Next GPU releases: 2022 edition
- Thread starter pauli
- Start date
Well, the mass market spectacles will definitely be showing whatever they want you to see.
China can feel slighted all they want, that doesn't mean they have any alternatives. China has always known what they need to do to lift the sanctions, they just don't want to. It will probably take a major economic collapse that the west has to bail them out from to finally break the protectionism, just like happened in Japan and Korea.
They will definitely release a 5090 for consumers, every single hardware and game reviewer and content creator uses 4090s for 90% of their content and Nvidia won't want to give that up. They can always order more dies if demand is high. AI demand will fall off at some point but it's not going to crash like crypto mining did, there are too many varied and already released products that use it. If anything the massive demand for AI GPUs will even out the supply for gaming GPUs because there will be so many dies getting made they can easily shift over however many they need to meet the gaming demand. That's how things work when manufacturing volume scales up. They will probably even start looking for other markets they can stick high end GPU dies.
No one said they wouldn't, just that it was being delayed slightly so the 5080 can launch alongside it.
Ehhh... that depends on how the 5090's being put together. If it's relying on CoWoS to glue two dies together, then it's not about the dies, it's about the advanced packaging. Packaging is the bottleneck for anything that is multi-chip.
Capital Economics is saying the AI bubble will burst in 2026, deflating the S&P.
Existing "AI" isn't intelligent, it's just more advanced text prediction. It serves no real purpose other than to delay actual tasks.
It's not sentient, but if you can't tell the difference for a specific task, is there actually a practical difference? If you don't think current AI has a massive amount of real world uses then you aren't paying any attention at all. Stock photos for random "stuff" are basically dead, it will be all AI from here on. Commodity voice acting is similarly dead, why pay someone when the computer can say it and takes direction better? Tools like Generative Fill in Photoshop are already critical to people's workflows and basically impossible to detect so people don't realize it's already everywhere. Go look some of the videos of ChatGPT-4o that released yesterday and tell me that's not something dozens of industries are going to jump on immediately. AI isn't great at coding whole applications, but it is great at giving answers to specific programming tasks tailored to your existing code, it's already in a lot of programmer's workflows.
AI is not a bubble and it's not going to crash. There may be an investment bubble at some point but honestly it seems like we have barely even started up that mountain and I expect investment in AI will still increase by an order of magnitude or two before it starts to taper off. Even when that happens, companies will still need buildings full of Nvidia hardware to run all the services they already successfully have, and they will continually replace it with the new faster more energy efficient versions.
Analysts say otherwise. For society's sake, I hope they're correct.
The AI hype is a bubble and it will burst. A few of the trained machine learning models will cross some threshold of reliable usefulness, so they will stay around. It may or may not be worth it to refine them, but without needing the mountain of AI hardware that was required for their initial training. Most of the trained models will turn out to be mostly useless due to "hallucination" or other unwanted outputs that are impossible to control. We trained them with everything we found on the internet, so the old adage will hold: "garbage in, garbage out".
Some of the actually disruptive uses, like protein folding, optimized chip layout, and many others that I am unaware of, will live on. But those are more narrowly targeted, and are often constrained by correctness criteria that are computed/validated with traditional algorithms, so they will not require a mountain of AI hardware either. We trained these with very carefully curated example data from the best human experts. "Quality in, quality out", if you will.
Finally, but this is IMHO decades to centuries out, a lot of AI inference accelerators will be replaced by less generic hardware, because human minds will meticulously analyze the most valuable trained models in order to identify and understand which correlations exactly were found during training, and are enabling the model to do its job. Then we'll design special function hardware to do that job much more efficiently. There will be some models that will defy analysis ... those might lead to more fundamentally groundbreaking insights eventually.
Some of the actually disruptive uses, like protein folding, optimized chip layout, and many others that I am unaware of, will live on. But those are more narrowly targeted, and are often constrained by correctness criteria that are computed/validated with traditional algorithms, so they will not require a mountain of AI hardware either. We trained these with very carefully curated example data from the best human experts. "Quality in, quality out", if you will.
