NVIDIA 4000 Series

[brostradamus]

whitepaper:

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf

lol the way they start with a 2pg highlight of their trackrecord of innovation..

summary:
- BVH structures have been simplified, probably they have sacrificed BVH depth for a much faster data structure that sacrifices traversals for set-up times - the trade-off has been favorable (also some talk abt a new kind of primitive, perhaps a game dev can explain this bit better)
- RT cores now also include alpha testing capability which was previously done using shaders, this is going to be a big advantage for nvidia in RT - more games with more foliage, looks like nvidia's tessellation moment in RT
- SER: another big innovation in RT, and looks intuitive - nvidia has found a way to treat secondary rays as primary rays, other than the ability to control divergence in execution this is going to make RT more scalable in future.. its just looking like a hardware accelerated recursive structure - pretty neat
- DLSS 3 has been detailed in a separate whitepaper

 

[AthlonXP1800]

whitepaper:

https://images.nvidia.com/aem-dam/Solutions/geforce/ada/nvidia-ada-gpu-architecture.pdf

lol the way they start with a 2pg highlight of their trackrecord of innovation..

Ah interesting.

RTX 4080 12GB and 16GB was confused before the whitepaper published. Everybody thought RTX 4080 12GB was changed from RTX 4070 12GB at last min but the whitepaper finally made it clear that RTX 4080 12GB will replace RTX 3080 10/12GB, not RTX 3070 and RTX 4080 16GB is actually Ti class will replace RTX 3080 Ti, not RTX 3080 10/12GB. I thought RTX 4080 Ti will launch later in early 2023 to replace RTX 3080 Ti.

Well guess I will have to wait until after Navi 31 launch to read reviews to see RTX 4080 12GB and 16GB benchmarks before decide to pick MSI RTX 4080 Suprim X 12GB or Ti class 16GB.

 

[Grim5]

This is cool

A video containing just the generated dlss3 frames from DFs Spider-Man video.

Considering that every single frame in this video is completely fake/generated, the image quality is very good

Watch output_even | Streamable

Watch "output_even" on Streamable.

streamable.com

 

[Twinz]

Ah interesting.

RTX 4080 12GB and 16GB was confused before the whitepaper published. Everybody thought RTX 4080 12GB was changed from RTX 4070 12GB at last min but the whitepaper finally made it clear that RTX 4080 12GB will replace RTX 3080 10/12GB, not RTX 3070 and RTX 4080 16GB is actually Ti class will replace RTX 3080 Ti, not RTX 3080 10/12GB. I thought RTX 4080 Ti will launch later in early 2023 to replace RTX 3080 Ti.

Well guess I will have to wait until after Navi 31 launch to read reviews to see RTX 4080 12GB and 16GB benchmarks before decide to pick MSI RTX 4080 Suprim X 12GB or Ti class 16GB.

I don't think anyone made the 4070 accusation based on what Nvidia puts in columns next each other.

 

[OG Don Mecca]

so in the US a load of AiB's are selling the low end models for the same as founders MSRP.......... whats the likleyhood of that happening here too?

 

[nomadd]

am I allowed to think that $1599 is a fair price or it has now become obligatory to join "I hate Nvidia" herd?

It depends..

..on what AMD does next.

 

[Gibbo]

so in the US a load of AiB's are selling the low end models for the same as founders MSRP.......... whats the likleyhood of that happening here too?

This is correct but said models will be shipping in extremely small numbers so wait times will be weeks and potentially months.

The majority and volumes of stock will be on more expensive OC models as that is where the AIBs have better margins.

 

[Purgatory]

Ah interesting.

RTX 4080 12GB and 16GB was confused before the whitepaper published. Everybody thought RTX 4080 12GB was changed from RTX 4070 12GB at last min but the whitepaper finally made it clear that RTX 4080 12GB will replace RTX 3080 10/12GB, not RTX 3070 and RTX 4080 16GB is actually Ti class will replace RTX 3080 Ti, not RTX 3080 10/12GB. I thought RTX 4080 Ti will launch later in early 2023 to replace RTX 3080 Ti.

