Naming the chip Jalapeño is great. Not only does it play into hot potato chips, but also Jalapeño peppers come from Xalapa, a city in Mexico. Xalapa comes from the Nahuatl word Xallapan, which roughly translates to "place of sandy water." These chips are made from quartz sand, and a data center then uses water to cool them. A data center is a place of sandy water!
This is a very verifiable task with well defined metrics and verifiable results. Chip designers are going to have just a rough time of it as software engineers.
It's just so obvious that with access to a timing and power simulator they will be able to do this just as well, as let's say opus 5.5 is at blackwell optimization. Which is to say, effectively superhuman.
By “rough time of it” do you mean chip designers are going to be inundated with slop that “works” but is overly complex, slowly degrades the system over time, and which nobody understands and this can fix? Or was it that they’d have to speak to insufferable AI agents who pretend to do things they don’t and also write in the most obtuse language possible?
I think these designers will be forced to deal with this slop, when they all collectively realize llms can produce more efficient designs then they could before the project tape out date.
People who work on high performance code in software had a similar attitude, but the problem is that they cannot compete with the ability of llms at this point.
I've heard this too, all the IC designers I know think there's some fundamental training missing that prevents the models from creating highly optimized circuits, which is what's needed for most tasks of interest.
Am I missing something because DEC apparently cracked this via expert systems in the 80s?
> SID (Synthesis of Integral Design) was a logic synthesis program used to generate logic gates for the VAX 9000. From high-level behavioral and register-transfer level sources, approximately 93% of the CPU scalar and vector units, over 700,000 gates, were synthesized.
> SID was an artificial intelligence rule-based system and expert system with over 1000 hand-written rules. In addition to logic gate creation, SID took the design to the wiring level, allocating loads to nets and providing parameters for place and route CAD tools. As the program ran, it generated and expanded its own rule-base to 384,000 low-level rules.
> Initially it was somewhat controversial but was accepted in order to reduce the overall VAX 9000 project budget. Some engineers refused to use it. Others compared their own gate-level designs to those created by SID, eventually accepting SID for the gate-level design job. Since SID rules were written by expert logic designers and with input from the best designers on the team, excellent results were achieved. As the project progressed and new rules were written, SID-generated results became equal to or better than manual results for both area and timing. For example, SID produced a 64-bit adder that was faster than the manually-designed one. Manually-designed areas averaged 1 bug per 200 gates, whereas SID-generated logic averaged 1 bug per 20,000 gates. After finding a bug, SID rules were corrected, resulting in 0 bugs on subsequent runs.
A lot trickier to do self supervised RL on chip design than Navier Stokes, or maybe they aren't trying hard enough. I mean even if you check their outputs with Altium and KiCad DRCs it won't capture everything that goes into a good board. Still need human training data which is slow and expensive.
I suspect before the end of next year they will have enough data to bootstrap a production level self supervised learning environment.
I wonder if they do training verification with real pick and place circuits too.
do you understand how large the training sets are for LLMs? there is not enough RTL in the world (let alone public RTL) to usefully contribute to the training of a O(T) model. and contrary to popular belief, HDLs are nothing like programming languages so nothing transfers.
It will be interesting to see if AI really starts innovating in circuit design or just continues doing the wrong things faster. But I'd bet that it will soon run circles around the status quo, and rightly so.
We're way out on the long tail of a dead-end approach to computation that started in the late 1990s with GPUs when we chose to optimize SIMD at the expense of general-purpose computing. On top of that, Moore's Law ended in roughly 2007 with the arrival of smartphones, when the industry chose cost and power minimization over compute maximization via a multicore approach.
Now only servers have access to highly multicore chips like AMD EPYC, at 10-100 times the cost and with poorer performance than if they maximized local/shared memories with content-addressable caching. Apple's M line of processors is ok, but its specialized cores just stand in the way of the higher abstractions available with symmetric multiprocessing, upon which specialized algorithms can be built. We're still missing the multicore foundation.
I used Synopsys and Cadence back in college in the late 1990s to automatically place and route circuits from diagrams in SPICE (I think, my memory is foggy) and they already performed 80% as well as humans. I don't remember if that was for chip area, speed or power. But the big players were obviously using tools at least as capable as whatever we were trained on.
Then in the 2000s I had hoped to do what OpenAI is doing and simulate small/open CPU designs on FPGAs. But the economics just weren't there. Even though some people like me (the poors) knew just exactly what could work, there was a tech winter between roughly the Dot Bomb around 2000 to the start of the Mobile Bubble in the early 2010s. So we didn't have a way to communicate what was possible to the people with capital, without having to network to get through the gatekeeping. Which may be fun for enthusiasts, but it's anathema to hackers wanting to get real work done.