Finally, but this is IMHO decades to centuries out, a lot of AI inference accelerators will be replaced by less generic hardware, because human minds will meticulously analyze the most valuable trained models in order to identify and understand which correlations exactly were found during training, and are enabling the model to do its job. Then we'll design special function hardware to do that job much more efficiently. There will be some models that will defy analysis ... those might lead to more fundamentally groundbreaking insights eventually.
AI is a bubble in the same sense that dotcom was a bubble in 2000; it's overhyped and money is pouring in a veritable torrent not justified by any remotely reasonable return on investment but the underlying product is very much real. Right now we're seeing the equivalent of Pets.com paying to ship cat food to everybody's door before logistics supported it-- but eventually it very much did and lots of people buy cat food from Amazon every month with Subscribe and Save. Amazon.com opened its doors in 1994 but wasn't profitable until 2003.
General purpose AI, these incomprehensibly huge models containing everything under the sun like GPT-4o, have less utility and certainly avenues for monetization unless you consider perhaps eventually sci-fi AGI, but directed generative AI, smaller models trained and fine-tuned on specific needs, have extreme utility in pretty much every human endeavor, and won't need human-led sanity checking forever. They'll run locally or on the edge and they'll be cheap. Gen AI will be broadly useful. It already is, really, but while today the hallucinations are funny and make everybody shake their heads chuckling-- they'll get better.
Lemme give you an example. Medical scribes. This is a job paying around $50k/year where you sit in the doctor's office and basically document patient visits in great detail. Gen AI can do that job, once those hallucinations are ironed-out. It's just watching a couple cameras "Dr. Peterson checked Mr. Samuelson for gout by palpating his left big toe", speech to text with attribution, and note taking.
General purpose AI, these incomprehensibly huge models containing everything under the sun like GPT-4o, have less utility and certainly avenues for monetization unless you consider perhaps eventually sci-fi AGI, but directed generative AI, smaller models trained and fine-tuned on specific needs, have extreme utility in pretty much every human endeavor, and won't need human-led sanity checking forever. They'll run locally or on the edge and they'll be cheap. Gen AI will be broadly useful. It already is, really, but while today the hallucinations are funny and make everybody shake their heads chuckling-- they'll get better.
Lemme give you an example. Medical scribes. This is a job paying around $50k/year where you sit in the doctor's office and basically document patient visits in great detail. Gen AI can do that job, once those hallucinations are ironed-out. It's just watching a couple cameras "Dr. Peterson checked Mr. Samuelson for gout by palpating his left big toe", speech to text with attribution, and note taking.
Last edited:
Except NVidia is the Amazon of AI, not the Pets.com of AI, or even better because they are immensely profitable already.
Nvidia is more infrastructure, they're like Intel supplying all the CPUs Amazon's servers run on to allow for its innovation. The analogy doesn't really transfer to Nvidia as they aren't themselves a player in gen AI, rather an enabler.
Anyway, in the analogy Nvidia would be Pets.com. Investors are acting irrationally enthusiastically based on wide-eyed prognostications. Facebook would be Pets.com too, buying 350k H100s at $30k apiece with no clear route to monetization-- although FB does say it helped their ad targeting after the Applocalypse, and that may even be true. OpenAI is definitely Pets.com.
There's tremendous amounts of money, crazy money, being thrown around. Huge enthusiasm for those crazy prognostications. And it's partially justified; AI will be transformative across many aspects of life. The problem is it won't come from those giant models, that transformation will be enabled by small, custom-trained and fine-tuned, fast, heavily-quantized models running on your device or at most on the edge.
Anyway, in the analogy Nvidia would be Pets.com. Investors are acting irrationally enthusiastically based on wide-eyed prognostications. Facebook would be Pets.com too, buying 350k H100s at $30k apiece with no clear route to monetization-- although FB does say it helped their ad targeting after the Applocalypse, and that may even be true. OpenAI is definitely Pets.com.
