Will The AI Wars Be The Next Core Wars ?
video description
Date: 2023-06-27
Related videos
Comments and reviews: 20
MrVipitis
AI might be able to fix intros, you should see the cutting edge research that is demoed on the trending huggingface spaces or twitter.
If not using the tech demoed by nvidia
AI Benchmarks does mean something different for me. On the research side, we benchmark how good different models are - not the hardware. And it's a really big question if those evaluations actually mean something. And the technical backend of those is a horrible struggle for me. I got an Intel A750 in my home machine, and that just doesn't do PyTorch. At least not without IPEX which isn't released on Windows. There is OpenVINO which claims CausalLM support --- but it's much slower than CPU for me. Perhaps due to my old CPU. Meaning I have to do my evaluations on smaller datasets (300 instead of 10000) or request some HPC time at the university. So I am really really critical towards Intel and what they want to do for consumers in their next products. Especially if the developer support right now isn't great. Intel does host and OpenVINO webinar on language model inference this coming Wednesday and I do have a bunch of questions.
reply
AI might be able to fix intros, you should see the cutting edge research that is demoed on the trending huggingface spaces or twitter.
If not using the tech demoed by nvidia
AI Benchmarks does mean something different for me. On the research side, we benchmark how good different models are - not the hardware. And it's a really big question if those evaluations actually mean something. And the technical backend of those is a horrible struggle for me. I got an Intel A750 in my home machine, and that just doesn't do PyTorch. At least not without IPEX which isn't released on Windows. There is OpenVINO which claims CausalLM support --- but it's much slower than CPU for me. Perhaps due to my old CPU. Meaning I have to do my evaluations on smaller datasets (300 instead of 10000) or request some HPC time at the university. So I am really really critical towards Intel and what they want to do for consumers in their next products. Especially if the developer support right now isn't great. Intel does host and OpenVINO webinar on language model inference this coming Wednesday and I do have a bunch of questions.
reply
John
People dont understand what AI is..
Picture upscaling is AI.. But it is basically a software..
A TV that upscales and uses AI, is using a software that analyses the content and choses an upscaling algorithm. Nothing more.
Software that can emulate a voice is AI.. Again, it is just a software
Software that can put a face on another's body and make a video is AI, but again, it is just a software..
Software can help create fake pictures and it is called AI, but it is software..
The real danger of AI is that in a decade you will not be able to tell if a picture or a video is real or fake. Thats it.
Today you could already do fake bad quality surveillance videos.. People can take your picture and suddenly you are drunk on the street on a video.. And it is all fake..
This is the danger of AI.. Not what AI is doing, but what people are using the AI software for..
All these programs are called AI. But they are all software.
And the reason for the hype is self made. The programmers want to sell their program. Nothing more.
There is no benefit of AI to the average person.
reply
People dont understand what AI is..
Picture upscaling is AI.. But it is basically a software..
A TV that upscales and uses AI, is using a software that analyses the content and choses an upscaling algorithm. Nothing more.
Software that can emulate a voice is AI.. Again, it is just a software
Software that can put a face on another's body and make a video is AI, but again, it is just a software..
Software can help create fake pictures and it is called AI, but it is software..
The real danger of AI is that in a decade you will not be able to tell if a picture or a video is real or fake. Thats it.
Today you could already do fake bad quality surveillance videos.. People can take your picture and suddenly you are drunk on the street on a video.. And it is all fake..
This is the danger of AI.. Not what AI is doing, but what people are using the AI software for..
All these programs are called AI. But they are all software.
And the reason for the hype is self made. The programmers want to sell their program. Nothing more.
There is no benefit of AI to the average person.
reply
John
So, if AI is in my PC, I want it to do something useful for me. Otherwise I don't care and it's just added expense.
I want acceleration in the CPU, but I want it to have real effects on applications I use. If not it's just added expense.
More than anything I want AMD to make big-little for desktop, because there IS a big advantage when you look at a 5600 vs. a 13600K. Forgetting power consumption which isn't the Intel architecture but the node that it's on, the 13600K is a superior product because it can do a bit more than a 5600 or complete multi-threaded apps faster or handle more multi-tasked apps (running at the same time) easier than the 5600.
