A 4070 SUPER is better than a 3090? Wow. Nvidia showcase SUPER GPUs

[filpo]

Summary

Today Nvidia showed their new SUPER GPUs off and they're looking kinda SUPER going from the top

 

Quotes

Specifications

Quote

Comparison of SUPERs and non-SUPERs

 

Quote

4080 SUPER

The NVIDIA GeForce RTX 4080 SUPER utilizes the AD103-400 GPU (PG139 SKU 355) with the full 10,240 cores, 320 TMUs, 112 ROPs, and 64 MB of L2 cache. The GPU will also be clocked slightly higher at a 2295 MHz base and 2550 MHz boost versus the 2205 MHz base and 2505 MHz boost of the RTX 4080 Non-SUPER variant.

 

The graphics card will use a 256-bit bus interface with 16 GB GDDR6X memory, the same as the RTX 4080 Non-SUPER but boosted to 23 Gbps, making it the fastest G6X solution in the world, offering up to 736 GB/s of bandwidth which is around a 3% boost over the Non-SUPER variant. The NVIDIA GeForce RTX 4080 SUPER will feature the same TGP as the RTX 4080 Non-SUPER at 320W. The card offers 52 Shader TFLOPs, 121 RT TFLOPs, and 836 AI TOPs while supporting all the latest NV encode technologies such as AV1 and H.264.

 

Quote

4070 Ti SUPER

The NVIDIA GeForce RTX 4070 Ti SUPER will feature a choice of AD103-275 / AD102-175 GPU SKUs (PG141 SKU 323) with 8,448 cores across both variants & 48 MB of L2 cache. The graphics card will also feature 16 GB of GDDR6 memory which will be a 4 GB VRAM upgrade vs the RTX 4070 Ti Non-SUPER and support a 256-bit bus interface. The graphics card will retain a 285 Watt TGP which is the same as the Non-SUPER variant.

 

The card offers 44 Shader TFLOPs, 102 RT TFLOPs, and 706 AI TOPs while supporting all the latest NV encode technologies such as AV1 and H.264. All of these combined put the RTX 4070 Ti within 90% of the performance of the original RTX 4080 and it essentially replaces the RTX 4070 Ti which also had the same MSRP.

 

Quote

4070 SUPER

 

The NVIDIA GeForce RTX 4070 SUPER will feature AD104-350 or AD103-175 GPU (PG141 SKU 335) configurations with the same 7,168 cores across both variants. The GPU is expected to get 36 MB of L2 cache. The card will stick with the 12 GB GDDR6X memory across the same 192-bit bus interface. For the RTX 4070 SUPER, the card will have a 20W higher TBP than the existing 4070 at 220W.

 

The card offers 36 Shader TFLOPs, 82 RT TFLOPs, and 568 AI TOPs while supporting all the latest NV encode technologies such as AV1 and H.264. All of these combined put the RTX 4070 within 95% of the performance of the 4070 Ti.

Gaming Performance

Quote

4080 SUPER

In terms of gaming performance, the NVIDIA GeForce RTX 4080 SUPER is said to be around 3% faster than the RTX 4080 Non-SUPER which isn't a huge gain but custom models will deliver close to 5% performance boosts. NVIDIA also mentioned that in CPU-bound scenarios, the RTX 4080 SUPER and RTX 4080 Non-SUPER are very similar in performance but the card offers a 2x boost over the RTX 3080 Ti with frame-gen applied, something that the older Ampere and Turing cards don't offer. The AMD Radeon RX 7900 XTX GPU is the direct competition for this card as it retails for the same price. The RTX 4080 and 7900 XTX trade blows in rasterized performance with the RDNA 3 GPU taking an edge in certain titles but with RT and more advanced graphical features, the NVIDIA GPU blazes ahead so the RTX 4080 SUPER is going to take the lead further at an even better value. Personally, this should've been the specs, price, and performance for the RTX 4080 since the start.

