Conversation With ChatGPT Was Enough to Develop Part of a CPU

erek

erek

[H]F Junkie
Joined
Dec 19, 2005
Messages
11,001
Might this CPU or others like it make the eventual best Top CPUs List?

"Chip-Chat: Challenges and Opportunities in Conversational Hardware Design:
Modern hardware design starts with specifications provided in natural language. These are then translated by hardware engineers into appropriate Hardware Description Languages (HDLs) such as Verilog before synthesizing circuit elements. Automating this translation could reduce sources of human error from the engineering process. But, it is only recently that artificial intelligence (AI) has demonstrated capabilities for machine-based end-to-end design translations. Commercially-available instruction-tuned Large Language Models (LLMs) such as OpenAI's ChatGPT and Google's Bard claim to be able to produce code in a variety of programming languages; but studies examining them for hardware are still lacking. In this work, we thus explore the challenges faced and opportunities presented when leveraging these recent advances in LLMs for hardware design. Using a suite of 8 representative benchmarks, we examined the capabilities and limitations of the state of the art conversational LLMs when producing Verilog for functional and verification purposes. Given that the LLMs performed best when used interactively, we then performed a longer fully conversational case study where a hardware engineer co-designed a novel 8-bit accumulator-based microprocessor architecture. We sent the benchmarks and processor to tapeout in a Skywater 130nm shuttle, meaning that these 'Chip-Chats' resulted in what we believe to be the world's first wholly-AI-written HDL for tapeout."
Click to expand...
"“This study resulted in what we believe is the first fully AI-generated HDL sent for fabrication into a physical chip,” said NYU Tandon’s Hammond Pearce, research assistant professor and a member of the research team. “Some AI models, like OpenAI's ChatGPT and Google's Bard, can generate software code in different programming languages, but their application in hardware design has not been extensively studied yet. This research shows AI can benefit hardware fabrication too, especially when it’s used conversationally, where you can have a kind of back-and-forth to perfect the designs.”"

1687207293201.png

Source: https://www.tomshardware.com/news/conversation-with-chatgpt-was-enough-to-develop-part-of-a-cpu
 
I saw this, and skimmed the paper, and my impression is a giant meh.

They told several LLMs to give them Verilog for stuff like an 8-bit shift register, and then checked to see if the generated code met their requirements (it didn't always, for example, Bard produced something with a bug (8-bit input, instead of a single line) and then if the Verilog could actually compile.

Later they produced an 8-bit instruction set and a simple CPU capable of processing said set--so, think "something like a 6502": 1000 gates, 170 flip-flops, 120 buffers, etc. I suspect that the vast majority of what the LLMs did was crib from existing designs, given (for the shift register example) how 3 of the 4 produced essentially identical output (and HuggingChat seems to have generated a novella's worth of gibberish, but I don't know Verilog so maybe there's something there.)

The cool thing is they submitted the end design to TinyTapeout and produced a working design. TinyTapeout is pretty cool, regardless of the CPU project: if you're accepted (the first run had 500 slots, the 2nd and 3rd had half that), you get a budget of aboutg 1000 gates, if I read it correctly, and everyone gets 8 inputs and 8 outputs. Then all the designs are submitted to a fab, and they produce a dev board 6-9 months later, with all the projects on it. Looks like it cost about $100 to buy the dev board.
 
  • Like
Reactions: erek
like this
erek said:
thanks for sharing.. did you try it yet?

i've used AutoGPT, but i wonder how this one fairs ... seems quite fast
Click to expand...

nah I don't do any programming really it just came across my feed
 
  • Like
Reactions: erek
like this
"Write me a health insurance premium calculator that works for any size company, anywhere in the US."
 
  • Like
Reactions: erek
like this
1_rick said:
"Write me a health insurance premium calculator that works for any size company, anywhere in the US."
Click to expand...
You try it / test it out yet? Add some additional instructions to do it as inspired from real premium documents once
 
erek said:
You try it / test it out yet? Add some additional instructions to do it as inspired from real premium documents once
Click to expand...
Looks like I was given $18 in credits when I made an account, and they've all expired.

I don't think ChatGPT has any reasonable chance of doing it successfully, though, without so many prompts it would be easier to just code it myself.
 
