bus speeds in a stable fashion which moves into the higher end, not sure if anyone here knows a given FPGA inside and out to make a test sample, likely have to start simple like the transputer 386-486 module before we move to higher speeds.

Wonder if the “Atom Folks” have the design so one could copy just the bus translation side of things?

Some little thought would be what's essentially a 50MHz bus speed S3 486 with 10x multiplier (on 2x setting) and 8 or 16MB L1 cache.
Or a 386 at 40MHz x10 with 4MB L1...
Or a 133x10 S7 with 32MB L1.
Or a 2GHz Pentium 3...

Tech has shrunk to minimum so with let's say 65nm tech production for 10.000 units would be more expensive than desirable but doable as proof of concept. But a base design should be made available by the big boys first...

rmay635703 wrote on 2022-01-01, 14:49:

Despite all evidence to the contrary Moors law broke in 2009, we have been coasting on more highly parallel units for a decade, that is extraordinarily inefficient and has impacted bloat to the point you can’t run a plain graphics version of Solitaire on a 3ghz p4.
This slow down expands out to everything, so I’m not sure how many years it will take for a robust enough FPGA to move downstream to us mere mortals to upgrade beyond a 600mhz k6-3 for example. Or make faster voodoo cards for that matter.

Moore's Law hasn't been dead that long, what's dying is Dennard Scaling, and we also pretty much ran out of ideas for making singlecore CPUs faster. Moore's Law is why we can have more transistors in the same space for things like 8 CPU cores on a single die, Dennard Scaling is the law that states those smaller transistors use less power. Less power -> less heat -> higher achievable clocks.

Bloat is unrelated to multiple CPUs and is pretty much entirely a software choice. Multithreaded software is necessarily more complex to program, resulting in developers using layers of abstraction, but there's also no reason for a solitaire game to multithread.

Bloat increases are primarily a result of higher and higher level programming and greater levels of abstraction, which results in programs that would run slower on the same hardware, but since hardware keeps getting faster all the time it basically runs the same.

This is to save on developer time when programming it, not every dev can afford to be Chris Sawyer and write their entire game from scratch in assembly. For similar reasons, not every dev can afford to write their entire game from scratch at all, and may end up using bloated overkill junk like Unity for a 2D game.

Meanwhile FPGAs actually have gotten a lot better in the past 12 years, though consumer level stuff isn't at the level of the ultra high-end from back then. FPGAs have benefited a lot from the continued march of semiconductor density, which is why things like FPGA clone consoles and MISTer are possible when they weren't in decades past.

x86 CPUs beyond the 8086/8088 are just a really hard thing to try to put in an FPGA as a result of their CISC nature, high clock speeds, and overall bloated and legacy design. They have boatloads of logic and they try to clock it really fast, and the FPGA implementation can't afford to make many functionality compromises to fit better since we're aiming for better-than-original performance and 100% compatibility. RISCs and simpler and slower CISCs like 6502, z80, 8088, 68k fit much better on FPGAs as a result of being easier to implement and being less demanding in terms of logic and/or clock speed requirements. In regards to implementation complexity, it kinda speaks for itself that there are a bunch of 6502, 8088, and Risc-V cores, and only one ao486.

If the IBM PC lineage had used the 68k we'd already be able to put new CPUs in old motherboards with the help of the 68080 softcore (though confusingly according to the advertised "model number" it isn't actually as good or better than the 68060, and it's proprietary so you'd have to persuade them to let you put it on a socket)

Interesting point about RISC/CISC, but even in FPGA RISC cores are slower than custom silicon. With later x86 cores becoming RISC with instruction set broadening done by microcode, I wonder if one could take a common RISC core, ARM or RISC V, with a high speed barebones implementation, then have a FPGA working as the microcode unit translating x86 instructions.

BitWrangler wrote on 2022-01-01, 22:14:

Interesting point about RISC/CISC, but even in FPGA RISC cores are slower than custom silicon.

Definitely true. A 65nm FPGA can just about squeak out 75MHz for a RISC-V core, roughly what was being done with 800nm processes by MIPS/SPARC.

