NVIDIA Is On a Different Planet
video description
Date: 2024-03-22
Related videos
Comments and reviews: 20
radugrigoras
I would say that Ryzen is not even remotely close to this. Their interconnect on the chip pcb is a shrunken down version of what multi socket motherboards have been doing for decades. A much closer comparison is Apple with the M2 Ultra. AMD is generations behind same goes for intel. Apple and NVIDIA are just so far ahead and advancing so fast, mainly due to unlimited funding that it’s going to be very difficult for anyone to catch up, if another company has some new innovation Apple and NVIDIA can afford to buy them cash with no questions just for their IP. Cost savings are for shareholders not for consumers, at the end of the day they don’t even need to make gaming gpus anymore to make profit, probably it’s actually the opposite when you factor in the fact that chips are a finite resource, if you can divert from gaming gpus to AI accelerators your profits would actually go up. So I think the 5000 series might not even really be that big of an upgrade as people are hoping for, it will just be the scraps that are left over and can’t be packaged for AI due to hardware defects. Which those scraps would still destroy what AMD or intel will come out with. At this point if they release the 4000 series as 5000 and just bump everything down one SKU like 4090 becomes 5080 and make the 5090 with some Blackwell scrap they would still dominate. They don’t need chiplets and HBM3 to win in the gaming race.
reply
I would say that Ryzen is not even remotely close to this. Their interconnect on the chip pcb is a shrunken down version of what multi socket motherboards have been doing for decades. A much closer comparison is Apple with the M2 Ultra. AMD is generations behind same goes for intel. Apple and NVIDIA are just so far ahead and advancing so fast, mainly due to unlimited funding that it’s going to be very difficult for anyone to catch up, if another company has some new innovation Apple and NVIDIA can afford to buy them cash with no questions just for their IP. Cost savings are for shareholders not for consumers, at the end of the day they don’t even need to make gaming gpus anymore to make profit, probably it’s actually the opposite when you factor in the fact that chips are a finite resource, if you can divert from gaming gpus to AI accelerators your profits would actually go up. So I think the 5000 series might not even really be that big of an upgrade as people are hoping for, it will just be the scraps that are left over and can’t be packaged for AI due to hardware defects. Which those scraps would still destroy what AMD or intel will come out with. At this point if they release the 4000 series as 5000 and just bump everything down one SKU like 4090 becomes 5080 and make the 5090 with some Blackwell scrap they would still dominate. They don’t need chiplets and HBM3 to win in the gaming race.
reply
erictayet
TSMC has multiple ways to connect chiplet to chiplet. Just went to TSMC website to look at their 3DFabric technology. There are many versions of the tech, some looks to be cheap and already used by AMD's Infinity Fabric. These are the tech using PCB substrate. Some looks to be way more expensive using interposer to connect chiplets together, then to the PCB substrate.
From my memory, AMD Zen 3 & RDNA3 & Apple M3 uses this tech to join chiplets together. This is different from an interposer which stacks wafers on wafers to form an SoC, which becomes said chiplet.
Zen 3 didn't need to share much data across CCX so performance was excellent even though it's using the cheaper process.
I speculate that RDNA3 designers seem to have cheap out on the tech offered by TSMC and paid a heavy price on a higher-than-expected latency which caused RDNA3 to miss performance goals.
Apple M3 Pro/Max seem to be using the more expensive process and obviously Blackwell is as well. So don't expect RTX 5000 to be cheap. And Nvidia can charge whatever they want for their GB200 solutions.
reply
TSMC has multiple ways to connect chiplet to chiplet. Just went to TSMC website to look at their 3DFabric technology. There are many versions of the tech, some looks to be cheap and already used by AMD's Infinity Fabric. These are the tech using PCB substrate. Some looks to be way more expensive using interposer to connect chiplets together, then to the PCB substrate.
From my memory, AMD Zen 3 & RDNA3 & Apple M3 uses this tech to join chiplets together. This is different from an interposer which stacks wafers on wafers to form an SoC, which becomes said chiplet.
Zen 3 didn't need to share much data across CCX so performance was excellent even though it's using the cheaper process.