 

Quote

4070 Ti SUPER

In terms of gaming performance, the NVIDIA GeForce RTX 4070 Ti SUPER is said to be around 15% faster than the RTX 4070 Ti Non-SUPER which isn't a huge gain but custom models will deliver close to 15-20% performance boosts. The card is also stated to offer a 2.5x boost over the RTX 3070 Ti with frame-gen enabled and a 60% uplift versus the same card without frame-gen. The AMD Radeon RX 7900 XT GPU is the direct competition for this card as it retails super close at $849 US. The RX 7900 XT still holds the higher VRAM advantage with its 20 GB capacity while the RTX 4070 Ti will tackle it in rasterized and ray-traced games considering it is around 90% of the performance of a 4080 which rivalled the 7900 XTX. 

 

Quote

4070 SUPER

In terms of gaming performance, the NVIDIA GeForce RTX 4070 SUPER is said to be around 15% faster than the RTX 4070 Non-SUPER, marking the biggest uplifts in the SUPER lineup. The card is also stated to be faster than an RTX 3090 without frame-generation though the performance is going to vary on a game-to-game basis. NVIDIA also confirms that the card is just 5% slower than an RTX 4070 Ti which retailed for $799 US. That too has been replaced by the RTX 4070 Ti SUPER. The card will be challenging the AMD Radeon RX 7900 GRE and RX 7800 XT. The GRE retails for $649 US while the 7800 XT retails for $499 US. Having 4070 Ti or 3090-class performance at $599 US is a very aggressive approach by NVIDIA and one that should definitely be appreciated by gamers. At the same time, the NVIDIA GeForce RTX 4070 will continue to live on with a $549 US price tag but at this point, it is better to go for the new SUPER variant.

Power efficiency - Another talking point for the NVIDIA GeForce RT 40 SUPER GPUs is its class-leading performance efficiency

Quote

4080 SUPER

While the GPU is rated at 320W, the actual gaming wattage is around 246W on average which is almost half the TGP of the RTX 4090 and much lower than what the Radeon RX 7900 series consumes. The video playback power is also 22W out of the box while idle power is 15W, something that the AMD RDNA 3 GPUs had trouble fixing during launch.

4070 Ti SUPER

While the GPU is rated at 285W, the actual gaming wattage is around 240W on average which is almost half the TGP of the RTX 4090 and much lower than what the Radeon RX 7900 series consumes. The video playback power is also 20W out of the box while idle power is 12W, something that the AMD RDNA 3 GPUs had trouble fixing during launch.

4070 SUPER

While the GPU is rated at 220W, the actual gaming wattage is around 200W on average which is less than half the TGP of the RTX 4090 and much lower than what the Radeon RX 7900 series consumes. Even the 7800 XT is rated at a much higher 263W TBP. The video playback power is also 16W out of the box while idle power is 11W, something that the AMD RDNA 3 GPUs had trouble fixing during launch.

(Diagram is same for all the GPUs)

 

Quote

4070 Ti SUPER

While the GPU is rated at 285W, the actual gaming wattage is around 240W on average which is almost half the TGP of the RTX 4090 and much lower than what the Radeon RX 7900 series consumes. The video playback power is also 20W out of the box while idle power is 12W, something that the AMD RDNA 3 GPUs had trouble fixing during launch.

AI Performance

Quote

Goes for all of the SUPER GPUs

AI is a major talking point for the NVIDIA GeForce RTX 4080 SUPER and the RTX 40 SUPER series in general, delivering up to 2x faster performance with DLSS 3.5, 70% faster image generation & 50% faster video generation speeds using ML optimizations through the Tensor-RT and Tensor-RT LLM stack that has been available on Windows and is being updated on the go.

Pricing and Availability

Quote

The NVIDIA GeForce RTX 4080 SUPER GPU will be available on the 31st of January in both Founders Edition and Non-FE models, starting at $999 US. The NVIDIA GeForce RTX 4070 Ti SUPER GPU will be available on the 24th of January in custom variants only since there's no Founders Edition of the card. It will start at $799 US. The NVIDIA GeForce RTX 4070 SUPER GPU will be available on the 17th of January in both Founders Edition and custom variants, starting at $599 US.