Well guess I will have to wait until after Navi 31 launch to read reviews to see RTX 4080 12GB and 16GB benchmarks before decide to pick MSI RTX 4080 Suprim X 12GB or Ti class 16GB.

The 4080 ti will be on a AD102 chip with 16GB or 20GB VRAM (but expecting it to be 16GB and not 20GB), the AD103 chip that the 16GB 4080 is on will be the 4080ti laptop chip as is the case with the 3080ti laptop chip is a GA103 with 16GB too.


Don't be fooled by the naming, the same was meant to happen on Ampere but AMD messed that up for Nvidia.

The real Ampere should have been :-

AD102 Full die Titan AP and A series cards (Quadros of the past).
AD102 Slightly cut down again a Titan A and A series cards.
AD102 3080 Ti More cut down than the Titan A and half the VRAM and A series cards.

AD103 3080 desktop and 3080ti laptop.

AD104 3070 and 3060's and 3080,3070,3060s laptop.


The 3090 was never meant to exist, all 90 class cards before were dual chip gpus on a single card.

So what the 3090 showed Nvidia is people are willing to buy this type of class card that was normally a Titan before without the Titan drivers, so now they have a new product range the single chip 90 class cards for the people that want the best without titan drivers, but problem is the 4090 is a neutered 3090 this time as no NVLINK the big selling point to pros is now gone that wanted dual gpu setups with 48GB VRAM and double the performance.

Nvidia learned a lot from the 30 class series and have made sure not to make the same mistakes, also by them doing certain things to the 4090 says to me AMD don't have anything as good as 6000 series this time compared to 30 series before, so Nvidia is taking a bet they can get away with the 40 series range that they can sell less for more with the usual up tick in performance.

4080 12GB is looking like a 3080 12GB to me with a new name and not much performance difference and DLSS 3 support, 16GB 4080 and 4090 are in a different league from their previous cards as they should be with a proper up tick in performance and new features but also remember some features removed from 4090 compared to 3090 too. So you win some and lose some with the 4090.

 

[Colonel_Klinck]

When can we expect performance reviews for the 4090? Anyone know when the embargo lifts?

 

[JediFragger]

When can we expect performance reviews for the 4090? Anyone know when the embargo lifts?

Day before, the 11th.

 

[Colonel_Klinck]

Day before, the 11th.

Thanks

 

[Bill Turnip]

Embargoes are a little mad when you think of it. I get the point, but really there are a whole load of press releases that run along the lines of "We have a new product! It's brilliant! It'll revolutionise your experience! This is definitely a game changer! It is safe to upgrade, the more you buy the more you save!"

"Oh cool, can we see it perform?"

"....No."

 

[Chuk_Chuk]

So what the 3090 showed Nvidia is people are willing to buy this type of class card that was normally a Titan before without the Titan drivers, so now they have a new product range the single chip 90 class cards for the people that want the best without titan drivers, but problem is the 4090 is a neutered 3090 this time as no NVLINK the big selling point to pros is now gone that wanted dual gpu setups with 48GB VRAM and double the performance.

Nvlink has been removed from the A series Ada cards as well. It seems like Nvidia may try and do something through PCIE 5

 

[Purgatory]

Nvlink has been removed from the A series Ada cards as well. It seems like Nvidia may try and do something through PCIE 5

I doubt it as the previous NVLINK was faster and had more bandwidth than what they could do on PCIE 5, but we will see , so far what I understand is Nvidia has removed all the silicon on the Ada chips related to NVLINK and only on Hopper chips this gen. Guessing we will see it on a Hopper version A series soon enough so they can add another range of cards with different features and higher prices, their way to knock up the prices of a feature we took for granted and now becoming a luxury feature for top end cards from now. Nvidia silly games again from the gamer market to the pro market.

