Tuesday, September 20th 2022
NVIDIA Project Beyond GTC Keynote Address: Expect the Expected (RTX 4090)
NVIDIA just kicked off the GTC Autumn 2022 Keynote address that culminates in Project Beyond, the company's launch vehicle for its next-generation GeForce RTX 40-series graphics cards based on the "Ada" architecture. These are expected to nearly double the performance over the present generation, ushering in a new era of photo-real graphics as we inch closer to the metaverse. NVIDIA CEO Jensen Huang is expected to take center-stage to launch these cards.15:00 UTC: The show is on the road.15:00 UTC: AI remains the center focus, including how it plays with gaming.
15:01 UTC: Racer X is a real-time interactive tech demo. Coming soon.
15:02 UTC: Future games will be simulations, not pre-baked- Jensen Huang15:03 UTC: This is seriously good stuff (RacerX). It runs on a single GPU, in real-time, uses RTX Neural Rendering15:05 UTC: Ada Lovelace is a huge GPU15:06 UTC: 76 billion transistors, over 18,000 shaders. 76 billion transistors, Micron GDDR6X memory. Shader execution reordering is major innovation, as big as out-of-order execution for CPUs, gains up to 25% in-game performance. Ada built on TSMC 4 nm, using 4N, a custom process designed in together with NVIDIA.
There's a new streaming multiprocessor design, with a total of 90 TFLOPS. Power efficiency is doubled over Ampere.
Ray Tracing is on the third generation now, with 200 RT TFLOPS and twice the triangle intersection speed.
Deep Learning AI uses 4th gen Tensor Cores, 1400 TFLOPS, "Optical Flow Accelerator"15:07 UTC: Shader Execution Reordering similar to the one we saw with Intel Xe-HPG15:08 UTC: Several new hardware-accelerated ray tracing innovations with 3rd gen RTX.15:09 UTC: DLSS 3 is announced. It brings with it several new innovations, including temporal components, and Reflex latency optimizations. Generates new frames without involving the graphics pipeline.15:11 UTC: Cyberpunk 2077 to get DLSS 3 and SER. 16 times increase in effective performance using DLSS 3 vs. DLSS 1. MS Flight Simulator to get DLSS 3 support15:13 UTC: Portal RTX, a remaster just like Quake II RTX, available from November, created with Omniverse RTX Remix.15:14 UTC: Ada offers a giant leap in total performance. Everything has been increased 40 -> 90 TFLOPS shader, 78 -> 200 TFLOPS RTX, 126 -> 300 TFLOPS OFA, 320 -> 1400 TFLOPS Tensor.15:17 UTC: Power efficiency is more than doubled, but power goes up to 450 W now.15:18 UTC: GeForce RTX 4090 will be available on October 12, priced at $1600. It comes with 24 GB GDDR6X and is 2-4x faster than RTX 3090 Ti.15:18 UTC: RTX 4080 is available in two versions, 16 GB and 12 GB. The 16 GB version starts at $1200, the 12 GB at $900. 2-4x faster than RTX 3080 Ti.15:19 UTC: New pricing for RTX 30-series, "for mainstream gamers", RTX 40-series "for enthusiasts".15:19 UTC: "Ada is a quantum leap for gamers"—improved ray tracing, shader execution reordering, DLSS 3.15:20 UTC: Updates to Omniverse
15:26 UTC: Racer X demo was built by a few dozen artists in just 3 months.15:31 UTC: Digital twins would play a vital sole in product development and lifecycle maintenence.15:31 UTC: Over 150 connectors to Omniverse.15:33 UTC: GDN (graphics delivery network) is the new CDN. Graphics rendering over the Internet will be as big in the future as streaming video is today.15:37 UTC: Omniverse Cloud, a planetary-scale GDN15:37 UTC: THOR SuperChip for automotive applications.15:41 UTC: NVIDIA next-generation Drive
15:01 UTC: Racer X is a real-time interactive tech demo. Coming soon.