 

My thoughts

It's a bit strange that they're using 5 and nearly 6 year old GPUs in their presentation. I understand that these GPUs are supposed to be looked at as an upgrade from those GPUs but if you're watching this presentation, you're probably going to be knowledgeable enough most likely already know the capacity of a 2070 for example. Then again it makes sense when they're the same class of GPU, but then shouldn't the 2070 SUPER have been used here? And the 3070 Ti (since it wasn't a large upgrade over the 3070)? Anyway, I like to see that they've learnt from they're pricing mistake from last year. Overall, I'd say good job Nvidia (no sarcasm unless launch day tells us otherwise)

 

Sources

NVIDIA launches GeForce RTX 40 SUPER series: $999 RTX 4080S, $799 RTX 4070 TiS and $599 RTX 4070S - VideoCardz.com

NVIDIA GeForce RTX 4080 SUPER GPU Comes Crushing Down on AMD's 7900 XTX With $999 US Price: Over 10K Cores, 16 GB GDDR6X Memory (wccftech.com)

NVIDIA GeForce RTX 4070 Ti SUPER Unleashed For $799 US: 16 GB G6X Memory, AD103 GPU, Aims For 7900 XT (wccftech.com)

(Link for 4070 SUPER Wccftech post isn't pasting for me so here's a screenshot of the page, if you wanna see the whole page go to Wccftech and it should be one of the first ones, you might need to scroll down a bit)

Edited January 9 by filpo
Edited title and removed misinformation

[Dellenn]

What I found weird was their comparisons. 

They compared a 4080 Super to a 3080ti, but the 4070 Super vs the 3090. 

[filpo]

1 minute ago, Dellenn said:

What I found weird was their comparisons. 

They compared a 4080 Super to a 3080ti, but the 4070 Super vs the 3090. 

That's Nvidia for you. I miss the good old days of the 10 Series. Even the 30 series wasn't all that bad (apart from transient spikes)

[GuiltySpark_]

Pricing on some is good and 4080 and 4070Ti being discontinued/replaced is also good. AMD pricing for 7900 series cards needs to be looked at.

 

Hopping around to various marketplaces, 7900xtx's already popping up for sale.

 

[filpo]

5 minutes ago, GuiltySpark_ said:

Hopping around to various marketplaces, 7900xtx's already popping up for sale.

in the US? I saw the 7900 XT for $760 XFX Speedster MERC 310 Black Edition Radeon RX 7900 XT 20 GB Video Card (RX-79TMERCB9) - PCPartPicker and the 7800 XT quite close to it's launch price (at $510 now) but the 7900 XTX seems to be steady at it's now supposedly $960-$970 MSRP

[RejZoR]

It's all pointless with astronomic prices that all the cards have. Also comparing it to idiotically inflated price of RTX 3090Ti, everything is a bargain. But don't kid yourselves that pricing of RTX 4000 series is normal. It's idiotic and it'll just be worse. If anyone thinks new cards will be cheaper, then you're mistaken. They are positioned between regular and Ti models. They'll be priced in between. Literally no one asked for any of this crap.

 

Also the way lower end cards basically don't scale in performance compared to old generation should be a warning sign.

[Agall]

RTX 3000	CUDA Cores	%cores	%cores	CUDA Cores	RTX 4000
GA102	10752	100%	100%	18432	AD102
3090ti 24GB $2000	10752	100%
3090 24GB $1500	10496	97.6%
3080ti 12GB $1200	10240	95.2%
			88.9%	16384	4090 24GB $1600
3080 12GB $800	8960	83.3%
3080 10GB $700	8704	81.0%	79.2%	14592	4090D 24GB $1600
3070ti 8GB $600	6144	57.1%	55.6%	10240	4080 S 16GB $1000
3070 8GB $500	5888	54.8%	52.8%	9728	4080 16GB $1200
3060ti 8GB $400	4864	45.2%	45.8%	8448	4070ti S 16GB $800
			41.7%	7680	4070ti 12GB $800
			38.9%	7168	4070 S 12GB $600
3060 12GB $330	3584	33.3%	31.9%	5888	4070 12GB $600
3050 8GB $250	2560	23.8%	23.6%	4352	4060ti 8GB $400
3050 6GB $250	2048	19.0%	16.7%	3072	4060 8GB $300
			13.9%?	2560?	4050 6GB? $250?