There's tremendous amounts of money, crazy money, being thrown around. Huge enthusiasm for those crazy prognostications. And it's partially justified; AI will be transformative across many aspects of life. The problem is it won't come from those giant models, that transformation will be enabled by small, custom-trained and fine-tuned, fast, heavily-quantized models running on your device or at most on the edge.
Last edited:
They are both. They train their own models for gaming, and for self driving cars, and probably Silicon design and who knows what else.
None of that is generative AI, though. Very different thing. Nvidia doesn't heavily develop in gen AI itself; just stuff like ChatRTX for home users to mess around with and some non-competitive image generation. At least as far as has been publicly disclosed, of course. They prefer to assist developers by providing proprietary (but very, very good) libraries like TensorRT so their hardware is always the best choice for whatever you're looking to do. Same philosophy as in gaming, really.
Last edited:
Generative AI is not the only AI, and arguably not the most useful AI.
Calling NVidia the Pets.com of AI is just absurd. Pets.com is punch line for failure, when NVidia is the exemplar of success.
It's not just over-exuberant share holders over valuing NVidia, NVidia is generating massive profits from selling real products.
The derail began from Hobold saying AI was a bubble and would only be used in very narrowly targeted segments. But sure, Nvidia works on all kinds of AI, just predominantly not generative. Point is potential AI uses are expansive, awesomely so, not narrow.
Nvidia is selling real products, yeah, analogy falls apart there. UPS wasn't part of the dotcom bubble, they just enabled part of it.
Nvidia is selling real products, yeah, analogy falls apart there. UPS wasn't part of the dotcom bubble, they just enabled part of it.
Ah yes, the "it's hobold's fault" meme
. The labelling of AI as "bubble" happened before, by a forum user who has moderation privileges. My take is that there is currently an AI bubble happening, not that AI itself is a bubble. I mentioned two examples where machine learning has been disruptively successful, and neither of those are large language models.I also dipped into science-fiction when I wrote that at some point down the road, humans would start to dig deeper into whichever successful trained AI models will remain after the shake-out. It is (to me, at least) a very plausible scenario that people would spend effort to learn and understand what the AI has learned earlier by brute force.
It's a bubble in that it's overvalued and will "burst". But I expect that it will "burst" into what was previously considered to be a normal value range, not sink beneath it. There's definite value in LLM, but the current "it will cure cancer in 3 months" valuation can't hold long.
We all agree, then. Certainly overvalued, but not by any means without value.
Anyway back to GPU news, Nvidia+ARM handhelds coming. Surely everybody interested in a gaming handheld has one by now? What they need is better performance, and DLSS will help. And every single new one absolutely must have a VRR display, anything less is unacceptable.
I wonder what software ecosystem they have in mind. Nintendo has one, and has demonstrated with the Switch that they can bootstrap from scratch. Steam Deck and Windows portables have a gamut of Windows games. What'll Nvidia do; x86 emulation? Cooperation with Valve?
There are mature Linux and Android ecosystems for cheap emulation devices, and if Nvidia started selling chips capable of later gen 3d console emulation they would own that market. Mobile gaming is big enough that I'm sure a bunch of people would prefer a steamdeck like device instead of a phone. Proton also works on ARM, so theoretically an ARM device running SteamOS is a possibility. That would require a bunch of work on Valve or the community's part, but most of the hard work is done already.
It's a legal grey area, emulation means obtaining ROMs and firmwares from unauthorized sources and less than 0.1% of the users actually have the knowledge and hardware needed to dump anything themselves. I doubt any big business is crazy enough to market a chip to that user base.
SteamOS is sponsored by their store just like Switch with their software; sure you can run third party software on it but it has a legit main source of games and software. Arm, well, surely nobody needs more FPS in Candy Crush.
My point is (I wasn't very clear about it) that Nvidia must have a business model in mind. They would not spend any resources if they didn't believe in a return. So where will the games come from?
They just need to market it to legal emulation devices and the community usually takes it from there. When you round up the market as a whole N64 and below emulation devices are a big market, both legal and the dodgy chinese ones, and people are willing to pay a premium for solid devices. If an Nvidia ARM chip means those types of devices can push into N64/GC/PS1/PS2/XBox emulation, that's a lot of potential money on the table.