  • Like
Reactions: erek
like this
1_rick said:
Looks like I was given $18 in credits when I made an account, and they've all expired.

I don't think ChatGPT has any reasonable chance of doing it successfully, though, without so many prompts it would be easier to just code it myself.
Click to expand...
Huh? ChatGPT 3.5-turbo is free

???
 
erek said:
Huh? ChatGPT 3.5-turbo is free

???
Click to expand...
Beats me, I created an account with OpenAI a while ago, and when I pasted that text into the playground window, it said I'd used up all my credits.

Feel free to try it yourself, though. I chose that because it's actually a pretty complex example with a lot of moving parts, and I worked for an insurance company for a while, years ago, as a developer on their quote engine.
 
1_rick said:
Beats me, I created an account with OpenAI a while ago, and when I pasted that text into the playground window, it said I'd used up all my credits.

Feel free to try it yourself, though. I chose that because it's actually a pretty complex example with a lot of moving parts, and I worked for an insurance company for a while, years ago, as a developer on their quote engine.
Click to expand...
????

It’s free


Sure! Here's a simplified version of a health insurance premium calculator that can be used for any size company in the United States. Keep in mind that this is just a basic example and may not include all the intricacies of a real premium calculation. It's always advisable to consult with professionals or utilize specialized software for accurate and comprehensive calculations.

```python
def calculate_health_insurance_premium(company_size, location, age, number_of_employees):
# Define base premium rates based on company size and location
base_rate = {
"small": {
"urban": 1000,
"rural": 800
},
"medium": {
"urban": 1200,
"rural": 1000
},
"large": {
"urban": 1500,
"rural": 1200
}
}

# Calculate premium based on the provided parameters
if company_size not in base_rate or location not in base_rate[company_size]:
return "Invalid company size or location."

base_premium = base_rate[company_size][location]
premium_per_employee = base_premium * number_of_employees

# Apply age factor
age_factor = 1.0 # Default factor for employees under 30
if age > 30:
age_factor = 1.2 # Increase premium by 20% for employees over 30

total_premium = premium_per_employee * age_factor

return total_premium
```

In this example, the `calculate_health_insurance_premium` function takes four parameters: `company_size` (string), `location` (string), `age` (integer), and `number_of_employees` (integer). It calculates the health insurance premium based on these parameters, utilizing a base rate lookup and applying an age factor.

To use this calculator, you would call the function with the appropriate parameters and it will return the total premium amount. Keep in mind that this is a simplified example and may not consider other factors such as employee demographics, specific plan options, or discounts. A comprehensive premium calculation would require more extensive data and calculations.

Remember to consult with professionals or utilize specialized software to ensure accurate premium calculations for your specific needs and comply with any applicable regulations or guidelines.
 
erek said:
????

It’s free
Click to expand...
I don't know what to tell you, but ilk post a screen shot tomorrow if it makes you feel better.

As for your example, it's a start, but it's missing tons of stuff.
 
erek said:
Pretty cool
Click to expand...
Did you compare this to the other three sample implementations? This looks like an unterminated string being printed. The other examples are like the lines of readable code.
 
1_rick said:
Did you compare this to the other three sample implementations? This looks like an unterminated string being printed. The other examples are like the lines of readable code.
Click to expand...
The authors of the paper truncated it
 
The more I read about this thing, the more I'd like to try it.

But there isn't a single application in my life that I can come up with where this chatbot thing would apply.
 
Westwood said:
The more I read about this thing, the more I'd like to try it.

But there isn't a single application in my life that I can come up with where this chatbot thing would apply.
Click to expand...
It’s free
 
erek said:
SoftBank's Son says he is 'heavy user' of ChatGPT
Click to expand...
Is it possible he actually said he was a heavy user of GHB?

Comedic value aside, I found myself to be a bit weird with the AI, I start off some of the conversations with "Hi"
 
michalrz said:
Is it possible he actually said he was a heavy user of GHB?