 

Update my chart in gauging these by their binning scheme. RTX 4070ti Super is just an RTX 3060ti tier bin for double the price. Still likely end up a good card at $800, but probably not worth it over another $200 for the 4080S.

[Timmahnz]

On 1/9/2024 at 6:44 AM, RejZoR said:

It's all pointless with astronomic prices that all the cards have. Also comparing it to idiotically inflated price of RTX 3090Ti, everything is a bargain. But don't kid yourselves that pricing of RTX 4000 series is normal. It's idiotic and it'll just be worse. If anyone thinks new cards will be cheaper, then you're mistaken. They are positioned between regular and Ti models. They'll be priced in between. Literally no one asked for any of this crap.

 

Also the way lower end cards basically don't scale in performance compared to old generation should be a warning sign.

I paid $1100 NZD for my 1080Ti as the mining boom started.. I'm likely to pay near $2000NZD for a 4070Ti Super as it's replacement. Suffice to say.. Consumers have decided that these prices are acceptable, as well as demand for their AI capabilities are dictating prices. Add that in today's economic climate and Nvidia's clever marketing, we are not going to see prices pre-covid, pre-mining or price wars for a long long time, if ever.

[StDragon]

5 hours ago, Timmahnz said:

Consumers have decided that these prices are acceptable, as well as demand for their AI capabilities are dictating prices. Add that in today's economic climate and Nvidia's clever marketing, we are not going to see prices pre-covid, pre-mining or price wars for a long long time, if ever.

Nvidia will get squeezed.

Once AI specific hardware floods the market, and iGPU/APUs reach the point of "good nuff" for the majority of the gaming market, Nvidia won't be able price their GPUs that high. Even China will have a consumer GPU that will eventually reach the shores of the West.

Don't get me wrong, Nvidia is the best there is for gaming, but with alternatives and changing HW dominance, they will be forced to reevaluate their pricing structure.

[Kisai]

7 hours ago, Timmahnz said:

Suffice to say.. Consumers have decided that these prices are acceptable, as well as demand for their AI capabilities are dictating prices. Add that in today's economic climate and Nvidia's clever marketing, we are not going to see prices pre-covid, pre-mining or price wars for a long long time, if ever.

I don't think the AI is going to drive consumer cards at all. Mining drove shortages of consumer GPU's because people built shoddy mining rigs from whatever they could buy or steal.

 

AI, nearly all training is going to be done on data center GPU's, and the only "realtime" inference that's worth using consumer GPU's on are things like TTS and similar audio generative AI projects because they aren't taxing on the GPU resources to get a realtime auto generation, but you still need something like a x70/x80 part to be able to get that low latency. If you do it on a x50 or x60 product, the latency is significant.

 

Everyone writing their own chatbot/assistant/waifubot's can do this on a computer with 3 GPU's. Here's a use case, and bear in mind this isn't middle-of-the road use case, it's an optimal one that you'd expect real time output from, not the joke of a response time you get from Siri/Alexa:

- x80Super/Ti series part - ASR (Audio recognition) , Whisper needs 16GB VRAM to attain usable accuracy, other options exist (eg Vosk) but they still need to run on the GPU with a large model to be useful. "lite" models to run on mobile phones are barely accurate.

- x70 series part for TTS (8GB = 140 characters), x90 series if you want emotional configurability in realtime, since it does a zero-shot train on every sentence. There are several models ranging from radtts to VITS that produce good enough audio reproduction

- 4090 series for an LLM (70B LLaMA 2 model requires 24GB x90 part, and the 4090 is the only card that can run it at present, 13B model can run on a 12GB GPU)

 

So before we get into any CV parts, we already need 3 RTX x90's (maybe two 3090's and a 4090) if you want to build a low-latency chatbot without any cloud parts, or 3 x 16GB x80's if you want to build a less flexible one.