That was my point, Proton runs on ARM now, so theoretically so could SteamOS and therefore most Steam games. ARM chips have decent performance these days, but the GPUs still suck pretty hard compared to Nvidia and even AMD GPUs, so if NVidia enters the market they could probably sell a lot of chips/devices if it's cheaper and more power efficient than x86 CPUs.
Also, nobody needs more FPS in Candy Crush, but these days the big games on mobile are CoD, Fortnite, Minecraft, Genshin, etc, where more FPS are actually a thing.
And with 'x86 emulation, including hardware assist like Apple built into their flavour of ARM CPU? In the latter case, Nvidia might get polite mails from Apple's lawyers. Without the hardware assist, the Steam Deck's Zen 2 cores will compete much better than most people would expect.
Well, that's the question, isn't it? From what I've seen recently, Windows x86 on ARM is actually pretty performant, similar to Rosetta on Mx, so assumedly Qualcomm does have similar hardware assisting the translation. Everybody just missed it because the actual hardware MS released alongside that version of Windows was laughable compared to the M1. So I'd expect Nvidia, AMD, Samsung, Broadcom, hell, TI, anyone else that wants to get in on this to do the same thing.
ARM isn't like x86, anyone who wants to pay for the IP can start selling chips. If ARM on Windows works and makes sense, offering similar advantages to Apple's migration, it'll be freakin' open-season. Everybody under the sun competing, potentially real choices amongst 5-6 different CPU manufacturers for your next computer. How cool would that be?
ARM isn't like x86, anyone who wants to pay for the IP can start selling chips. If ARM on Windows works and makes sense, offering similar advantages to Apple's migration, it'll be freakin' open-season. Everybody under the sun competing, potentially real choices amongst 5-6 different CPU manufacturers for your next computer. How cool would that be?
Very cool... until half your applications don't work right.
My boss tries to run windows stuff on a Mac M2 and has emulation issues all the freaking time. Super annoying. I'm sure there is some magic combination of updates and new releases that will fix it all but it means re-buying a lot of software (financial management packages, math stuff, some office software and some very specialized flight software or something he uses). Basically you would have to spend $10k+ to get things working to get a few extra hours of battery life vs. running it all on his Intel Macbook Pro.
I'm sure it will all get sorted eventually to some degree but it's going to be a bumpy road for some people who simply see 'upgrade and get double the battery life!' and pull the trigger.
My boss tries to run windows stuff on a Mac M2 and has emulation issues all the freaking time. Super annoying. I'm sure there is some magic combination of updates and new releases that will fix it all but it means re-buying a lot of software (financial management packages, math stuff, some office software and some very specialized flight software or something he uses). Basically you would have to spend $10k+ to get things working to get a few extra hours of battery life vs. running it all on his Intel Macbook Pro.
I'm sure it will all get sorted eventually to some degree but it's going to be a bumpy road for some people who simply see 'upgrade and get double the battery life!' and pull the trigger.
Not sure I buy Nvidia trying to make a play for the x86 handheld market with emulation. There seems to be some market for Android gaming and we've not had a true replacement for the Shield tablet. Maybe they want to take another shot at it.
My wild speculation is the MediaTek partnership is for CV "AI" stuff, for integration into TV's, smart toasters, etc. and the internal CPU is directly or a minor respin of what they're making for Switch 2, letting Nintendo have it for cheap in exchange for no exclusivity. Maybe a port of one of the smaller Orin's to TSMC, I suspect Thor isn't near/ready enough.
edit: mildly related, I'm curious to see where AMD's handheld efforts go. Recent financials showed a dip in gaming revenue, partly attributed to downturn in custom silicon. There is also fair amount of scuttlebutt around supply issues of their SoCs, providing further hints that they're shifting allocations to DC. Which makes sense, Intel isn't nearly as competitive in that segment and it's where the margins currently are.