I speculate that RDNA3 designers seem to have cheap out on the tech offered by TSMC and paid a heavy price on a higher-than-expected latency which caused RDNA3 to miss performance goals.
Apple M3 Pro/Max seem to be using the more expensive process and obviously Blackwell is as well. So don't expect RTX 5000 to be cheap. And Nvidia can charge whatever they want for their GB200 solutions.
reply
aBoogivogi
I don't see the real benefit of the Nvidia design. For most applications were you want high end AI performance you are dealing with multiple cores from the programming side of things because a single chip with our current tech is slow, regardless how fast Jensen might make it. If this interconnect method scaled infinitely it would be another story, but for what he is selling this news is barely worth mentioning as the customer is already dealing with multiple GPU hosts or clusters already. The one exception may be with automakers were space is very limited, but then again as many cars as they push they are still a minor consumer of these parts in the grand scheme of things.
reply
I don't see the real benefit of the Nvidia design. For most applications were you want high end AI performance you are dealing with multiple cores from the programming side of things because a single chip with our current tech is slow, regardless how fast Jensen might make it. If this interconnect method scaled infinitely it would be another story, but for what he is selling this news is barely worth mentioning as the customer is already dealing with multiple GPU hosts or clusters already. The one exception may be with automakers were space is very limited, but then again as many cars as they push they are still a minor consumer of these parts in the grand scheme of things.
reply
Hybris51129
This goes to show just how embarrassing it has been to keep defending AMD while blaming Nvidia for being a Evil Corporation. When the only other company in the market isn't even bothering to compete with you, why not see how far you can go AI is the hot new tech and money maker and AMD isn't even on the radar and unlike gamers the companies in the market for AI GPU's aren't going to back the lame horse for some nebulous moral reasoning.
This is why I have been pissed at AMD's GPU division since the failure of Vega, they as a company have given up and only now are the shills waking up to the realization that not everyone is buying their BS.
reply
This goes to show just how embarrassing it has been to keep defending AMD while blaming Nvidia for being a Evil Corporation. When the only other company in the market isn't even bothering to compete with you, why not see how far you can go AI is the hot new tech and money maker and AMD isn't even on the radar and unlike gamers the companies in the market for AI GPU's aren't going to back the lame horse for some nebulous moral reasoning.
This is why I have been pissed at AMD's GPU division since the failure of Vega, they as a company have given up and only now are the shills waking up to the realization that not everyone is buying their BS.
reply
mindinversions4487
The big question is: Do we NEED Nvidia in the GPU space anymore Ok, I love my big expensive GPU as much as the next guy, but with the performance we're getting out of integrated solutions these days... Intel developing GPUs [and one would assume using the data to improve their iGPU solutions]. Seriously though, between frame gen and technological advancements, how far are we from a steam deck sequel that can play about anything at 1080 60 Yeah, that's the FIRST bar to hurdle, but in a few years
reply
The big question is: Do we NEED Nvidia in the GPU space anymore Ok, I love my big expensive GPU as much as the next guy, but with the performance we're getting out of integrated solutions these days... Intel developing GPUs [and one would assume using the data to improve their iGPU solutions]. Seriously though, between frame gen and technological advancements, how far are we from a steam deck sequel that can play about anything at 1080 60 Yeah, that's the FIRST bar to hurdle, but in a few years
reply
predabot__6778
I'm quite skeptical about Nvidia's claims here... Mainly in how it's supposed to act like one chip - if you remember AMD has actually ALREADY done dual Compute-dies in their CDNA-architectures, but because of limitations in the fabric/implementation, it was only suitable for Server/Compute, and not for Gaming and Rendering. If nVidia uses chiplets in their consumer-cards, I'll eat crow, but until I see it, then I will still see this as merely used in the datacenter.
reply
I'm quite skeptical about Nvidia's claims here... Mainly in how it's supposed to act like one chip - if you remember AMD has actually ALREADY done dual Compute-dies in their CDNA-architectures, but because of limitations in the fabric/implementation, it was only suitable for Server/Compute, and not for Gaming and Rendering. If nVidia uses chiplets in their consumer-cards, I'll eat crow, but until I see it, then I will still see this as merely used in the datacenter.