 

[Chuk_Chuk]

I doubt it as the previous NVLINK was faster and had more bandwidth than what they could do on PCIE 5, but we will see , so far what I understand is Nvidia has removed all the silicon on the Ada chips related to NVLINK and only on Hopper chips this gen. Guessing we will see it on a Hopper version A series soon enough so they can add another range of cards with different features and higher prices, their way to knock up the prices of a feature we took for granted and now becoming a luxury feature for top end cards from now. Nvidia silly games again from the gamer market to the pro market.

So the quote I saw on another forums were. Unfortunately the person who posted it didn't give a source.

Jensen stated that the reason behind removing the NVLink connector was because they needed the I/O for “something else,”

Jen-Hsun continued with “and also, because Ada is based on Gen 5, PCIe Gen 5, we now have the ability to do peer-to-peer cross-Gen 5 that’s sufficiently fast that it was a better tradeoff”

 

[AthlonXP1800]

I don't think anyone made the 4070 accusation based on what Nvidia puts in columns next each other.

You obviously not read many posts on this forum and other forums, websites, reddit or watching youtube etc.

Yes people wrongfully made overpriced $899 4070 accusations.

Here is Why the NVIDIA RTX 4080 12GB is an Overpriced RTX 4070 Ti | Hardware Times

NVIDIA announced its RTX 40 series GPUs earlier this month to a moderate to lukewarm welcome. The RTX 4090 looks like a proper flagship with the RTX 4080 16GB holding its own, but its 12GB sibling is a bit of an abomination. Featuring the AD103 die, it packs 7,680 shaders (less than the RTX 3080 …

www.hardwaretimes.com

Reddit - Dive into anything

www.reddit.com

Reddit - Dive into anything

www.reddit.com

Why the RTX 4080 12GB feels a lot like a rebranded RTX 4070 | Digital Trends

Nvidia has announced two versions of the RTX 4080. But the more we find out about these two GPUs, the more different they seem.

www.digitaltrends.com

 

[Joxeon]

Ah interesting.

RTX 4080 12GB and 16GB was confused before the whitepaper published. Everybody thought RTX 4080 12GB was changed from RTX 4070 12GB at last min but the whitepaper finally made it clear that RTX 4080 12GB will replace RTX 3080 10/12GB, not RTX 3070 and RTX 4080 16GB is actually Ti class will replace RTX 3080 Ti, not RTX 3080 10/12GB. I thought RTX 4080 Ti will launch later in early 2023 to replace RTX 3080 Ti.

Well guess I will have to wait until after Navi 31 launch to read reviews to see RTX 4080 12GB and 16GB benchmarks before decide to pick MSI RTX 4080 Suprim X 12GB or Ti class 16GB.

Just look at the die sizes, neither are 80 class, ones a 60ti the others a 70. If you plan to buy something then get the 4090 as it'll be 40% faster than even the more expensive 4080 16gb which in turn will be around 25% faster than your 3080ti, the 4080 12gb version will be no faster than the 3080ti.

 

[gpuerrilla]

Just look at the die sizes, neither are 80 class, ones a 60ti the others a 70. If you plan to buy something then get the 4090 as it'll be 40% faster than even the more expensive 4080 16gb which in turn will be around 25% faster than your 3080ti, the 4080 12gb version will be no faster than the 3080ti.

This was the scary part. If your high end from last gen your not even sidegrading. A 4080 should be faster than a 3080Ti.

 

[Purgatory]

Jen-Hsun continued with “and also, because Ada is based on Gen 5, PCIe Gen 5, we now have the ability to do peer-to-peer cross-Gen 5 that’s sufficiently fast that it was a better tradeoff”

We know Ada is only using PCIe 4 on the slot so that doesn't make sense as you know then. Ada is PCIe 4 16 lanes only and PCIe 5 power connectors only. I think people are getting confused by all these PCIe numbers flying around the specs for Ada are already up and even states no SLI/NVLINK and PCIe Gen 4 16x. Nvidia removed it to make it a pro thing only for their top of the top cards and server gear. Even the A series has been neutered in the same way, only Hopper has NVLINK and they are not doing it with GEN 5 PCIe either but seems it's in the spec but not used for NVLINK just normal data over PCIE 5 and NVLINK 4.0 to link gpus. They are even using Nvlink switches now to do the linking.