THIS excites me, because I could be able to buy something like a 8800X with 8-4 and 16c/24t, or even more exciting an 8800X3D which is 16c/24t, runs games faster and can run lightly threaded apps in the background that aren't using the same resources as the game. That excites me.
Acceleration and AI only excite me if they provide REAL benefit to the apps I'm running or even if it helps with gaming (AI probably never acceleration).
reply
So, if AI is in my PC, I want it to do something useful for me. Otherwise I don't care and it's just added expense.
I want acceleration in the CPU, but I want it to have real effects on applications I use. If not it's just added expense.
More than anything I want AMD to make big-little for desktop, because there IS a big advantage when you look at a 5600 vs. a 13600K. Forgetting power consumption which isn't the Intel architecture but the node that it's on, the 13600K is a superior product because it can do a bit more than a 5600 or complete multi-threaded apps faster or handle more multi-tasked apps (running at the same time) easier than the 5600.
THIS excites me, because I could be able to buy something like a 8800X with 8-4 and 16c/24t, or even more exciting an 8800X3D which is 16c/24t, runs games faster and can run lightly threaded apps in the background that aren't using the same resources as the game. That excites me.
Acceleration and AI only excite me if they provide REAL benefit to the apps I'm running or even if it helps with gaming (AI probably never acceleration).
reply
Josh
I gotta say I think AI is distinctly different because the value is always in the eyes of the beholder, and AI will be of varying use to everyone, depending on their use case. Straightforward performance on the other hand is universally applicable. I gotta say I'm not on the AI hype train, if only because I think companies will implement it for more aesthetic reasons rather than improving the amount of work you can do with a computer, at least in the consumer space, not so much on the enterprise level. I think one of Dr. Cutress's earliest comments suggests at the main point here: 'AI' is a poorly defined buzzword which cannot mean anything to the consumer in-and-of itself. Machine learning undoubtedly has potential in the computer space, but I think I'll wait until the tech is mature enough for companies to emphasize use case instead of 'experiences'.
reply
I gotta say I think AI is distinctly different because the value is always in the eyes of the beholder, and AI will be of varying use to everyone, depending on their use case. Straightforward performance on the other hand is universally applicable. I gotta say I'm not on the AI hype train, if only because I think companies will implement it for more aesthetic reasons rather than improving the amount of work you can do with a computer, at least in the consumer space, not so much on the enterprise level. I think one of Dr. Cutress's earliest comments suggests at the main point here: 'AI' is a poorly defined buzzword which cannot mean anything to the consumer in-and-of itself. Machine learning undoubtedly has potential in the computer space, but I think I'll wait until the tech is mature enough for companies to emphasize use case instead of 'experiences'.
reply
yumri4
For what to show for the benchmark i think an image you can make always happen with Stable diffusion 1.5 would be good. Iterations per second over 1000 images average also showing the curve of when it got better and worse as most AI front ends will make that happen.. The reason 1000 is the newer cards will make those images quickly.
The reason why Stable diffusion 1.5 is it is still the base to most AIs on huggingface aa research AI model site. So you can just pull one from that for free like most others probably did.
The interfaces i have used require you to select the device to use in setup / install though you can change them in most in the initialization file.
reply
For what to show for the benchmark i think an image you can make always happen with Stable diffusion 1.5 would be good. Iterations per second over 1000 images average also showing the curve of when it got better and worse as most AI front ends will make that happen.. The reason 1000 is the newer cards will make those images quickly.
The reason why Stable diffusion 1.5 is it is still the base to most AIs on huggingface aa research AI model site. So you can just pull one from that for free like most others probably did.
The interfaces i have used require you to select the device to use in setup / install though you can change them in most in the initialization file.
reply
TekEagle21
Well, at least I am 2nd. Just got off work and will watch the video when I get back to my dorm.
Edit: I wonder how the AI Wars will turn out. Will it be similar to CPU cores? More cores with the same amount of power produces higher benchmarks, but with lower stability, like with the i9-12900HX vs i9-13900HX laptop CPUs (RIP old naming scheme. This will make my massive comparison spreadsheets more difficult). Also, what is the difference between AI acceleration and hardware acceleration? I am a tech person, and I am still confused with recent changes in technology.
reply
Well, at least I am 2nd. Just got off work and will watch the video when I get back to my dorm.
