No one will end anyone.

We are in dire need of a diverse GPU market.

Intel, AMD and now Qualcomm will be just fine.

So Nvidia wants to patent order of operations, kinda reminds me of apple trying to patent the rectangle.

Just costs 30 pikachus to do that operation....

It's funny how we now need fixed function accelerators for matrices after 15 years of turning GPUs (fixed function FP accelerators) into programmable devices. Also, we went from 12/20/36bit word computers to 4bit and 8bit micros, to 16bit RISC processors and FP engines, to 32bit and now 64. Only to discover we now need much less precision, FP 32 and fp16, now 8 and 4bit. We could probably go down to ternary for large model nodes, or 2bit.

Sounds cool, but all I want is more VRAM under 10k

5:12-5:44, is there a Reason for the blackout in the video? Interesting video as always

at some point. 3d and ai need to fork into dedicated architectures instead of having a general do it all GPU

Hate them or love their corporation, NVidia really has some brilliant engineers.

Darn - hate to be held hostage to leather jacket man but it is a good way to go - local AI processing instead of relying on large corporates for AI services on ones data.

jesus nvidia does not rest for the competition.

I would bet my life coreteks owns nvidia stock

CEO of Nvidia only Said IA . IA... IA dozen times, only one gaming...SAD TIMES

This time you are wrong! Inference is where AMD and Intel play in a plainfield with NVIDIA. NPUs have much more simple apis so the vertical stack is thin, almost irrelevant and all the applications are going to support NPUs from intel, amd and nvidia very easily and AMD and Intel have already NPUs enabled processors and PCIE cards on the wild

how is that new and how this will end amd or intel? like whats stopping amd from getting their xillinx accel to do the same job?

4:44 oh no, they went from 30 Pikachus to only 1.5 Pikachus 😮 where did all the Pikachus go then??

10:50 Yes, I'd like to use AI to turn my cat photo into a protein.

I think the NPU accelerator is very much a open market. Both Apple and Qualcomm are embedding NPU accelerators into their ARM V9 SOCs. Also Groq has alternative approach to inference which is much more power efficient?

Haha title reminded me of 30 series launch rumor, I really wanted it to happen but all we got was upgrade in prices lol

Accelerator is not the same thing that you said the 40 series was going to have?

28nm to 4nm is only a 3x density increase? But isnt the 4nm 8700G like 2.5x more transistors than the 7nm 5700G? (both around 180mm²) I mean i guess clock speeds makes a difference, but weren't clock speed lower on 28nm?

Super interesting. Will be following developments.

I might actually end Nvidia if they keep up their ridiculous pricing.

Very informative Ty

When did we change times to x? I've been hearing it more and more lately.. gamers nexus said it earlier (technically yesterday) and a few months ago I remember hearing it.. maybe it's been standing out more and more to me but I think it's now a pet peeve. It's ten times as much sounds like the proper way to say it rather than ten x as much.

So it's an NPU? Or did I miss something?

What about 1.58 bit architectures? same accuracy, way lower memory and transister footprint, be interesting to see how that works out.

*1000 ex performance uplift" I'm annoyed by people using ex instead of times, been hearing it more lately I guess.. I know it's been a thing for awhile but maybe I'm getting old?

This is nothing more than a joke! "Google, Intel, Microsoft, Meta, AMD, Hewlett-Packard Enterprise, Cisco and Broadcom have announced the formation of the catchily titled "Ultra Accelerator Link Promoter Group", with the goal of creating a new interconnect standard for AI accelerator chips." People are tired of Nvidia gimmicks and they will shut them out.

Either I'm bad at listening or I just didn't understand but inference or running neural models locally is all the rage about NPUs and TOPS in current SoC's isn't it? Apple with M3/4, AMD Strix with 50 TOPS and Snapdragon Elite X and Windows 12 with Copilot is exactly that use-case, running models locally isn't it? So why not just cram in these NPUs or new type of accelerators into your CPUs or discrete GPUs and call it a day? What's so revolutionary about this new type of accelerators from NV, that the chips that are hitting the market TODAY, don't have? It's my understanding that optimisations happen on all fronts all the time, transistor level, instruction level, compiler level and software level. When I look for open job positions in IT it hits me how many compiler and kernel optimisation roles are opening for drivers, CUDA and ROCm... Don't get me wrong I love your videos but I just don't see the NV surprise, when everyone is releasing ai accelerators today vs NV promising them in maybe 1 year. NV was focused on the server market, while AMD was actually present in both server and client. Also notice, that NV was already using neural accelerators for their ray tracing workloads, which significantly lowered the required budget of rays, that needed to be cast as they could reconstruct the proper signal quite believably with neural networks. We'd need to assume that TOPS/W metric is only understood by NV and that everyone else will sit idle and be blind to it. I doubt that, judging on what is happening right now. Also we assume, that models will keep growing, at least the cost of learning. There are some diminishing returns somewhere, so I expect models to also shrink and be optimised as opposed to only grow in size. As more people/companies start releasing more models they really need to think how to protect the IP, which is the weights in neurons of these networks because transfer learning is a "biatch" for them :) With progress happening so fast, yesterdays models become commodities. As they become commodities they are likely also to become open-sourced. As such you can expect a lot of transfer learning activities happening, that will act as a force which leads to democratization of older still very good models. So this is a head wind for server HW as I can cheaply transfer learn locally... For me local models are mostly important in two areas of my life: coding aid and photography processing. I really follow what fylm.ai does with color extraction and color matching. As NPUs proliferate more and more cloud based features can be run locally.... (for example Lightroom, Photoshop, fylm ai or copilot like models to aid programmers).

Ah so it's more AI fluff that will amount to more hardware in landfills after the AI bubble bursts.

I feel like the next big step to properly implement the coming technologies in AI and acceleration would be to integrate such architectures directly into the motherboard. especially in light of how large NVidias highest end cards are becoming and how much more space efficient MoBos have gotten through the last half decade, and not to mention the power and efficiency of APUs and NPUs that are coming out this year. To physically offloading those calculations onto a dedicated spot on the motherboard could provide an upper hand in computer hardware. This also doesnt seem all too farfetched when you take into account the industry is planning to implement an "NPU standard" across mobile devices and various OS. Also the fact mobo manufacturers are already re configuring things like RAM from DIMM slots to CAM2 on desktops. combine all of this plus the fact that the technologies could potentially be tied closer to the CPU on the north bridge and it feels like a no brainer to work with mobo manufacturing to further push the limitations of computing power

Thanks. Great video

Fetch and Decode does not move data they only process instruction not data. Also instructions are tiny 16~32bit vs. 512~1024bit SIMD vector data

Do you think Tesla will jump into this market at some point? Would the FSD HW4 be competitive? I don't know if this is a market they are interested in, but it seems like they have already done a lot of the work as far as low power usage inference.

I just realized that Pikachu is just a pun for Picojoules.

When did this become a meme page?

For what I'm gathering this accelerator is a type of cash for the GPU which means it won't be a dedicated card and consumer products it will probably be parts of the video card itself.

Would be interesting to hear your take on Qualcom's "PC Reborn" pitch too.

Have any of CoreTek's predictions ever come true just as described?

isn't that "accelerator" the NPU everybody is talking about nowaday? specialized in local low power inference. Nvidia may have the best prototype, but Qualcomm and AMD are already starting to ship CPU with NPU doing 40-50TOPS. all back up by Microsoft within W11. so even if nvidia comes to market in 2025 it may be too late.

the audio is kinda weird

Well done again, I want to be able to add several cards to my pc and cluster them, Might need a three phase plug.

Video error-blacks out

When do you think NVIDIA will release their consumer desktop pc GH200 Grace Hopper Superchip style product? I'd love an Nvidia-ARM all-in-one Linux beast.