My wild speculation is the MediaTek partnership is for CV "AI" stuff, for integration into TV's, smart toasters, etc. and the internal CPU is directly or a minor respin of what they're making for Switch 2, letting Nintendo have it for cheap in exchange for no exclusivity. Maybe a port of one of the smaller Orin's to TSMC, I suspect Thor isn't near/ready enough.
edit: mildly related, I'm curious to see where AMD's handheld efforts go. Recent financials showed a dip in gaming revenue, partly attributed to downturn in custom silicon. There is also fair amount of scuttlebutt around supply issues of their SoCs, providing further hints that they're shifting allocations to DC. Which makes sense, Intel isn't nearly as competitive in that segment and it's where the margins currently are.
Last edited:
I'll be happy if they just release a new Shield TV device to upgrade my Shield TV Pro. It still works great but it has always needed just a bit more RAM, faster flash and (now days) better encoding/decoding support for AV1 and all that in hardware and HDMI2.1 or whatever they can do by the time it comes out.
Alchemist was a pile of junk:
View: https://www.youtube.com/watch?v=MGD41i5QCyk&t=87s
GN goes over the highlights of updates to Intel's GPU architecture in Xe2 (Battlemage). One of them is moving from SIMD8 to SIMD16, and they state that this will result in, "More and more often, you'll see games running right out of the box."
This wasn't rocket science, Intel. AMD was on SIMD16, and the engineers they poached knew this...
Unless you really need AV1 encoding and can't afford a 4060, just stay far, far away from Alchemist. I can see them dropping support for it like a hot potato once Battlemage is out.
View: https://www.youtube.com/watch?v=MGD41i5QCyk&t=87s
GN goes over the highlights of updates to Intel's GPU architecture in Xe2 (Battlemage). One of them is moving from SIMD8 to SIMD16, and they state that this will result in, "More and more often, you'll see games running right out of the box."
This wasn't rocket science, Intel. AMD was on SIMD16, and the engineers they poached knew this...
Unless you really need AV1 encoding and can't afford a 4060, just stay far, far away from Alchemist. I can see them dropping support for it like a hot potato once Battlemage is out.
Xe has to be an incredible hard sell to Intel's board - a VERY expensive product that's just a money pit and will continue to be for likely several generations. I guess these days so many people are frothing for anything you can tack the work 'AI' on, so maybe it can be sold as a steppingstone to global AI domination or something.
Disappointed that we didn't see any consumer GPU announcements from Computex - at least some people were expecting Nvidia to announce Blackwell-based RTX cards, but Nvidia basically just did a victory lap talking about previous announced products.
Disappointed that we didn't see any consumer GPU announcements from Computex - at least some people were expecting Nvidia to announce Blackwell-based RTX cards, but Nvidia basically just did a victory lap talking about previous announced products.
I find it immensely amusing that the stated reason for Intel getting back into GPUs was that they needed to fill their plants - and then they fab it at TSMC.
Nobody has anything ready to ship. AMD didn’t paperlaunch because they wanted people to pay attention to Zen 5. Intel didn’t paperlaunch because they wanted people to pay attention to Lunar Lake. Nvidia didn’t paperlaunch because they didn’t want to reduce demand when the next gen is still 6 months out and they just did a facelift.
And who expected new Nvidia GPUs now? I honestly haven’t seen anyone seriously suggest that.
The funny thing is that from a purely technical point of view, shader programs should neither know nor care about the SIMD with of the specific hardware they are running on. If this is really such a big compatibility issue, then it means a surprisingly large fraction of real world shader programs are technically incorrect. Of course the programmers never knew, because there were no test cases exposing such bugs.
And then indeed the pragmatically best way to deal with the issue is to become "bug compatible" to whoever is setting the de facto standard (probably Nvidia; AMD would have converged on SIMD16 earlier for the same reason).