I'd nearly given up all hope of progress until last year when AI went mainstream. Now I see 3 steps to the epiphany:
[x] Time
[ ] Money
[ ] Resources
AI solves the time component, since specs (ideas) are orders of magnitude easier to create than implementations.
But money and resources may not be available for quite some time. It's looking like the music may stop for AI funding next year, now that the election is over and the writing's on the wall as far as doing things without consent goes. Regulation is coming, and with it perhaps the last AI winter, although admittedly I'm probably exaggerating.
So as usual, we may have to pull ourselves up by our bootstraps and skip the money step to provide resources directly to ourselves and the rest of humanity. There's so much capital now that all it can do is inflate Wall Street. There's little to no carryover onto Main Street. So in a very real way, it's easier now to make money by borrowing and investing than going to a job. That's a dangerous situation for any country, as it slips into being a giant casino of haves and have-nots. We'll let the farms go under before raising taxes.
The interesting thing is that without my involvement, the universe went ahead and made AI without me, which is ..surprising. But automating problem solving was perhaps the hardest problem there was. The remaining problems are easier. We may be ready for Morris's Law of Manifestation, where the time it takes to manifest whatever you can think of keeps halving. So whatever tech or medical breakthrough you're wishing for will arrive in 12 years now, not 25. Then if you don't get it, it arrives in just 6 more years. Then 3, then 2, then 1, then we reach the Singularity. So whatever all this is will certainly end by 2050. I just hope that whatever's on the other side has less suffering rather than more.
Sorry this got so long. I've been thinking about it for a long time.
I agree with a lot of what you wrote as I've been thinking about this a long time too.
The conclusion I came to is none of this really solves the problem hardware designers like me have.
The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
And with people spitting out even larger gobs of RTL, and that getting that force multiplied by AI, is not making the verification problem any easier.
Here's how you can tell: every statement and unit in that doc is worthless.
Every graph is in terms of a reward or a loss. If you have an actual amazing improvement you put it on the y axis and show how much better you can drive it. Instead of a unit that is undefined.
And every statement is hedged to the 9s. Like "AMD also verified single-digit percentage gains in PPA" That's awesome! Except.. it continues with "across key blocks of its upcoming next-generation flagship CPU using designs written by our model". The improvement could have been zero. It could have by chance proposed something. Also... designs written by our model could have been inspiration for a human. It's totally impossible to know.
If they actually had something they would be bragging about it. Instead of whatever nonsense this is.
The fact your comment is grey 19 minutes in is irritating. It feels you're making at least a good faith post but presumably OpenAI employees are here ready to downvote
Tinfoil hat, whatever. Industry people definitely are here wanting to shape narratives
Eh. Announcing stuff is like pointing in a direction. When people don't like the direction (or do not believe that you can get there), it does not have to be to your benefit.
I hope it’s true
prompt: come up with a name for an inference chip that works across multiple leveling, think hard
It's just so obvious that with access to a timing and power simulator they will be able to do this just as well, as let's say opus 5.5 is at blackwell optimization. Which is to say, effectively superhuman.
People who work on high performance code in software had a similar attitude, but the problem is that they cannot compete with the ability of llms at this point.
> SID (Synthesis of Integral Design) was a logic synthesis program used to generate logic gates for the VAX 9000. From high-level behavioral and register-transfer level sources, approximately 93% of the CPU scalar and vector units, over 700,000 gates, were synthesized.
> SID was an artificial intelligence rule-based system and expert system with over 1000 hand-written rules. In addition to logic gate creation, SID took the design to the wiring level, allocating loads to nets and providing parameters for place and route CAD tools. As the program ran, it generated and expanded its own rule-base to 384,000 low-level rules.
> Initially it was somewhat controversial but was accepted in order to reduce the overall VAX 9000 project budget. Some engineers refused to use it. Others compared their own gate-level designs to those created by SID, eventually accepting SID for the gate-level design job. Since SID rules were written by expert logic designers and with input from the best designers on the team, excellent results were achieved. As the project progressed and new rules were written, SID-generated results became equal to or better than manual results for both area and timing. For example, SID produced a 64-bit adder that was faster than the manually-designed one. Manually-designed areas averaged 1 bug per 200 gates, whereas SID-generated logic averaged 1 bug per 20,000 gates. After finding a bug, SID rules were corrected, resulting in 0 bugs on subsequent runs.
https://en.wikipedia.org/wiki/VAX_9000#SID_Scalar_and_Vector...
Mathematics is generally done in the open. SOTA chip design is done by a handful of people all behind closed private doors.