 

Then if you want to get into CV, you also need:

CLIP (is what forms the basis of Dall-E Stable Diffusion, and pretty much all AI models that identify things and objects at present. (For how poor these models really are, see https://www.theverge.com/2021/3/8/22319173/openai-machine-vision-adversarial-typographic-attacka-clip-multimodal-neuron )

openCLIP exists, but I don't have any data on the requirements.

 

And also a way to locally fine tune those AI's with new data every day so they don't become stale and can react to news.

 

Then if you want a moving "AI" character, you need two additional GPU's, one x90 again for doing the movement, and one x60 for rendering it.

 

So we're up to 6 GPU's now. While you don't need these to be in one computer, you do need to have them on the same network, on the same rack. Which means you need at least one more computer to tie everything together.

 

[RejZoR]

10 hours ago, Timmahnz said:

I paid $1100 NZD for my 1080Ti as the mining boom started.. I'm likely to pay near $2000NZD for a 4070Ti Super as it's replacement. Suffice to say.. Consumers have decided that these prices are acceptable, as well as demand for their AI capabilities are dictating prices. Add that in today's economic climate and Nvidia's clever marketing, we are not going to see prices pre-covid, pre-mining or price wars for a long long time, if ever.

GTX 1080Ti was 850€ when I bought it basically after launch. I paid for RTX 3080 999€. I've decided I'll never pay over 1000€ for a graphic card. And I'm gonna be damn stubborn about it. Also I follow my own rule of changing graphic card only when performance is doubled. RTX 3080 pretty much doubled performance of GTX 1080Ti. So I pulled a trigger on it. RTX 4080 is far from it so it's just gonna wait. I can play all games at max super smoothly with ray tracing at 1440p so I'm in no rush to upgrade either.

[Stahlmann]

This announcement gives me a little bit of hope that the 5000-series will actually be a meaningful fps/$ improvement again. There are already games where my 3080 struggles at 4K, but any meaningful upgrade currently available is just unjustifiably expensive.

 

If the 5080 can make the same +50% performance jump as the 4080 (compared to the 3080) and also stays at 1000€ then i'll consider an upgrade again.

[thorhammerz]

2 hours ago, Stahlmann said:

but any meaningful upgrade currently available is just unjustifiably expensive.

On the flip side: because meaningful upgrades are expensive (if at all present), while at the same time the hardware itself can last near a decade (if maintained in good care), over the long run consumers won't be spending that much more (if anything, possibly less) as the need to upgrade for most people is simply not urgent anymore.

[Stahlmann]

7 minutes ago, thorhammerz said:

On the flip side: because meaningful upgrades are expensive (if at all present), while at the same time the hardware itself can last near a decade (if maintained in good care), over the long run consumers won't be spending that much more (if anything, possibly less) as the need to upgrade for most people is simply not urgent anymore.

I agree. Video game graphics also have plateaued over the last few years, so that also makes the upgrade cycle significantly slower than 5 years ago for example. My 3080 would probably also last significantly longer if i wouldn't use a 4K display to begin with.

[thorhammerz]

13 minutes ago, Stahlmann said:

I agree. Video game graphics also have plateaued over the last few years, so that also makes the upgrade cycle significantly slower than 5 years ago for example. My 3080 would probably also last significantly longer if i wouldn't use a 4K display to begin with.

Agree; sitting on a 10 series GPU and am only starting to consider shopping when the 5000 series drops (8 years will have been a good run ).

 

I'd argue the most important factor for the upgrade cycle slowing is on the PC side: most rigs built in the past several years are very unlike the ones from the early/mid 2010's as widespread adoption of SSDs means PCs are no longer become unbearably slow after a few years of use.

 

Tech enthusiasts have not come to terms that what's considered "normal" (wrt to rate of performance change & relative pricing) is just as liable to change against their preferences, as it was for their preference last decade.

[Isuck Assimov]

Ti SUPER ???
make it even more confusing, why don't you nvidia..