15:02 UTC: Future games will be simulations, not pre-baked- Jensen Huang15:03 UTC: This is seriously good stuff (RacerX). It runs on a single GPU, in real-time, uses RTX Neural Rendering15:05 UTC: Ada Lovelace is a huge GPU15:06 UTC: 76 billion transistors, over 18,000 shaders. 76 billion transistors, Micron GDDR6X memory. Shader execution reordering is major innovation, as big as out-of-order execution for CPUs, gains up to 25% in-game performance. Ada built on TSMC 4 nm, using 4N, a custom process designed in together with NVIDIA.
There's a new streaming multiprocessor design, with a total of 90 TFLOPS. Power efficiency is doubled over Ampere.
Ray Tracing is on the third generation now, with 200 RT TFLOPS and twice the triangle intersection speed.
Deep Learning AI uses 4th gen Tensor Cores, 1400 TFLOPS, "Optical Flow Accelerator"15:07 UTC: Shader Execution Reordering similar to the one we saw with Intel Xe-HPG15:08 UTC: Several new hardware-accelerated ray tracing innovations with 3rd gen RTX.15:09 UTC: DLSS 3 is announced. It brings with it several new innovations, including temporal components, and Reflex latency optimizations. Generates new frames without involving the graphics pipeline.15:11 UTC: Cyberpunk 2077 to get DLSS 3 and SER. 16 times increase in effective performance using DLSS 3 vs. DLSS 1. MS Flight Simulator to get DLSS 3 support15:13 UTC: Portal RTX, a remaster just like Quake II RTX, available from November, created with Omniverse RTX Remix.15:14 UTC: Ada offers a giant leap in total performance. Everything has been increased 40 -> 90 TFLOPS shader, 78 -> 200 TFLOPS RTX, 126 -> 300 TFLOPS OFA, 320 -> 1400 TFLOPS Tensor.15:17 UTC: Power efficiency is more than doubled, but power goes up to 450 W now.15:18 UTC: GeForce RTX 4090 will be available on October 12, priced at $1600. It comes with 24 GB GDDR6X and is 2-4x faster than RTX 3090 Ti.15:18 UTC: RTX 4080 is available in two versions, 16 GB and 12 GB. The 16 GB version starts at $1200, the 12 GB at $900. 2-4x faster than RTX 3080 Ti.15:19 UTC: New pricing for RTX 30-series, "for mainstream gamers", RTX 40-series "for enthusiasts".15:19 UTC: "Ada is a quantum leap for gamers"—improved ray tracing, shader execution reordering, DLSS 3.15:20 UTC: Updates to Omniverse
15:26 UTC: Racer X demo was built by a few dozen artists in just 3 months.15:31 UTC: Digital twins would play a vital sole in product development and lifecycle maintenence.15:31 UTC: Over 150 connectors to Omniverse.15:33 UTC: GDN (graphics delivery network) is the new CDN. Graphics rendering over the Internet will be as big in the future as streaming video is today.15:37 UTC: Omniverse Cloud, a planetary-scale GDN15:37 UTC: THOR SuperChip for automotive applications.15:41 UTC: NVIDIA next-generation Drive
333 Comments on NVIDIA Project Beyond GTC Keynote Address: Expect the Expected (RTX 4090)
As I understand, the hardware is there right from Turing, but whether DLSS 3 works on it or not, we'll see. Let's be honest:
1. The economic recession didn't start with covid. It started with the way governments around the world reacted to it. They could have implemented sensible safety measures, but noooo... they had to press the big red stop button on the economy. They created an artificial supply chain problem that didn't have a reason to exist. Of course companies like Nvidia jumped on it to make profit.
2. With international companies stopping business with Russia, they are ridding themselves of a huge market. It's a loss that has to be recouped by artificial price hikes. Fewer products sold = higher margins per unit. All of this while Russia still gets everything through black/grey import and China. Who's the real loser in this situation? We are. No Western company should have ever stopped business with Russia.
We tend to say that we're so much more enlightened than we were during the world wars and before, yet here we are, suffering the costs of political ideology once again.
Consider prices in Norway for electricity. x10 the normal price. Literally 10 times more. Previously it was due to covid x2 x3 or so. (infected cables I suppose) Now there is war in Ukraine. There is always a valid reason that will be used for the elites to justify their doings and normal people will have to pay for it. So these elites can feel superior. And yet Norway has so much renewable energy which people paid for in the first place in order for the renewable energy to contribute to the economy. So in order to get rich you have to take it from someone else unfortunately.