I like the Pokemon reference you're making

This is extremely interesting, I hope they will release discrete accelerators for desktop users

5:20 black screen?

Didn't Google use tensor cores before nvidia

excellent and professional video, ty for ur work!

Some wise man said that Jensen wears leather jacket even on the hottest day of summer

All of these hw are pretty equivalent. AMD showcased similar jumps and you'll see pretty similar jumps next year from Blackwell am next gen MI. Question is how much do you want to pay for it.

Can't wait to buy a PhysX, i mean, Nvidia-AI PCIe card :)

18:39 Truth bombs being dropped.

Ill pay to who ever does not have any AI shit.

I'm already so tired of AI this AI that AI underwear AI toothbrush AI AI AI AI

if you still need to send data ta data center to be processed there and then get the result back, how it is different from any cloud based service? this sounds like a normal cloud-based computing absurdly overcomplicated with unnecessary AI shit instead of encryption.

i think CXL and a gpu would solve the inference problem with right software.

The more they push all this A.I. the more likely I am to end up switching off and going back outside.

Let me guess a traversal coprocessor?

AMD also entering tick tock cadence for the MIx00 series. Ex: M1300X (CDNA3+HBM3) with MI325X (CDNA3+HBM3E) after that, MI350X (CDNA4+HBM3E/HBM4)

I think much will depend on how aggressively NVIDIA builds and defends it (leading edge) patent portfolio. First is best. :)

Mostly for Server side I do not think consumers have as much interest in AI as these corpos make it seem

Yeah, given the performance of their last 2 gens, I'm sure AMD is shaking in their boots.

you can also get win 11 for free and activate it with cmd.

Also, love in the local AI space, Apple has accidentally (or strategically) made their Mac Studio a rather economic choice for local inferencing. That and Nvidia's profit margin (nearly twice that of Apple's) making Tim Apple gushing also shows how dominating Nvidia is in the market right now.

It's interesting how nVidia is still locked out of the Desktop & Laptop CPU market - with AMD, Intel and now Qualcomm pushing Co-Pilot PCs & Laptops - I know Qualcomm had an exclusive on Windows ARM CPU development, but that ends this year (?) - so obvious nVidia should be making SOCs for this market

could... nah

it is kinda nuts how much nvidia has gotten out of that cuda core

Do You think that there will be a competitor or alternative to NVidia within the next 18 months ?

I expect nvidia to soon produce Decelrators and have these slowdown any computer immensely unless the owners pays horrible subscription fees ..

No mention of games until 20:09

I think it will be just as Successful as the Traversal Coprocessor 😈

NVidia making $30,000 AI chips that consumers are not going to purchase should not affect AMD/Intel who aren't currently competing in AI. To "end" AMD and Intel, the NVidia chips would have to be under $500 as that is the limit most consumers will spend on a graphic card or CPU. Somehow, I don't see that happening anytime soon.

Great work! Thank you for sharing! If this materializes it'll be very exciting

New? Supercomputers have existed forever.

What does it look like brainiac

I like the part where he talks about Pikachu

From memory, yesterday, Jensen was speaking about 1000 GWh consumption (1 TWh) for GPT4. I found that HUGE. FYI a battery Gigafactory will consume about 1GWh of electricity per year to produce about 40 GWh of Li on cells , enough to produce 50kWh packs for 800 000 small cars each year. Or maybe, i misunderstood him.... maybe a subject for a video .... how crazy energy you need to run such large LLMs on a server.... compared with what consume our little brain.... (BTW, i do not need to solve plenty of matrices full of sin, cos, tan to move my shoulder, arm, hand, finger every time i want to scratch my head 😉)

i think Nvidia/Jensen are banking so much on AI, because they want generate all the possible money/margin, to be used for CPU R&D, so they will have complete solution/stack 🧐

I Wana geek out about this but I'm convinced that all the ai advancement and inference refinement is just snake oil for over 90% of buyers. So I hate it. I'm down for new approaches, but fr can any non-salesman or non fan boi explain how this will benefit any more than 1-3 companies??