BitWrangler wrote on 2022-01-01, 22:14:

With later x86 cores becoming RISC with instruction set broadening done by microcode, I wonder if one could take a common RISC core, ARM or RISC V, with a high speed barebones implementation, then have a FPGA working as the microcode unit translating x86 instructions.

Maybe. It might also be possible to just emulate an x86 CPU in software on one of the Zynq FPGAs that has a real hardware ARM core integrated into it. On that note there's also some Intel FPGAs with hardware Atoms on them as a separate die, but I don't think they're integrated as tightly as the ARM on the Zynq, which may make implementation difficult. The Intel ones also definitely not wired in a way that would let you use the FPGA as a bus translator from the Atom FSB to Socket5/Socket7.

Looks like we might get those intel 4004 garage made clones after all xD

https://www.wired.com/story/22-year-old-build … ents-garage/amp

Zeloof recently upgraded his photolithography machine to print details as small as about 0.3 microns, or 300 nanometers—roughly on par with the commercial chip industry in the mid-'90s. Now, he’s thinking about the functions he could build into a chip on the scale of Intel’s historic 4004. “I want to push garage silicon further and open people’s minds to the possibility that we can do some of this stuff at home,” he says.

I repeat it again...

W.x. wrote on 2021-11-12, 18:37:

So, chip like Intel 4004 is possible doing at home by DIY, maybe by one genial , rich individual, or maybe small team, but I think, 8008 is highest, where you can get. You wont get even on xt, or 286 chips, in house conditions...

But I doubt he will make even working 4004 chip. He will make it only with such similiar miniaturized process (in other words, transistors will be about as small, as in 4004), but that will be all. His chip will be worse than Intel one, and more important topic, I would say, you don't get a one. So I will be suprised, if he even get on similiar chip as 4004, but let it give a chance. As I've said, it can be done by one genial (that means one of 1 milion) indivudual, there is a chance. You won't get enough such people, so it can be produced in mass quantities, so it will be impossible to get to working chip anyway. In relation in this topics, which is about you'll get those chips and cheaper than on current retro market, which was the point. It's too difficult to collect such mini-laboratory and factory, there will be not many such people. So even if he manage to make Intel 4004 chip, it will be very expensive.

AppleSauce wrote on 2022-01-21, 08:04:

Looks like we might get those intel 4004 garage made clones after all xD

https://www.wired.com/story/22-year-old-build … ents-garage/amp

Zeloof recently upgraded his photolithography machine to print details as small as about 0.3 microns, or 300 nanometers—roughly on par with the commercial chip industry in the mid-'90s. Now, he’s thinking about the functions he could build into a chip on the scale of Intel’s historic 4004. “I want to push garage silicon further and open people’s minds to the possibility that we can do some of this stuff at home,” he says.

His jump from 10um to 350nm is a mistake, IMO, even if he has loads of money to buy used equipment. At 350um, that is a process node still being used in parts of the world, and used equipment will likely command a price that makes this not something a casual person will ever do in the foreseeable future. All this is proving really, is he has money to spend, and yes, if you can purchase at this level, you probably can start thinking about opening your own fab anyway. But who knows, with the semi shortage, it's possible these machines are flying out of warehouses right now, so this level is probably not a good idea for a hobbyist.

The 350nm node is a transitional period. Lots of smart people spent alot of effort correcting it and perfecting the design techniques to get chips done right. These are design techniques still in use today, albeit using much faster PCs to do the work. So while he might be able to pull off a very simple CPU like the 4004 on his machines that can output at 350nm eventually, since he has the machine, the cost of money and time, and the difficulty of getting there won't really be one that would gather the attention of people wanting to make chips, but only of where he can say, hey I did it, cause I have money and time.

My recommendation would be for him to target say, 1um process, one that is very capable for the level he is targeting, and learn how to manufacture NEW equipment using off the shelf modern parts, so that whatever he is doing is repeatable. That is something I'd look at. Or even better, have a 1um process that makes chips for equipment capable at 1um or beyond. Then we can discuss whether this is a thing hobbyist can start doing. I mean, honestly, if TTL / 5V CMOS chips are manageable with what he has done, the more people doing it the better. I was looking to get some logic ICs, and they were all on back order on digikey, so I ended up paying 2X (about $1 per chip) from ebay instead.