[RejZoR]

4 hours ago, Isuck Assimov said:

Ti SUPER ???
make it even more confusing, why don't you nvidia..

RTX 4070Ti Super MX GTS Ultra Titan

[Mark Kaine]

what's funny my gpu still goes south of 1000 bucks here in scam paradise.

 

i might be able to sell mine for ~600 and get this new fangled (none founders) Super also for 600 or slightly more?

 

 

[starsmine]

13 hours ago, Isuck Assimov said:

Ti SUPER ???
make it even more confusing, why don't you nvidia..

super is only refresh, TI is not ever a refresh

[bramturismo]

I am confused about the naming of these cards.

 

The 4070 Ti Super is faster than a 4070 Ti, but not as fast as a 4080? Or how does this work? Is it like a half-step up to 5000 series, or are these additional models in the range?

 

It's like Audi naming their cars... You want an A6? Which one, a 30, 35, 40, 50, 55, TFSI, e-tron, quattro, what you want?

[filpo]

Just now, bramturismo said:

or are these additional model in the range?

additional models within the range off 40 series

3 minutes ago, bramturismo said:

The 4070 Ti Super is faster than a 4070 Ti, but not as fast as a 4080?

If I'm not mistaken most of the performance uplift of the 4070 Ti SUPER is going to 16GB of vram

It's a bit like they did with the 20 series, a mild increase. Not so big it's a Ti but not just an addition of VRAM (in the 4070 Ti SUPER's case)

Just now, bramturismo said:

quattro

better yet a quattro is actually an add on to their cars (four wheel drive)

 

Just now, bramturismo said:

t's like Audi naming their cars

but tbh this could go for all car companies. For example you've got BMW with their 318i, 320i, 330i, 330e, M340i and then M3

[bramturismo]

2 minutes ago, filpo said:

additional models within the range off 40 series

If I'm not mistaken most of the performance uplift of the 4070 Ti SUPER is going to 16GB of vram

It's a bit like they did with the 20 series, a mild increase. Not so big it's a Ti but not just an addition of VRAM (in the 4070 Ti SUPER's case)

better yet a quattro is actually an add on to their cars (four wheel drive)

 

but tbh this could go for all car companies. For example you've got BMW with their 318i, 320i, 330i, 330e, M340i and then M3

Alright, so it's a slightly tickled Ti, like buying an M340i and then chip-tuning it. Not quite an M3, but bloody fast.

 

I still find it confusing to have so many different cards in a single generation, but maybe that's just me. Should've put 16Gb of VRAM in the cards to start with to be honest.

[filpo]

4 minutes ago, bramturismo said:

like buying an M340i and then chip-tuning it

more like getting an M340i and putting it into a permanent Sport+ mode instead of leaving it in comfort 

 

4 minutes ago, bramturismo said:

Should've put 16Gb of VRAM in the cards to start with to be honest.

definitely

 

4 minutes ago, bramturismo said:

I still find it confusing to have so many different cards in a single generation

same, I mean I get why they did SUPER in the 20 series (they're only Ti card was the 2080 Ti). They were trying something new. Then when 40 series came along and I saw the 4060 Ti and 4070 Ti I wondered if the SUPER was gonna make a comeback but I don't think anyone could have known they were gonna make a Ti SUPER

[starsmine]

15 minutes ago, bramturismo said:

I am confused about the naming of these cards.

 

The 4070 Ti Super is faster than a 4070 Ti, but not as fast as a 4080? Or how does this work? Is it like a half-step up to 5000 series, or are these additional models in the range?

 

It's like Audi naming their cars... You want an A6? Which one, a 30, 35, 40, 50, 55, TFSI, e-tron, quattro, what you want?

its a refresh. why would a 4070ti refresh be faster than a 4080 and still be called a 4070ti?

[porina]

12 minutes ago, filpo said:

If I'm not mistaken most of the performance uplift of the 4070 Ti SUPER is going to 16GB of vram

10% more cores at near enough same base/boost clocks. However it is same TBP so unless it is binned better it might boost less all else being equal.

33% more VRAM bandwidth, along with the quantity.