NVLink - Wikipedia

en.wikipedia.org

NVLink 3.0	50 Gbit/s		~6.25 GB/s		Ampere, NVSwitch for Ampere
NVLink 4.0	50 Gbit/s		~6.25 GB/s		Hopper, Nvidia Grace Datacenter/Server CPU NVSwitch for Hopper

PCIe 5.0

32 GT/s[7]

128b/130b

~4 GB/s

Hopper

Performance​

The following table shows a basic metrics comparison based upon standard specifications:

Interconnect	Transfer rate	Line code	Effective payload rate per lane per direction	Max total lane length (PCIe: incl. 5" for PCBs)	Realized in design
PCIe 1.x	2.5 GT/s	8b/10b	~0.25 GB/s	20" = ~51 cm
PCIe 2.x	5 GT/s	8b/10b	~0.5 GB/s	20" = ~51 cm
PCIe 3.x	8 GT/s	128b/130b	~1 GB/s	20" = ~51 cm[6]	Pascal, Volta, Turing
PCIe 4.0	16 GT/s	128b/130b	~2 GB/s	8−12" = ~20−30 cm[6]	Volta on Xavier (8x, 4x, 1x), Ampere, Power 9
PCIe 5.0	32 GT/s[7]	128b/130b	~4 GB/s		Hopper
PCIe 6.0	64 GT/s	128b/130b	~8 GB/s
NVLink 1.0	20 Gbit/s		~2.5 GB/s		Pascal, Power 8+
NVLink 2.0	25 Gbit/s		~3.125 GB/s		Volta, NVSwitch for Volta Power 9
NVLink 3.0	50 Gbit/s		~6.25 GB/s		Ampere, NVSwitch for Ampere
NVLink 4.0	50 Gbit/s		~6.25 GB/s		Hopper, Nvidia Grace Datacenter/Server CPU NVSwitch for Hopper

The following table shows a comparison of relevant bus parameters for real world semiconductors that all offer NVLink as one of their options:

Semiconductor	Board/bus delivery variant	Interconnect	Transmission technology rate (per lane)	Lanes per sub-link (out + in)	Sub-link data rate (per data direction)	Sub-link or unit count	Total data rate (out + in)	Total lanes (out + in)	Total data rate (out + in)
Nvidia GP100	P100 SXM,[8] P100 PCI-E[9]	PCIe 3.0	8 GT/s	16 + 16 Ⓑ	128 Gbit/s = 16 GByte/s	1	16 + 16 GByte/s[10]	32 Ⓒ	32 GByte/s
Nvidia GV100	V100 SXM2,[11] V100 PCI-E[12]	PCIe 3.0	8 GT/s	16 + 16 Ⓑ	128 Gbit/s = 16 GByte/s	1	16 + 16 GByte/s	32 Ⓒ	32 GByte/s
Nvidia TU104	GeForce RTX 2080, Quadro RTX 5000	PCIe 3.0	8 GT/s	16 + 16 Ⓑ	128 Gbit/s = 16 GByte/s	1	16 + 16 GByte/s	32 Ⓒ	32 GByte/s
Nvidia TU102	GeForce RTX 2080 Ti, Quadro RTX 6000/8000	PCIe 3.0	8 GT/s	16 + 16 Ⓑ	128 Gbit/s = 16 GByte/s	1	16 + 16 GByte/s	32 Ⓒ	32 GByte/s
Nvidia Xavier[13]	(generic)	PCIe 4.0 Ⓓ 2 units: x8 (dual) 1 unit: x4 (dual) 3 units: x1[14][15]	16 GT/s	8 + 8 Ⓑ 4 + 4 Ⓑ 1 + 1	128 Gbit/s = 16 GByte/s 64 Gbit/s = 8 GByte/s 16 Gbit/s = 2 GByte/s	Ⓓ 2 1 3	Ⓓ 32 + 32 GByte/s 8 + 8 GByte/s 6 + 6 GByte/s	40 Ⓑ	80 GByte/s
IBM Power9[16]	(generic)	PCIe 4.0	16 GT/s	16 + 16 Ⓑ	256 Gbit/s = 32 GByte/s	3	96 + 96 GByte/s	96	192 GByte/s
Nvidia GA100[17][18] Nvidia GA102[19]	Ampere A100 (SXM4 & PCIe[20])	PCIe 4.0	16 GT/s	16 + 16 Ⓑ	256 Gbit/s = 32 GByte/s	1	32 + 32 GByte/s	32 Ⓒ	64 GByte/s
Nvidia GP100	P100 SXM, (not available with P100 PCI-E)[21]	NVLink 1.0	20 GT/s	8 + 8 Ⓐ	160 Gbit/s = 20 GByte/s	4	80 + 80 GByte/s	64	160 GByte/s
Nvidia Xavier	(generic)	NVLink 1.0[13]	20 GT/s[13]	8 + 8 Ⓐ	160 Gbit/s = 20 GByte/s[22]
IBM Power8+	(generic)	NVLink 1.0	20 GT/s	8 + 8 Ⓐ	160 Gbit/s = 20 GByte/s	4	80 + 80 GByte/s	64	160 GByte/s
Nvidia GV100	V100 SXM2[23] (not available with V100 PCI-E)	NVLink 2.0	25 GT/s	8 + 8 Ⓐ	200 Gbit/s = 25 GByte/s	6[24]	150 + 150 GByte/s	96	300 GByte/s
IBM Power9[25]	(generic)	NVLink 2.0 (BlueLink ports)	25 GT/s	8 + 8 Ⓐ	200 Gbit/s = 25 GByte/s	6	150 + 150 GByte/s	96	300 GByte/s
NVSwitch for Volta[26]	(generic) (fully connected 18x18 switch)	NVLink 2.0	25 GT/s	8 + 8 Ⓐ	200 Gbit/s = 25 GByte/s	2 * 8 + 2 = 18	450 + 450 GByte/s	288	900 GByte/s
Nvidia TU104	GeForce RTX 2080, Quadro RTX 5000[27]	NVLink 2.0	25 GT/s	8 + 8 Ⓐ	200 Gbit/s = 25 GByte/s	1	25 + 25 GByte/s	16	50 GByte/s
Nvidia TU102	GeForce RTX 2080 Ti, Quadro RTX 6000/8000[27]	NVLink 2.0	25 GT/s	8 + 8 Ⓐ	200 Gbit/s = 25 GByte/s	2	50 + 50 GByte/s	32	100 GByte/s
Nvidia GA100[17][18]	Ampere A100 (SXM4 & PCIe[20])	NVLink 3.0	50 GT/s	4 + 4 Ⓐ	200 Gbit/s = 25 GByte/s	12[28]	300 + 300 GByte/s	96	600 GByte/s
Nvidia GA102[19]	GeForce RTX 3090 Quadro RTX A6000	NVLink 3.0	28.125 GT/s	4 + 4 Ⓐ	112.5 Gbit/s = 14.0625 GByte/s	4	56.25 + 56.25 GByte/s	16	112.5 GByte/s
NVSwitch for Ampere[29]	(generic) (fully connected 18x18 switch)	NVLink 3.0	50 GT/s	8 + 8 Ⓐ	400 Gbit/s = 50 GByte/s	2 * 8 + 2 = 18	900 + 900 GByte/s	288	1800 GByte/s

 

[von Humboldt]

This explains the price - there were obviously billions spent on R&D for the move from a 9 blade fan to a 7.

Seriously though, are these mock-ups? The angles of the fan blades are different.

Anyone have an explanation for this? Not a big deal - I'm just curious as to why the blade orientation would have been reversed.