NV is doing this and I'm sure AMD will follow all because someone wants to have more power or more money a lot of people have to pay for it.
But back to topic: I like the idea of the 4080 12 GB, I really do. But I can't justify spending nearly a grand on a graphics card alone. It's more like the price of a complete system overhaul for me - which I might do if AMD comes up with better pricing. Or maybe the Arc A770 ends up being dirt cheap (it'll have to be to compete), so maybe I'll just buy one and call it a day. At least it'll be something different.
On a separate note even with these crazy priced GPUs I'll still be happy for the tech enthusiasts who pick them up this fall/winter.
It's a shame Nvidia is pricing a lot of people out but I still wonder if that's due to them likely having over a year of ampere stock... The 4080 12GB looks like it's at best gonna match a 3090 that's already priced about the same, yeah you get DLSS 3.0 but that's still basically 0 progress at a similar price other than RT so Turing 2.0 it seems. The next 4-6 months should be interesting.
As far as AMD goes one of two things happened. Nvidia caught wind of actual RDNA3 performance and wasn't very impressed or they are super arrogant and will be blindsided by it's performance... Hoping for the latter.
It would also be trivially simple to solve this: go from a "highest bid sets the universal price" system to a system that distributes the price difference from more expensive forms of energy onto cheaper ones, averaging out prices rather than raising them all to match the most expensive. Sure, this would be more complex than the current system (you wouldn't be able to set final prices until after all sales were settled), but it really wouldn't be that hard - it would just require active regulation. Instead, we're maintaining a system that now works explicitly to shovel money into the coffers of whoever is producing cheap energy, as they literally can't sell it anywhere near cost. (Of course, there's also a somewhat unique situation in southern Norway with a very dry summer and thus little water in hydroelectric reservoirs, which drives up costs through fear of a future shortage, alongside increases in consumption from increased electrification of industry and widespread adoption of electric cars, which have yet to be met by increased production of electricity.) From what I gathered, DLSS 3 is reliant on a degree of FP8 support that no generation previous to Lovelace has. I mean, they're very explicit about previous generations only supporting DLSS2 on their site:
I think the 4080 12GB will probably be a good GPU in a vacuum - it's just such an insulting proposition alongside the 16GB, the 4090, and the 30 series. A $200 price increase - and this isn't even the 8-tier card, but an explicitly and significantly cut-down variant? The only "value" proposition of this is "hey, look, it's supposed to match the 3090 at a much lower price" - which of course ignores the fact that the 3090 was a stupidly overpriced card to begin with.
If they called this the 4070 Ti and sold it for $600, it would be fantastic even if that would still represent a significant per-tier price hike. Instead, they're framing this as "8-tier for those of you who can't afford the full-fat version", on top of a $200 price increase. It's just such an obvious slap in the face. I'm not saying that you did, but the rhetoric of "I worked hard for my wealth" can't be extracted from its inherent politics or implications. Claiming that such statements aren't making a strong implication that less wealthy people are either not working as hard, and/or are just less deserving overall is ... well, IMO impossible - after all, what you said strongly references meritocracy and the idea that there's a linear(ish) relation between work and rewards (both of which are pure and utter BS). Which is what I was pointing out. I'm not saying that you were explicitly saying this, but the implication is there whether you like it or not. There is no other logically sound argument to be extracted from what you said - and after all, you were quite explicit in your "if it's too expensive, then this just isn't for you" approach. I'm not putting words in your mouth, just making the very strong implications from your rhetoric and point of view explicit. You're very welcome to argue for how you mean something different than that - but that would require actual arguments. Beyond that, putting the two statements of your post together very strongly implies "well, if you can't afford this, you haven't been working hard enough, unlike me". That might not be what you intended to say, but it's right there in what you said regardless of intention.
Please look at that,
If Lovelace delivers packed FP8 instructions (i.e. 1xFP32, 2xFP16, 4xFP8) through the tensor cores, then for anything actually running FP8 that is an instant 2x performance increase over previous generations running the same workload, with no real way of working around it.