reply
151mcx
with musk and meta zuck buying every high end nvidia hardware they can get their greedy fat fingers on.... It's easy cake for nvidia. The question is, how long will it go on, and when it dries up, and it probably will. I wonder how easy they can retool the dies to jump right back into gaming GPU's Interesting times. A lot of people aren't just using these GPU's for gaming. Hobby scientist and minors. Will be interesting to see the direction it all runs to.
reply
with musk and meta zuck buying every high end nvidia hardware they can get their greedy fat fingers on.... It's easy cake for nvidia. The question is, how long will it go on, and when it dries up, and it probably will. I wonder how easy they can retool the dies to jump right back into gaming GPU's Interesting times. A lot of people aren't just using these GPU's for gaming. Hobby scientist and minors. Will be interesting to see the direction it all runs to.
reply
LisaSamaritan
I have seen the analogy with selling mining equipment in the old Gold Rush. But that analogy is kind of flawed. Because NVIDIA use the mining equipment to make better mining equipment and to create new type of tools. The ones that sold the equipment, back in the 1800s didn't produce or improve on the equipment. They only sold equipment.
Because NVIDIA have the equipment and tools before any costumer, they can always stay a step ahead.
reply
I have seen the analogy with selling mining equipment in the old Gold Rush. But that analogy is kind of flawed. Because NVIDIA use the mining equipment to make better mining equipment and to create new type of tools. The ones that sold the equipment, back in the 1800s didn't produce or improve on the equipment. They only sold equipment.
Because NVIDIA have the equipment and tools before any costumer, they can always stay a step ahead.
reply
ianc4178
It's pretty well established now that Nvidia doesn't give a stuff about gamers. But the irony here is that if Nvidia continue to short-change gamers, as they did with the 4000 series & continue to hike their GPU prices, it will effectively kill RTX & their other technologies in gaming because gamers will jump ship to more affordable GPUs ... that is if AMD price their GPUs more appropriately & Intel get their act together.
reply
It's pretty well established now that Nvidia doesn't give a stuff about gamers. But the irony here is that if Nvidia continue to short-change gamers, as they did with the 4000 series & continue to hike their GPU prices, it will effectively kill RTX & their other technologies in gaming because gamers will jump ship to more affordable GPUs ... that is if AMD price their GPUs more appropriately & Intel get their act together.
reply
jjnavacool
I'm on the AI wagon using or building things in and out since 2000's, I really want to believe that the current gold rush is here for good, but there's something in my gut that says that this is a bubble and is going to pop, something in me -maybe is denial- believe that companies are going to find quick that this thing is less useful that they want to sell.
reply
I'm on the AI wagon using or building things in and out since 2000's, I really want to believe that the current gold rush is here for good, but there's something in my gut that says that this is a bubble and is going to pop, something in me -maybe is denial- believe that companies are going to find quick that this thing is less useful that they want to sell.
reply
erin7551
During the gold rush there was a lot less money than people thought in finding gold, and a lot more money in selling pickaxes to people who were going to go broke. I like to think this nod to history is built into the host's analogy. There certainly are gains to be had in AI, it's not the cure all/replacement for humans in every job it's hyped up to be though.
reply
During the gold rush there was a lot less money than people thought in finding gold, and a lot more money in selling pickaxes to people who were going to go broke. I like to think this nod to history is built into the host's analogy. There certainly are gains to be had in AI, it's not the cure all/replacement for humans in every job it's hyped up to be though.
reply
martheunen
If my calculations are correct then:
mod mat calculation long ton usage to kg is 2.56026819125kg (with average african male adult bush elephant body mass of 6.05t [using long ton] calculated from info off of wiki).
But according to your information the mat is approx. 4lbs or 1,8kg.
Could you please clarify so as to explain this discrepancy please
reply
If my calculations are correct then:
mod mat calculation long ton usage to kg is 2.56026819125kg (with average african male adult bush elephant body mass of 6.05t [using long ton] calculated from info off of wiki).