Can someone show me this forum's projects the OP is talking about ?
I found the open source projects list but that's it.
Thank you.

JustJulião wrote on 2022-05-05, 15:35:

Can someone show me this forum's projects the OP is talking about ?
I found the open source projects list but that's it.
Thank you.

The comprehensive list of all modern reproductions of vintage sound cards

Newly made 4MB 30pin SIMMs

BitWrangler wrote on 2022-01-01, 22:14:

Interesting point about RISC/CISC, but even in FPGA RISC cores are slower than custom silicon. With later x86 cores becoming RISC with instruction set broadening done by microcode, I wonder if one could take a common RISC core, ARM or RISC V, with a high speed barebones implementation, then have a FPGA working as the microcode unit translating x86 instructions.

I believe the AMD K5 is also a RISC processor (Am29000) running the x86 core? Personally, I need to get one of those at some point to expand my collection of RISC processors, and now subcollection of RISC Processors That Emulate The x86 Instruction Set, along with my Efficeon and I guess Pentium II and my i5-2400 and later, but those kinda feel like cheating. Either way, it's almost certain that that would work, the question is how quickly -- especially as that adds latency as any emulation does. I suppose it may not be any worse than a K5 or a Pentium Pro.

I kinda believe that we'll see FPGAs start to break the speed barrier in 2023 when the new fabs open up -- if for no other reason than prices lowering for the higher end models. I'm not holding my breath, though. In any case, FPGA processors are now fast enough to run a CPU core and Linux -- see the Microwatt.

Even though there would be basically zero software (including OSes) for it, I'd love to see e.g. something around a PowerPC 7450 or Microwatt or the likes on a Slot A/Slot 2 cartridge. Or maybe something along the lines of that PowerPC 6-something CPU I can't quite recall the name of that had an x86 core packaged with it. PowerPC is just my bread and butter, and I'd love to have more non-Apple examples of it.

I know it's definitely not one hobbyist in a room-level stuff, but it would still be cool to see. Modern CPUs are obviously kinda out of the scope of a motherboard of that like, but maybe some of the other architectures around at the time, or a cut down version of such, might be doable. Even if it needs new boards, so really the socket just stays the same but it's essentially just a new, low-end hobbyist PC along the lines of the sam440ep. It would certainly be cheaper than a Blackbird or Talos II.

The simple solution would seem to be license the design's from Intel or AMD and produce it on a smaller process node allowing it to run at higher clock speeds but that runs into other issues, but that seems the most likely method for a faster K6 or 486

candle_86 wrote on 2022-11-18, 04:43:

The simple solution would seem to be license the design's from Intel or AMD and produce it on a smaller process node allowing it to run at higher clock speeds but that runs into other issues, but that seems the most likely method for a faster K6 or 486

Wonder which one would be the go to in that case, k6-3+ or mII

3+ has cache but mII has some 686 instructions.

If both are on a new process clocked to the moon the cyrix may be able to compete.
Heck, I’ve played halo on a mII at 350mhz.

Sphere478 wrote on 2022-11-18, 05:30:

Wonder which one would be the go to in that case, k6-3+ or mII […]
Show full quote

candle_86 wrote on 2022-11-18, 04:43:

The simple solution would seem to be license the design's from Intel or AMD and produce it on a smaller process node allowing it to run at higher clock speeds but that runs into other issues, but that seems the most likely method for a faster K6 or 486

Wonder which one would be the go to in that case, k6-3+ or mII

3+ has cache but mII has some 686 instructions.

If both are on a new process clocked to the moon the cyrix may be able to compete.
Heck, I’ve played halo on a mII at 350mhz.

possibly, but the other issue with a smaller process would be voltage, you'd have to also build an interposer, a K6-3 on say 45nm would likely kill itself at 2.0V or at the more common 2.4V, it would likely need to run far lower at 1.2V This might be an issue even just going from 180nm to 130nm