Edit: I wonder how the AI Wars will turn out. Will it be similar to CPU cores? More cores with the same amount of power produces higher benchmarks, but with lower stability, like with the i9-12900HX vs i9-13900HX laptop CPUs (RIP old naming scheme. This will make my massive comparison spreadsheets more difficult). Also, what is the difference between AI acceleration and hardware acceleration? I am a tech person, and I am still confused with recent changes in technology.
reply
ikjadoon
Ah, I love these series. Cheers for having two great speakers for this conversation. On benchmarking, I'd hope for tight integration with OS AI SDKs, as those seem like the universal language that any useful AI hardware should be optimized for its background removal, noise cancelling, etc. I hope UL (3DMark.), Primate Labs (Geekbench), etc. will help spit out numbers. MacOS also has a similar AI / neural something SDK, but it will auto-select CPU vs GPU vs NPU for running the task, but I imagine that's also exposed in the SDK, somehow?
reply
Ah, I love these series. Cheers for having two great speakers for this conversation. On benchmarking, I'd hope for tight integration with OS AI SDKs, as those seem like the universal language that any useful AI hardware should be optimized for its background removal, noise cancelling, etc. I hope UL (3DMark.), Primate Labs (Geekbench), etc. will help spit out numbers. MacOS also has a similar AI / neural something SDK, but it will auto-select CPU vs GPU vs NPU for running the task, but I imagine that's also exposed in the SDK, somehow?
reply
DoublesAdvocate
The only benchmarks that people are going to understand for client/consumer devices like these are just their speed in mainstream apps (SDF, CGPT) like you said.
It'll be the same things we see for GPUs with games or CPUs with video rendering. I don't think most people care about gimmicks like studio quality background blur for zoom calls, I don't think most people would even notice the difference between a dedicated AI replacement vs whatever cheap background replacement zoom already has built in.
reply
The only benchmarks that people are going to understand for client/consumer devices like these are just their speed in mainstream apps (SDF, CGPT) like you said.
It'll be the same things we see for GPUs with games or CPUs with video rendering. I don't think most people care about gimmicks like studio quality background blur for zoom calls, I don't think most people would even notice the difference between a dedicated AI replacement vs whatever cheap background replacement zoom already has built in.
reply
Skele-Ton
The big sadness is that we're using AI so skin-deep for aesthetic endeavors. We COULD be using AI to effectively turn Windows into a Star Trek computer that can manage, evaluate, run tests, diagnose, and solve problems. But I think we all know the Windows Copilot won't be this. It will basically be Bing GPT with the ability to open your Settings window.
I don't think we'll ever have an AI where I can say Computer, record the GPU usage for the next 30 seconds and then compile it to a graph or something.
reply
The big sadness is that we're using AI so skin-deep for aesthetic endeavors. We COULD be using AI to effectively turn Windows into a Star Trek computer that can manage, evaluate, run tests, diagnose, and solve problems. But I think we all know the Windows Copilot won't be this. It will basically be Bing GPT with the ability to open your Settings window.
I don't think we'll ever have an AI where I can say Computer, record the GPU usage for the next 30 seconds and then compile it to a graph or something.
reply
Scott
Asked for comments so here goes, I d like to see a separate AI mini socket so I can pick and choose who s AI chip to use. (Maybe even ARM based AI chip.) So yeah, separate universal socket (asking too much?) on the motherboard and design it so you can put in an AI chip or just leave it empty and the mainboard still functions. This would also allow for AI chip to be replaced several times before e-wasting your PC and starting over. I know BIG ASK!
reply
Asked for comments so here goes, I d like to see a separate AI mini socket so I can pick and choose who s AI chip to use. (Maybe even ARM based AI chip.) So yeah, separate universal socket (asking too much?) on the motherboard and design it so you can put in an AI chip or just leave it empty and the mainboard still functions. This would also allow for AI chip to be replaced several times before e-wasting your PC and starting over. I know BIG ASK!