Shaders are a problem kind of like javascript, they are doing a job they were never designed to do at a scale nobody imagined when they were writing the specs. The majority of shaders in a game are going to be cobbled together automatically from UI buttons and sliders and recursively stacked and bolted together to get the effect the designer wanted. They are way too dynamic, which is what causes all the shader compilation issues, since you don't actually know what combinations of shaders need to be run on each object until runtime because everything is so conditional. To combat the performance mess that flexibility creates, the game engines and drivers try to do just in time optimization to make it all faster. All that means there's a whole lot of room for bugs especially when there are only two platforms to build to. You don't need code that works to spec, you need code that runs as fast as possible on Nvidia and AMD cards.
It stopped being funny when people who knew how de facto standards in the industry worked decided to ignore those standards and fuck things up.
Intel coming out and saying SIMD16 will help with games working out of the box is as much of an admission that the previous team fucked up as we're going to get.
After getting two black eyes and a bloody nose in their first attempt at a dGPU in ~20 years, they're hiding their second generation GPU core (Battlemage) behind the CPU at the outset. That's not going to help them gain dGPU experience, but maybe it gives the dGPU team an extra 3-6 months to iron out the dGPU-specific drivers for day 1.
That’s one way to look at it, the other is it’s miraculous they released something even vaguely usable on the first try. It’s an unfathomably challenging task.
Then they dedicated massive resources to catching up to twenty years of driver development from AMD and Nvidia, optimizing for hundreds of games released over the decades each with its own idiosyncrasies, another challenge the likes of which has rarely been seen anywhere, anywhen.
Sure, it failed in the market. It wasn’t enough. Even if it was, the chips are too dang big to make any profit. Yet they persevered and are aggressively moving to gen 2. This is the sort of determination required to give consumers a third choice, and I celebrate it.
Alchemist clearly had a metric shitton of low hanging fruit to pluck. I have real hopes for battlemage.
Then they dedicated massive resources to catching up to twenty years of driver development from AMD and Nvidia, optimizing for hundreds of games released over the decades each with its own idiosyncrasies, another challenge the likes of which has rarely been seen anywhere, anywhen.
Sure, it failed in the market. It wasn’t enough. Even if it was, the chips are too dang big to make any profit. Yet they persevered and are aggressively moving to gen 2. This is the sort of determination required to give consumers a third choice, and I celebrate it.
Alchemist clearly had a metric shitton of low hanging fruit to pluck. I have real hopes for battlemage.
NVidia is more akin to the Sun Microsystems of the AI bubble. They're generating massive profits but their profits and stock value is being inflated by the bubble. They may still survive and be profitable after the bubble bursts but their sales and stock value will likely decline considerably.
There is no miracle involved here.
Not their first try. We're on attempt three or four.
No, it's not. Qualcomm, Apple, nVidia, AMD, and many others pull the task off, every year.
No, they didn't. They used their IGP driver experience, and the IGP has been playing games for over a decade.
You're off by two orders of magnitude.
There are 11,010 titles in the PC Gaming Wiki from DirectX 2 to DirectDraw.
The PC Gaming Wiki is publicly available. They knew what they were walking into, and if they didn't the person in charge should have been fired. Oh wait, he was.
Stop giving Intel the benefit of the doubt. It's a multi-billion dollar company that has had aspirations of getting back into dGPU since 2009, and has been poaching industry talent the whole time. They've been working on this for 15 years, and continue to fuck it up due to their arrogance.
Last edited:
I think it's a mistake to call AI a bubble.
As far as NVidia being like Sun. Lots of companies get big, and stay big.
In choosing Sun as your example, I think you need to justify why you think NVidia would so badly as to be compared with Sun.
This is Intel's first discrete GPU, unless you count the lightspeed i740 or whatever twenty-five years ago. It's a new architecture.
Apple does not even consider compatibility with PC games. They don't even care about backwards compatibility with Mac games.
Qualcomm has a similar challenge with drivers now, we'll see how well they do.
Intel did optimize for hundreds of old games. That number doesn't encompass every game ever released, they're prioritizing.
Apple does not even consider compatibility with PC games. They don't even care about backwards compatibility with Mac games.
Qualcomm has a similar challenge with drivers now, we'll see how well they do.
Intel did optimize for hundreds of old games. That number doesn't encompass every game ever released, they're prioritizing.