I suspect before the end of next year they will have enough data to bootstrap a production level self supervised learning environment.
I wonder if they do training verification with real pick and place circuits too.
We're way out on the long tail of a dead-end approach to computation that started in the late 1990s with GPUs when we chose to optimize SIMD at the expense of general-purpose computing. On top of that, Moore's Law ended in roughly 2007 with the arrival of smartphones, when the industry chose cost and power minimization over compute maximization via a multicore approach.
Now only servers have access to highly multicore chips like AMD EPYC, at 10-100 times the cost and with poorer performance than if they maximized local/shared memories with content-addressable caching. Apple's M line of processors is ok, but its specialized cores just stand in the way of the higher abstractions available with symmetric multiprocessing, upon which specialized algorithms can be built. We're still missing the multicore foundation.
I used Synopsys and Cadence back in college in the late 1990s to automatically place and route circuits from diagrams in SPICE (I think, my memory is foggy) and they already performed 80% as well as humans. I don't remember if that was for chip area, speed or power. But the big players were obviously using tools at least as capable as whatever we were trained on.
Then in the 2000s I had hoped to do what OpenAI is doing and simulate small/open CPU designs on FPGAs. But the economics just weren't there. Even though some people like me (the poors) knew just exactly what could work, there was a tech winter between roughly the Dot Bomb around 2000 to the start of the Mobile Bubble in the early 2010s. So we didn't have a way to communicate what was possible to the people with capital, without having to network to get through the gatekeeping. Which may be fun for enthusiasts, but it's anathema to hackers wanting to get real work done.
I'd nearly given up all hope of progress until last year when AI went mainstream. Now I see 3 steps to the epiphany:
[x] Time
[ ] Money
[ ] Resources
AI solves the time component, since specs (ideas) are orders of magnitude easier to create than implementations.
But money and resources may not be available for quite some time. It's looking like the music may stop for AI funding next year, now that the election is over and the writing's on the wall as far as doing things without consent goes. Regulation is coming, and with it perhaps the last AI winter, although admittedly I'm probably exaggerating.
So as usual, we may have to pull ourselves up by our bootstraps and skip the money step to provide resources directly to ourselves and the rest of humanity. There's so much capital now that all it can do is inflate Wall Street. There's little to no carryover onto Main Street. So in a very real way, it's easier now to make money by borrowing and investing than going to a job. That's a dangerous situation for any country, as it slips into being a giant casino of haves and have-nots. We'll let the farms go under before raising taxes.
The interesting thing is that without my involvement, the universe went ahead and made AI without me, which is ..surprising. But automating problem solving was perhaps the hardest problem there was. The remaining problems are easier. We may be ready for Morris's Law of Manifestation, where the time it takes to manifest whatever you can think of keeps halving. So whatever tech or medical breakthrough you're wishing for will arrive in 12 years now, not 25. Then if you don't get it, it arrives in just 6 more years. Then 3, then 2, then 1, then we reach the Singularity. So whatever all this is will certainly end by 2050. I just hope that whatever's on the other side has less suffering rather than more.
Sorry this got so long. I've been thinking about it for a long time.
The conclusion I came to is none of this really solves the problem hardware designers like me have.
The problem is not circuits, or RTL, takes too long to write or design. The problem is circuits take too long to verify and validate. I haven't measured it, but I'm mostly sure even with a tool could take the spec to RTL time to zero, I would not be able to make even one more chip per year. Which is not to say that tool wouldn't be useful and a cost savings. It definitely would be. It's just to say the impact will be limited in current incarnations. Amdahl's law and all that.
And with people spitting out even larger gobs of RTL, and that getting that force multiplied by AI, is not making the verification problem any easier.
Curious how these partnerships work. AMD uploads its X Billion dollar IP and OpenAI pinky promises not to retain or train on it?
"... OpenTitan uses Ibex as the embedded processor ..."
Here's how you can tell: every statement and unit in that doc is worthless.
Every graph is in terms of a reward or a loss. If you have an actual amazing improvement you put it on the y axis and show how much better you can drive it. Instead of a unit that is undefined.
And every statement is hedged to the 9s. Like "AMD also verified single-digit percentage gains in PPA" That's awesome! Except.. it continues with "across key blocks of its upcoming next-generation flagship CPU using designs written by our model". The improvement could have been zero. It could have by chance proposed something. Also... designs written by our model could have been inspiration for a human. It's totally impossible to know.
If they actually had something they would be bragging about it. Instead of whatever nonsense this is.
Tinfoil hat, whatever. Industry people definitely are here wanting to shape narratives
I block ads everywhere else, yet this is somehow being paraded around as something of value.