But according to your information the mat is approx. 4lbs or 1,8kg.
Could you please clarify so as to explain this discrepancy please
reply
misterthegeoff9767
It sure sounds like everything Nvidia are claiming the AI Goldmine will offer is really similar to what big data evangelists were claiming 15 years ago, at least the ones that tried peddling their snake oil where I was working. Same shit different technology Monetise all the metadata your customers didn't realise you collected
reply
It sure sounds like everything Nvidia are claiming the AI Goldmine will offer is really similar to what big data evangelists were claiming 15 years ago, at least the ones that tried peddling their snake oil where I was working. Same shit different technology Monetise all the metadata your customers didn't realise you collected
reply
Airtight215
Don’t worry, it’s not just technology that corporate media has no clue about. It’s literally everything that they know nothing about. That’s why they all lie about everything and break the county down to a color gang war of red and blue. Only if it was actually bloods and crips they were lying about.
reply
Don’t worry, it’s not just technology that corporate media has no clue about. It’s literally everything that they know nothing about. That’s why they all lie about everything and break the county down to a color gang war of red and blue. Only if it was actually bloods and crips they were lying about.
reply
AgneDei
I'm sorry but I must do it.
Adult African bush elephants weigh on average 3000kg for a female and 6000kg for a male.
So let's assume that the average of those is 4500kg.
0.0004165 of 4500kg is 1.874kg, which is quite a lot for a mat, but at the same time quite close to what it actually may weigh
reply
I'm sorry but I must do it.
Adult African bush elephants weigh on average 3000kg for a female and 6000kg for a male.
So let's assume that the average of those is 4500kg.
0.0004165 of 4500kg is 1.874kg, which is quite a lot for a mat, but at the same time quite close to what it actually may weigh
reply
joaovitorLabres
About Gym, it's a Reinforcement Learning environment software achitecture. The digital twins are learning how to do things in a digital environment. So, yes, the blue robot is learning how to mock other robtos, but probably receiving a huge penalty and will, eventually, learn to not do it!
reply
About Gym, it's a Reinforcement Learning environment software achitecture. The digital twins are learning how to do things in a digital environment. So, yes, the blue robot is learning how to mock other robtos, but probably receiving a huge penalty and will, eventually, learn to not do it!
reply
masterluckyluke
At least the media in your country TRIES. Here noone cares about technical stuff in the news. Well, on the other hand, maybe it's better that way.
And for the AI Terminator part, to quote The Big Bang Theory: Nobody thinks it will happen, that's why it will definitely happen.
reply
At least the media in your country TRIES. Here noone cares about technical stuff in the news. Well, on the other hand, maybe it's better that way.
And for the AI Terminator part, to quote The Big Bang Theory: Nobody thinks it will happen, that's why it will definitely happen.
reply
cavedon.felipe
That's awesome. The mainstream media just found out a new company XD
Well, it was always supposed to happen anyways. Once you enter the server market and start to serve a bigger margin, there's no business reason to keep focused on a such small margin market as gaming.
reply
That's awesome. The mainstream media just found out a new company XD
Well, it was always supposed to happen anyways. Once you enter the server market and start to serve a bigger margin, there's no business reason to keep focused on a such small margin market as gaming.
reply
larrymacdonald4241
Nvidia is a garbage company, I will never buy one of their products because they stole it from 3dfx.... the technology, I remember when they started and what they did, get informed people....the more you people buy their products, the more they get away with it...
reply
Nvidia is a garbage company, I will never buy one of their products because they stole it from 3dfx.... the technology, I remember when they started and what they did, get informed people....the more you people buy their products, the more they get away with it...
reply
EliteRock
I fail to see why multiple dies and chiplets are supposed to make us go ooo, ahh!. It seems to me nothing more than an admission that silicon density/transistor count has maxed out, i.e. that current processor fabrication techniques have reached their practical limits.
reply
I fail to see why multiple dies and chiplets are supposed to make us go ooo, ahh!. It seems to me nothing more than an admission that silicon density/transistor count has maxed out, i.e. that current processor fabrication techniques have reached their practical limits.
reply
Add a review, comment
Other channel videos