reply
Darel
If AI on CPU can improve Branch prediction accuracy, then on chip cache can be reduced to reduce power requirements while improving utilization of next level cache, which can further be reduced because cache miss losses have been managed by the AI. The desktop has carte blanche on the availability of power management schemes and AI management of thread assignment will improve the reaction time of CPUs further.
reply
If AI on CPU can improve Branch prediction accuracy, then on chip cache can be reduced to reduce power requirements while improving utilization of next level cache, which can further be reduced because cache miss losses have been managed by the AI. The desktop has carte blanche on the availability of power management schemes and AI management of thread assignment will improve the reaction time of CPUs further.
reply
Mike
What does AI do besides face and finger recognition, in a mobile device, nothing. All these AI conversations with no address of the applications, the utility values so said AI provides the user. Maybe some speech recognition synthesis/correction . . . I don't get the hype. Right, what exact use cases Dr. Ian . . . mb
reply
What does AI do besides face and finger recognition, in a mobile device, nothing. All these AI conversations with no address of the applications, the utility values so said AI provides the user. Maybe some speech recognition synthesis/correction . . . I don't get the hype. Right, what exact use cases Dr. Ian . . . mb
reply
Ojas
Honestly branding these things as AI cores is unhelpful, because it makes the use cases sound very narrow. Maybe i have no AI workload that's run on my laptop, why should i care?
It's kinda like how more than 8 cores 16 theeads is unnecessary for most people at the moment, or highly situational
reply
Honestly branding these things as AI cores is unhelpful, because it makes the use cases sound very narrow. Maybe i have no AI workload that's run on my laptop, why should i care?
It's kinda like how more than 8 cores 16 theeads is unnecessary for most people at the moment, or highly situational
reply
zimelo
To the naysayers: Both Android and iPhone chips already have had AI hardware for years and it's been doing great things for them in terms of image and video processing, performance, power efficiency, security etc.
reply
To the naysayers: Both Android and iPhone chips already have had AI hardware for years and it's been doing great things for them in terms of image and video processing, performance, power efficiency, security etc.
reply
Leon
This PC world and Tech Tech potato colabo is great. Currently according to my view AI needs standardization or software and hardware manufacturers will be burning R&D on making propriety AI solutions.
reply
This PC world and Tech Tech potato colabo is great. Currently according to my view AI needs standardization or software and hardware manufacturers will be burning R&D on making propriety AI solutions.
reply
Brad
i dont see 'AI' helping on consumer sales especially above core count because we KNOW cores and clocks make a difference, but we have no clue what 'AI' hardware might help if it does anything at all
reply
i dont see 'AI' helping on consumer sales especially above core count because we KNOW cores and clocks make a difference, but we have no clue what 'AI' hardware might help if it does anything at all
reply
Abdul
Hey ENTED WID GORDEUNGUN GUN. Please update when will Ryzen 7940HS AI Graphics Optimized Driver will be released and which of AMD Ryzen Laptop processor has these UNGAN GUN AI capabilities.
reply
Hey ENTED WID GORDEUNGUN GUN. Please update when will Ryzen 7940HS AI Graphics Optimized Driver will be released and which of AMD Ryzen Laptop processor has these UNGAN GUN AI capabilities.
reply
pedro
Monitoring software for each AI engine would be great to assess if being efficiently used (or simply being used). Excellent documentation for devs always pay-off down the line.
reply
Monitoring software for each AI engine would be great to assess if being efficiently used (or simply being used). Excellent documentation for devs always pay-off down the line.
reply
Christopher
I would like an AI agent that can highlight the most useful insightful comments, especially on videos where there are 30,000 comments and 29,500 are Cool vid bro!
reply
I would like an AI agent that can highlight the most useful insightful comments, especially on videos where there are 30,000 comments and 29,500 are Cool vid bro!
reply
Aquatic
A lot of companies could accidentally dump their code and customer data into public AI databases, adding yet another security headache.
reply
A lot of companies could accidentally dump their code and customer data into public AI databases, adding yet another security headache.
reply
Add a review, comment
Other channel videos
1:33:12
1:50:27