Thats a lot of pikachus!!

Green clickbait cope, thanks!

In the last year you were saying that Nvidia was in a dead end and that their stock price should plunge, instead, their price has rise a 200%. Lucky for me at that time I did not listen to you and I bought, taking into account how important AI was going to be making a lot of profit.

But isn't Batman the same as Batman Batman

I don't care about llm ai assistants. If gaussian splatting takes over rendering than ok.

coreteksilikiaaaa

AI acceleration sounds as good as 8ch sound for my pair of ears. The presentation looks more eyecatching than RGB, maybe because I like spreadsheets and informed narrative.(Just telling that I viewed it with will to absorb as much as my brain accepts🤷🏼‍♀️, when AI make it to desktop PC and games I will understand 2x more) You know more how ground breaking Nvidia acceleration could be, but I am sure I will watch it from distance with my slim wallet 😂 GG, pretty news of this topic.. As usual by Core-news-tech 👍 Patents last legally only limited time, right? AMD and all other will develop their acceleration at law bureau soon.

Nvidia has plenty of cash to continue to be aggressive in its goal to disrupt and lead as many markets as possible. It's pretty much confirmed their next goal will be around edge and consumer AI which basically predicted their unsuccessful acquisition of the Arm in the past. It will be very interesting to see how their edge/consumer Arm SoC they r working on will compete with the rest of the players. Thx for the vid and for shedding more light about their future accelerator👍

But who actually needs this to the point of paying for it?

At this point of development nvidia might as well build a new computer concept with the GPU as the central processing unit. The CPU is becoming more and more a marginal gimmick. AI shifts things and intel and amd have to react, if they want to survive. Arm seems to understand and Jim Keller with tenstorrent understood it as well.

So the industry is having YET ANOTHER go at thin client BS. I hope, again, for them to, again, fail miserably.

Becouse of price I will always buy and GPU and CPU

Idk,, I should be as enthusiastic as you. This may be an investment opportunity. Still waiting for NVidia stock to pull back. Technical analysis says it's at it's cycle top. But I'm also a machine learning expert. While inference is important. It's currently the fastest thing compared to training. The problem is phones not laptops. Laptops can more than handle inference. A phone on the other hand struggles. So Samsung with it's focus on an AI chip is more important in this arena. Unless NVidia is going to start making phones, I don't see this, as an implementer, as that impactful. And memory is more important on phones than cpu for this type of work. On a side note I don't even use GPUs for my AI work. I did a comparison and it only gave me a 20% increase in performance and cost twice as much. So at scale I can buy more CPUs than GPUs. And one more CPU is 100% increase in performance compared to gaining the 20% increase at twice the cost. So I don't see the NVidia AI hype. 1x cpu 1x gpu is 120% at twice the cost 1x cpu is 100% at the same cost is 2x so I get 200% increase for the same price as 1 cpu and 1 gpu

If you're going to talk about LLM/AI inference speeds, why aren't you looking at other hardware like Groq's LPUs? 20x~ (average) cheaper than gpt-4o per million tokens, far faster inference. Basically, efficient. Would be nice to see a video from you on other companies like Groq.

Stop hyping

Accelerators have been around in CPU's for decades now, I don't get all the Nvidia fan fair.

Bruh, using accelerators ain't some genius 4D chess move, NV is just the only one for whom it's already a viable business strat, while AMD doesn't have the scale, and Intel, well, is doing Intel things.

With all due respect Coretex, first it was AMD going to destroy nVidia with chiplet design, then Intel were going to destroy AMD with new CPUs and now nVidia will destroy both! 🤔 Can you please make up your mind? 🙂 PLEASE! 🙏