This also doesn't really say that DLSS3 is necessarily laggy on Lovelace - as long as the FP8 can keep up, delivering that interpolated frame in the gap between the previous and next frame, then it should be completely fine. There's no reason why this should increase input lag unless the interpolated frame causes the next "real" frame to be delayed.
One can think of it as Nvidia innovating gaming technologies, but in a world where we also have FSR that runs even on Intel Xe integrated graphics, I'd rather not. Yes. It's a x70-tier card with a x70-tier GPU (AD104 - the xx4 has always been x60 or x70 level) and x70-tier VRAM, but with an x80 name so they can sell it for more money. Disgusting. I can see it being laggy if the generated frames don't rely on user input - and I don't see how they would.
The absurd thing here is the system as I described it above, and how it is built so that it shovels money into the pockets of whoever is generating cheap electricity during a shortage. This has nothing to do with taxation or who is paying for what, but is down to a fundamentally flawed system of trade regulation, which needs to be fixed. They've been pretty explicit about this from the launch - it's one of the main selling points of Lovelace. Yep, obviously. It does a lot of things at once: makes the $1200(!!!) 16GB look less completely absurd ("there's a cheaper 8-tier too!"); it deliver a "reasonable" 8-tier GPU ("it's only +$200 from the 3080, and it's much faster!"), it hides the massive performance difference between the two (they're both "the 4080"), and it allows Nvidia to keep pushing prices ever higher without implementing a much-needed change to their model tier scheme. I don't see it as a parallel to mobile though - mobile has real-world physical limitations that necessitate performance differences, and it's a distinct and discrete product segment fromd desktops even if they share core naming conventions. A mobile 3080 is still the 8-tier mobile GPU, the fastest or second fastest there is - you just can't expect a 150W mobile part to match a 330W desktop part. That is fundamentally different from this - the 12GB and 16GB 4080s are just very, very different hardware, with the 12GB having drastically less shaders, memory and memory bandwidth, with cost being the only explanation - but it's still wildly expensive. It just makes zero sense outside of a perspective that only cares about Nvidia's profits.
I hope AMD will do better and the price probably will be OK-ish but not good. The prices have been going up since Turing and they are accelerating.
www.tomshardware.com/news/nvidia-dlss-3-only-works-with-geforce-rtx-40-series-gpus-for-now
In Poland, my friend pays 320NOK per 2 MOTHS OF ELECTRICITY using around 250-300KW/H per month (you get it? ) Two fucking months. You know how much I paid for 300KW/H for last month? 2200NOK
Germany, the last time I checked they were only concerned about the gas price. When I mentioned Electricity, they have asked my. What electricity prices?
Energy meaning GAS PRICE not electricity per se. how do I know? My wife lived there for 7 months before moving in to Norway. Of course they have had a price hike but not ten times more like in Norway. No way you going to tell me that the price in Germany and Poland for instance (also France and Spain) went up ten times over the last 2 years.
So no it is not a nonsense to me rather it is nonsense of what the price for electricity is here in Norway when you clearly produce electricity from renewable sources which is supposed to be free. Main selling point or downfall when people see they get stiffed. DLSS 3 is out so who cares about DLSS 2 development?
At least vs Sweden and Denmark, Norway have a way higher percentage use spot prices, instead of fixed long term prices. I'm assuming it's similar in the mentioned places. I've seen people from UK being nervous about what their new fixed price plan will be when their current expires.
Of course prices in energy will increase, when demand is above/close to supply.
Won't DLSS fall away as a selling point now? Will new games stop getting DLSS2? Can 4xxx buyers trust being able to use DLSS3 (or next versions), or will it just disappear when 5xxx comes with DLSS4 (if it does, who knows, not buyers)?
Not sure I like it.
I can see two reasons for choosing this path: competitive pressure from AMD, and a desire to sell hyper-expensive ultra-flagships. Without either of those, they could have scaled AD102 down to a die size like the GP102 instead, making a much smaller and more reasonably sized die that would have delivered tons of performance still. But no - they chose to go balls-to-the-wall, on a super expensive node.