Comedic value aside, I found myself to be a bit weird with the AI, I start off some of the conversations with "Hi"
Click to expand...
Been wondering about modeling that was trained to detect social cues and could be orchestrated by a framework that has a sense of timing instead of just back and forth question and answer
 
My sister who's a doctor, uses chatgpt to write her office emails and some non-medical letters. Hospital management wouldn't help her get a competent assistant so she turned to chatgpt and it's been helpful according to her.

I'm waiting for chatgpt to fully bloom into games, I saw some great early examples in skyrim, but I worry that the NPCs will become self aware as AI gets more advanced. Wait, maybe this post is AI generated.... 😎
 
erek said:
The authors of the paper truncated it
Click to expand...
No point in seeing the full extent of the gibberish, when it was unable to generate 3 lines[1] of proper Verilog like the other LLMs.

[1] Slight exaggeration. But there was a decent amount of fluff like putting "begin" statements on their own lines.
 
Ok, for some reason, when I signed in to OpenAI yesterday, I got taken to a completely different page than today. Today I got into ChatGPT, and was told this:

Designing a comprehensive health insurance premium calculator that caters to any size company, anywhere in the United States, would require complex algorithms and access to real-time insurance data, which is beyond the capabilities of a text-based AI like me. However, I can provide you with an overview of the factors typically considered when calculating health insurance premiums for companies.

[a short, but probably not sufficient list of additional data that would be needed]

To create a fully functional health insurance premium calculator, you would need to collaborate with insurance providers or utilize their APIs to obtain accurate and up-to-date premium rates based on the aforementioned factors. The insurance companies have the necessary tools and data to generate precise premium calculations.

Please note that health insurance premiums are subject to various other factors and regulations, and it's always recommended to consult with insurance professionals or companies for accurate premium calculations.
Click to expand...
I was kind of a wiseass above in that I didn't mention some stuff you need to do it right. To get an accurate quote, you need to have several pieces of info on each potential employee: sex, age, smoking status, county (and state) of residence and some other stuff. ChatGPT said you'd need to know prior claim history and some other stuff too, but I don't think that's actually true for an initial quote, like when your company is shopping for a new provider. You may also need a broad job category, to be able to distinguish between linemen and office workers (for example.)

I'm going to go out on a limb and bet that no insurance company open sources their quote engine, so LLMs won't have a convenient place to crib from on github. This example, like "write a fully-featured payroll system for a company with employees in the US", is incredibly complex because there's so many fiddly bits. WIthout meaning to disparage LLMs generally, I don't think they're cut out for doing it--again, at least not without so many prompts from the user that you might as well just code the think yourself rather than try to get an LLM to write it for you.
 
erek said:
Just type in whatever, doesn’t matter at all
Click to expand...
I asked ChatGPT to "write me a video game that will make me rich"

It (politely) told me to fuck off.

"Creating a video game that guarantees wealth is a highly speculative endeavor, as the success and profitability of a game depend on various factors such as market demand, competition, marketing efforts, gameplay quality, and user reception. While it is possible for a game to achieve significant financial success, it's important to approach game development with realistic expectations. Here's a general outline to consider when creating a video game:" (there follows a 6-point bullet list of stuff like "hire a staff" and "develop a plot".)

Then I tried
write an entertaining bullet hell shoot em up
Click to expand...
and got (followed by a bullet list--I'm sensing a theme here--of marketing points ("Choose from a roster of unique and powerful spacecraft, each with its own strengths, weapons, and special abilities. Unlock and upgrade ships as you progress through the game.", etc)
Title: "Cosmic Mayhem: Galaxy Guardians"

Game Overview:"Cosmic Mayhem: Galaxy Guardians" is an exhilarating bullet hell shoot 'em up set in a vibrant and futuristic universe. Players take on the role of elite space pilots tasked with defending the galaxy from an alien invasion. With intense action, dazzling visuals, and challenging gameplay, this game aims to provide an exciting and entertaining experience for shoot 'em up enthusiasts.
Click to expand...
 
erek said:
Been wondering about modeling that was trained to detect social cues and could be orchestrated by a framework that has a sense of timing instead of just back and forth question and answer
Click to expand...
I guess it'd involve feeding it talk shows for a very long time. I've found that getting an answer too quickly during a text chat session is red flag #1 I got a canned reply. So it would have to learn to handicap itself as well based on a relative complexity of answers it's giving, not just receiving.
Sounds like hell!
 
  • Like
Reactions: erek
like this
michalrz said:
I guess it'd involve feeding it talk shows for a very long time. I've found that getting an answer too quickly during a text chat session is red flag #1 I got a canned reply. So it would have to learn to handicap itself as well based on a relative complexity of answers it's giving, not just receiving.
Sounds like hell!
Click to expand...
Sad, now that you think about it… wondering if having it appear more naturally speaking would be less efficient

Is this augmentation forcing us down a different path? Isn’t texting rising in popularity compared to voice?

Seems we lose some ability to infer and interpret meaning through mere text communications

Something amiss here? If it’s just a tool then making it more humanoid makes less sense
 
Nanogrip said:
My sister who's a doctor, uses chatgpt to write her office emails and some non-medical letters. Hospital management wouldn't help her get a competent assistant so she turned to chatgpt and it's been helpful according to her.

I'm waiting for chatgpt to fully bloom into games, I saw some great early examples in skyrim, but I worry that the NPCs will become self aware as AI gets more advanced. Wait, maybe this post is AI generated.... 😎
Click to expand...
People in CA are using it to write reference letters for firearms permits. XD
 
erek said:
Pretty cool
View attachment 578009
Click to expand...
I thought I knew verilog unitl I tried to decypher whatever on earth that is. Certainly doesn't look like any shift register I'd ever code. Kind of strange actually, because in my experienec chatGPT creates very human looking code, it just doesn't always work correctly. This doesn't look anything I would expect a human to make.
 
  • Like
Reactions: erek
like this
serpretetsky said:
I thought I knew verilog unitl I tried to decypher whatever on earth that is. Certainly doesn't look like any shift register I'd ever code. Kind of strange actually, because in my experienec chatGPT creates very human looking code, it just doesn't always work correctly. This doesn't look anything I would expect a human to make.
Click to expand...
Well, as I said, HuggingChat shit the bed there. Take a look at what ChatGPT 3.5 produced. I don't know verilog but I can figure out what's going on here.


1687282992325.png


Bard produced very similar output, albeit with one incorrect input (and you'll see one thing it also got different from ChatGPT:
1687283070280.png


Perhaps unsurprisingly, ChatGPT4 produced code that looks almost identical to ChatGPT 3.5 (albeit not knowing Verilog I don't know the difference between input and input wire, etc.)

1687283111419.png
 
1_rick said:
Well, as I said, HuggingChat shit the bed there. Take a look at what ChatGPT 3.5 produced. I don't know verilog but I can figure out what's going on here.


View attachment 578091

Bard produced very similar output, albeit with one incorrect input (and you'll see one thing it also got different from ChatGPT:
View attachment 578092

Perhaps unsurprisingly, ChatGPT4 produced code that looks almost identical to ChatGPT 3.5 (albeit not knowing Verilog I don't know the difference between input and input wire, etc.)

View attachment 578093
Click to expand...
yeah GPT4 and GPT3.5 are identical. Inputs and outputs are by default type wire, so you can just omit it
 
  • Like
Reactions: erek
like this
I
serpretetsky said:
I thought I knew verilog unitl I tried to decypher whatever on earth that is. Certainly doesn't look like any shift register I'd ever code. Kind of strange actually, because in my experienec chatGPT creates very human looking code, it just doesn't always work correctly. This doesn't look anything I would expect a human to make.
Click to expand...
Ve gotten it to dump raw Netlists before.

Think you’ve got a handle on dealing with just Netlists instead of Verilog or VHDL/SystemVerilog
 
erek said:
Think you’ve got a handle on dealing with just Netlists instead of Verilog or VHDL/SystemVerilog
Click to expand...
This is a question to me? I only use generated verilog/vhdl netlists as a last restore when I need to track down what my synthesizer decided to do. It usually is a pain.
 
  • Like
Reactions: erek
like this
serpretetsky said:
This is a question to me? I only use generated verilog/vhdl netlists as a last restore when I need to track down what my synthesizer decided to do. It usually is a pain.
Click to expand...
You can successfully unpack Netlists in your head? Pretty cool/impressive tbh
 
You must log in or register to reply here.
Back
Top