I mean, what if they went a bit more moderate with this instead - or at least didn't try to make a smooth gradient in pricing from a 380mm2 die to a 600mm2 one? If TPU's die size for the AD103 is correct, charging $1200 for that and $1600 for the 4090 - at 60% more die area, and 50% more VRAM - is downright absurd. But Nvidia has clearly chosen the path of "we'll sell on mindshare and flagship cred alone, screw any idea of value". Sorry, but where are you getting the idea that energy from renewable resources is supposed to be free? Do you imagine that building, running and maintaining a power plant doesn't have a cost? They're cheap in the long run, but not free.
Also, did you read anything at all of the post I responded to? The countries you mention do not use electricity for heating, they use gas for heating. Gas is their major domestic energy expenditure, and gas is traded in the same system as electricity in the EU. There are of course exceptions - there is a move towards less reliance on natural gas for heating, but pricing hasn't followed yet, meaning anyone there using electricity for heating is in a pretty good position.
As for gas prices: more than 5x higher according to this source (and historically, EU energy prices have been significantly higher than Norwegian prices, which makes any increase here look bigger). Here are reports of recent energy pricing protests in Germany; energy prices are a big part of the ongoing and hotly protested cost of living crisis across the continent. It's not getting much exclusive attention because it's tied into rising prices for food and other necessities as well, but energy prices are a central part of this crisis.
I mean, you're explicitly contradicting yourself here, on the one hand you say "they pay peanuts for energy", and on the other you say "they pay for gas, not electricity", which ... well, if you read my previous post, maybe you'd understand why this is effectively the same thing - it's the core domestic source of energy, is traded within the same system, and is used for the same things. Norway uses mainly electricity and that's where prices are high; continental Europe uses mostly gas and that's where prices are high. Trying to separate these two is actively misleading and just downright misrepresenting reality - the current situation in Norway is in no way unique. Yes, it is a private exchange - just as most stock markets are private - but it's still subject to EU trade regulations. And the EU could quite easily mandate a change in their trade policies. And crucially, nobody here is arguing that it's unnatural for prices to rise, it's just the magnitude of this rise and how it's intrinsically tied into the mechanisms of the trade system rather than any even remotely sensible system of price setting that makes this problematic. The solution is simple: price regulation targeted towards averaging out prices closer to the average cost of production/sourcing the energy. This will of course eat into corporate profits, but, well, tough luck. Corporations do not have a right to exploit people in a crisis. This might well be - but then, games that are playable at 30-45fps are typically not all that reliant on smooth and rapid input, so the difference will also be less noticeable. But if the "source" framerate is indeed as low as 15fps, this will most likely be unplayable, yes. True to some extent, but native resolution gaming is also a rather silly brute-force solution as resolutions scale higher, simply because the perceptible increase in detail and sharpness is pretty much inversely proportional to the resolution increase at this point. 4K has 4x the pixels of 1080p, and is clearly sharper even at 27", but it's not night and day. 8k is 4x the pixels of 4k, and the increase in sharpness is essentially imperceptible unless you're sitting very close to a very large TV. And as new nodes and increased transistor density becomes more difficult, we need to abandon the simplistic brute force solutions for improved visual fidelity - they're getting too expensive. If moving up one step in resolution has a 4x compute cost but a ... let's say 50% increase in perceptible detail/sharpness, then that is terrible, and never worth it. Upscaling is really the only viable way forward - though precisely how said upscaling will work is another question entirely.
I think cables are a good analogy for this: the signal requirements for the excessive bandwidth of DP 2.0 and HDMI 2.1 isn't bringing with it thumb-thick cables, but rather bringing about a shift to active cabling instead of passive copper. This takes us from a simple, brute-force solution to a more complex one. Where this analogy falls apart is that active cabling is easily 10x the BOM cost of passive, while upscaling is about as close to a free performance upgrade as you'll find. But it's another example of needing to find more complex, smarter solutions to a problem as the older brute-force ones are failing.
4070*khm* 4080 12 GB, which are also marketed as semi-flagship models at such prices. Never has a xx4 chip been in a flagship product. Ever. They're basically saying "it's not a flagship product, but you'll think it is and you'll buy one because you're stupid." That's why I still play at 1080p native. Brute force is the way. :cool: