Nvidia GeForce RTX 4000 – Release date – Specifications and features in detail
Nvidia GeForce RTX 4000 – Release date – Specifications and features in detail
Next-generation Nvidia graphics cards can deliver up to twice the performance of the RTX 3000 series. And although there is still a lot of time left until the premiere of the new systems, we decided to collect everything we know so far.
Updated April 14, 2022
Everything indicates that kopite7kimi’s first entry in April was not a joke at all, and more and more sources confirm this.
The author, who is known for many leaks on Twitter , has presented the specifications of three models of the best GPUs based on Ada Lovalace cores .
April Fool’s Day
**90 PG139-SKU310, AD102-300, 24G 24Gbps, 600W;
** 80 PG139-SKU330, AD102-200, 20G 24Gbps, 500W;
** 70 PG141-SKU310, AD104-400, 12G 24Gbps, 400W.
???????????????????????????????????????????????? ???????????????????????????????????????????????? ?????????????????? ????????????????????????????????????????????????– kopite7kimi (@kopite7kimi) April 1, 2022
Unfortunately, previous reports that the new most powerful RTX will have a TGP 600W by default have been confirmed.
The card receives 24 GB of GDDR6x memory .
Nvidia will likely use the new Micron chips, which achieve speeds of 24Gb/s. If the AD102 cores continue to have a 384-bit data bus, the memory bus bandwidth of the reference RTX 4090 (or 5090 ) models will reach 1152 GB/s.
Also check:
It should be noted that even the lower model mentioned in the representative version of the 70 series ( RTX 4070 or RTX 5070 ) will also get GDDR6x modules .
However, we can assume that the AD104 cores will have a 256-bit bus , which translates to 768GB/s of bandwidth for 12GB of memory.
Updated April 11, 2022
It is possible that Nvidia is preparing to change the naming of cards based on the Ada Lovalace architecture. Although the information about kopite7kimi was revealed for the first time – a well-known leaker from Twitter, because it happened on April 1 , no one took it seriously.
April Fool’s Day
**90 PG139-SKU310, AD102-300, 24G 24Gbps, 600W;
** 80 PG139-SKU330, AD102-200, 20G 24Gbps, 500W;
** 70 PG141-SKU310, AD104-400, 12G 24Gbps, 400W.
???????????????????????????????????????????????? ???????????????????????????????????????????????? ?????????????????? ????????????????????????????????????????????????– kopite7kimi (@kopite7kimi) April 1, 2022
However, the author of the first Twitter post, harukaze5719, continues to claim that such a change is possible. This wouldn’t be the first time Nvidia has renumbered its desktop GPUs.
We can observe a similar situation with the release of the GTX 400 series cards (the GTX 300 series has been deleted), and the GTX 900 (the GTX 800, which was only available for mobile devices) has been deleted.
RTX 4000 Series or RTX 5000 Series?
Well whatever …
– Posi Posi (@harukaze5719) April 10, 2022
**90 PG137/139-SKU330, AD102-300, 24Gb/s 21Gbps, 600W
– kopite7kimi (@kopite7kimi) April 11, 2022
Updated March 28, 2022
IgorsLab recently shared new information about upcoming graphics cards based on Ada Lovelace cores .
Portal recently received a picture of a clean PCB. Taking care of its source material, IgorsLab has made a sketch of the new RTX design to corroborate the information.
Everything indicates that we’re dealing with a reference design that Nvidia plans to use in the RTX 4090 and RTX 4080 .
On the basis of the information gathered, it was possible to confirm previous reports about the demand for 600 watts of power. For owners of older power supplies not equipped with the new 12 + 4 (12VHPWR) power connector, a 4 x 6 + 2 (8-pin PCIe) adapter must be added in the kit.
The 24-phase section should be responsible for powering the kernel .
However, given that at the moment the uP9512 chip is responsible for the control, and is able to manage only 8 phases, it can be assumed that not all fields will be used, or Nvidia will use the phase redundancy option.
The power section of the VRAM also looks impressive, because we are dealing with a four-phase system .
According to the schematic diagram, there are 12 solder sockets for the GDDR6X modules. This means that we will again receive RTX variants with 12 or 24 GB of memory .
Let’s move on to cooling information. Some portal sources say that NVIDIA will rely on three-slot air cooling in reference models or FE models.
We have already seen such a solution, for example in the RTX 3090 FE. “Green” partners are more likely to lean toward constructions that require up to 3.5 holes.
This means that many new graphics cards can weigh more than 2 kg . Although we will see versions equipped with AIO liquid cooling systems or equipped with water cooling blocks, they will not be treated as a standard solution.
Updated March 23, 2022
The beta version of HWiNFO, a very popular tool for gathering and presenting information about PC hardware, has added support for unreleased and unconfirmed NVIDIA GPUs.
The latest HWiNFO release notes list some Ada Lovelace, Blackwell, and Hopper GPUs.
According to the information on the website, support for the following GPUs has been added – AD102, AD103, AD104, AD106 and AD107.
These five GPUs will be the basis for the upcoming RTX 4000 series graphics cards. Unfortunately, it is not possible to verify the details of the systems.
The base names listed are consistent with those that have been leaking for some time – including the most recent ones created after the hack attack by the $Lapsus group, which exposed confidential NVIDIA data.
It is not clear if the HWiNFO tool can actually detect and report these chips.
It is entirely possible that the creators of the popular tool have added the names to the program’s database as placeholders and intend to add full support for them in the near future.
Updated March 8, 2022
Once again, Ada Lovelace’s chip specs have been made available on Twitter. User Olrack29_ has collected all the information that came to light after the recent theft of data by hackers from NVIDIA servers. The most powerful chip for consumer cards is the AD102.
AD102
12 GPCs * 6 TPCs * 2SMs * 128FP32 = 18432 cores
384-bit bus
96MB L2 cacheAD103
7 GPCs * 6 TPCs * 2SMs * 128FP32 = 10752 cores
256-bit bus
64MB L2 cacheAD104
5 GPCs * 6 TPCs * 2SMs * 128FP32 = 7680 Cores
192 bit bus
48 MB L2 cache pic.twitter.com/tkGqDBEudj– Olrak (@ Olrak29_) March 8, 2022
It consists of twelve GPC computing clusters, each containing six TPC (Texture / Processor Cluster) blocks. This means that the kernel will consist of 144 SM blocks (2 in each TPC) and 18432 CUDA units (128 in one SM). In addition, the AD102 must have 96MB of L2 memory and communicate with GDDR6X modules using a 384-bit bus.
AD106
3 GPCs * 6 TPCs * 2SMs * 128FP32 = 4608 cores
128-bit bus
32MB L2 cacheAD107
3 GPCs * 4 TPCs * 2SMs * 128FP32 = 3702 cores
128-bit bus
32MB L2 cache pic.twitter.com/Zt7BNNr8r5– Olrak (@ Olrak29_) March 8, 2022
Olrack29 also shared unofficial graphical schematics of future NVIDIA Ada Lovelace systems.
Updated March 3, 2022
More information about the upcoming RTX 4000 systems has been revealed. Kopite7kimi shared a post on Twitter containing the configuration of Ada Lovalece’s cores and showing the communication bus with the card’s memory. This is another confirmation of previous reports.
I know these before,
AD102 12 * 6 384bit ~ 600mm²
Not my sources,
AD103 7 * 6 256bit
AD104 5 * 6 192bit
AD106 3 * 6 128bit
AD107 3 * 4 128bit
Celebrate the leak!
You don’t need me anymore. ????????????????????????????????????????????????— kopite7kimi (@kopite7kimi) March 2, 2022
The most efficient chips will continue to use the 384-bit interface. However, given yesterday’s information about the second-level cache increase, we can suspect that the memory bandwidth has been greatly increased.
Everything indicates that NVIDIA has developed an equivalent of AMD’s Infinity Cache.
Jensen lowered some small features like NVLINK or extra dp links to reduce pin usage for adding GPU power.
— kopite7kimi (@kopite7kimi) March 2, 2022
The author of the entry also claims that the cards are deprived of smaller functions, such as NVLINK or additional signal lines for DP connectors.
All this to provide more strength to the core itself. We would like to remind you that information has already appeared according to which Ada Lavalace systems will consume up to 600 Watts.
Updated March 2, 2022
Recently, NVIDIA was hacked, leaking more than 1 terabyte of confidential information to the network.
It’s currently being announced, which means we’re learning about the possibility of, for example, bypassing LHR technology or code names for future chip architectures.
According to the latest reports, NVIDIA is preparing at least six cores that belong to the Ada Lovelece family. These are the AD102, AD103, AD104, AD106, AD107 and AD10B chips .
The first five will be found in future designs of both our desktop and laptop computers, and new GeForce cards and RTX modules for workstations will be built on their basis.
The last mentioned core – AD10B according to Kopite7kimi will be used in the latest version of the Tegra chips , which drive consoles, among others Nintendo Switch.
Again, we also received confirmation of the Ada Lovalece modules configuration.
From Nvidia Leak:
102: 144 SM + 384-bit + 96MB L2
103: 84 SM + 256-bit + 64MB L2
104: 60 SM + 192-bit + 48MB L2 106
: 36 SM + 128-bit + 32MB L2107
: 24 SM + 128-bit + 32MB L2– Nobody (@no_one180) March 2, 2022
Updated February 21, 2022
Everything indicates that previous reports of an RTX 4000 price hike have been confirmed. Once again, Graymon55 shared a message on Twitter suggesting that the RTX 4090 could cost as much as $2000.
As a reminder, the MSRP for the RTX 3090 was $1,199 on the day of the announcement.
This means an increase of 60%. It can therefore be assumed that other models will also be offered at a much higher price.
I would say the 4090 is about $1500-2000.
— Greymon55 (@greymon55) February 20, 2022
Fortunately, the author of the entry also shared good news. According to his data, next month we should see significant price cuts for current-generation graphics cards
It looks like GPU prices will drop dramatically next month. I hope this is true. ????
— Greymon55 (@greymon55) February 20, 2022
Updated February 14, 2022
Once again, Graymon55 has shared a new piece of information about future GPU models.
This time we received a list of Ada Lovelace chips along with which graphics card models they would go with.
Here are my guesses about the next generation chips and models.
AD102 (RTX 4080-90 series) 5nm
AD103 (RTX 4070series) 5nm
AD104 (RTX 4070series) 5nm AD106
(RTX 4060series) 5nm
N31 (RX 7900series) 5nm
N32 (RX 7800series) 5nm N33 (RX 7700s23
)
6nm 6nm 6nm– Greymon55 (@greymon55) February 13, 2022
According to previous leaks, the most powerful AD102 chip will be found in the RTX 4090 and 4080. Graymonn55 also believes that all chips will be made in TSMC’s 5nm manufacturing process.
Some Twitter users are of the opinion that due to the limited production capacity of Taiwanese facilities, Nvidia may decide to split production and use Samsung facilities that are currently producing Ampere chips.
Updated 02/08/2022
According to 3DCenter , NVIDIA has an opportunity to improve chip scaling in the upcoming Ada Lovalace architecture to get the most efficient chips.
According to Twitter, the AD102 core is expected to deliver 71% more performance than the GA102 based on the Ampere architecture. It is noteworthy that the comparative core of Ada Lovalace contains 12 GPC units , which translates to 144 SM blocks.
This means that the internal design of the units has not changed and we are still talking about a model equipped with 18,432 CUDA units .
Lovelace gives nVidia the opportunity to improve the poor amp gradient in high-end solutions:
GA104 -> 6 GPC, 48 SM
GA102 -> 7 GPC, 84 SM (+17%, +75%)AD102 -> 12 GPC, 144 SM
+ 100%, + 200% vs GA104
+ 71%, + 71% vs GA102– 3DCenter.org (@3DCenter_org) February 6, 2022
Again, Greymonn55 has a slightly different opinion. According to him, there were no major changes in the Lavalas architecture, and the difference in performance is due only to the larger number of computing units and the higher frequency of the system.
Lovelace’s architecture doesn’t change much.
— Greymon55 (@greymon55) February 5, 2022
Updated 01/28/2022
According to information shared again by Greymonn55 in a post on Twitter, everything indicates that the premiere of the new systems is not delayed in any way.
This means we should see new RTX 4000 series models in the fall of 2022.
If the lovelace goes to 2.5GHz, you’ll get about 90T from the FP32, and the 3090T only 40T. ????
— Greymon55 (@greymon55) January 25, 2022
In addition, the same user shared information that Ada Lovelace’s cores are capable of reaching a frequency of up to 2.5 GHz.
As a result, they achieve performance in FP32 calculations of up to 90 TFLOPs.
For comparison, the soon-to-be-launched RTX 3090 Ti is getting “only” 40 TFLOPs.
Updated November 10, 2021
Another entry about Ada Lovelce’s layouts appeared on the account of well-known leaker Greymon55 .
According to the latest reports, the card based on the AD102 core is expected to achieve performance in the range of 85-92 TFLOPs in single resolution (FP32) calculations.
All thanks to 18,432 CUDA cores that run at over 2.3GHz .
AD102
5nm TSMC
144SM 18432CUDA
384bit 24G GDDR6X 21Gbps?
2.3 ~ 2.5GHz?
85 ~ 92T FP32?– Greymon55 (@greymon55) November 9, 2021
However, according to kopite7kimi , we should see the RTX 4090, RTX 4080, and RTX 4070 systems released in the third quarter of 2022.
If nothing else, we’ll see the 4090, 4080, and 4070 in 2022Q3. I don’t think this is new information.
– kopite7kimi (@kopite7kimi) November 8, 2021
Updated September 22, 2021
Once again, we’ve received new information from leaks on Twitter.
User creper9000 decided to collect all of Ada Lovelce’s upcoming chip data into a table. Shows the assumed specifications of the cores of different generations and their performance compared to Ampere chips.
And while the entire content of the table is largely speculative, another well-known leaker, kopite7kimi, commented .
According to him, we can be sure of the cores of 103m , but we won’t see 107m . This means that, as with the current generation, we won’t be getting the lower-end series of new GPUs.
So far, Nvidia has preferred to reintroduce the 12nm RTX 2060 by bumping up its memory to 12GB of memory rather than creating a new lithograph chip that could be called the RTX 3050 .
AD 103 –
107 AD– kopite7kimi (@kopite7kimi) September 21, 2021
Updated September 20, 2021
Greymon55 once again shared the new rumors on Twitter .
According to new information, the AD102 chip will be equipped with a 384-bit GDDR6x memory bus . This means that Nvidia isn’t planning any major changes here compared to the Ampere cards.
AD102 Frequency:2.2GHz
– Greymon55 (@greymon55) September 19, 2021
In addition, Greymonn put information that the card’s clock will be more than 2.2 GHz .
Given that the current RTX 3000 chipset operates at 1800MHz , it can be assumed that the new rumors also apply to this situation. This would be a huge advance for Nvidia, similar to what AMD achieved with RDNA2 .
An interesting note was shared on Twitter by 3DCenter.org.
According to their calculations, an AD102 core equipped with 144 SMs running at 2.2 GHz would achieve a performance of 81 TFlops in FP32 (single precision) calculations.
This is a 122% increase over the score achieved by the GA102 units.
At 2.2GHz, nVidia’s AD102 will hit 81 raw power FP32 TFlops (with 144 SM fully activated).
With an SM increase of 71% over the GA102, the overall jump in raw power between the GA102 and AD102 would be +122% (not more than 22%, it’s a staggering 122% increase). https://t.co/cCuCeESrZn
– 3DCenter.org (@3DCenter_org) September 19, 2021
Updated 09/08/2021
More information has been shared about the RTX 4000 series, the successor to the Ampere codenamed Lovelace.
Unlike AMD’s RDNA3, the architecture does not use a multi-chip module (MCM) design and will be built with a single core .
According to Greymon55, NVIDIA plans to release new graphics cards in October 2022, but it hasn’t been specified if this is a planned release date or just a presentation date.
The next generation gpu will be in October next year, about a year from now.
– Greymon55 (@greymon55) September 7, 2021
Previously, it was also confirmed that the new chips will be manufactured at TSMC’s facilities using 5nm lithography.
RTX 4000
Nvidia’s GeForce RTX 3000 series offered a huge jump in performance compared to the RTX 2000 series. Rumors suggest that we can expect a similar increase for the RTX 4000 over the current generation.
However, we must remember that we have not received any official information from the manufacturer, and the entire production process is kept secret.
Although we are far from a release date, all rumors indicate that the new graphics cards will be produced in 5nm lithography , and the core will be based on the Ada Lovelce architecture .
This is to ensure a twofold increase in efficiency.
When will the RTX 4000 be released?
We can be sure of one thing, we can’t expect it to premiere in 2021 .
Typically, Nvidia decides to keep a two-year break between series. This is somewhat confirmed by Greymon55 – a frequent source of leaks and rumors on Twitter.
According to him, the cards will not come out before the end of 2022 , although he also mentioned the possibility of postponing this date until the first quarter of 2023 .
Which is logical, given that the premiere of Grace server systems, which will also be produced on the 5nm lithography process, is scheduled for 2023.
In recent years, the first new solutions for professional applications were presented, and only then those intended for the retail market.
However, this does not mean that we will not get new graphics cards from Nvidia by the end of 2022.
However, after all, successors to popular cheap systems like the GTX 1050 or GTX 1650 were not introduced.
And as usual from previous generations, we can also expect more updated and improved versions of the current generation RTX 3000. Everything to satisfy gamers and fill the market before the premiere of the new series.
The release of the RTX 4000 in the first quarter of 2023 could be good for all of us .
It remains to be hoped that Nvidia and all other manufacturers will have enough time to overcome the current shortage of components, which is causing production chaos throughout the technology industry .
How much does the Nvidia RTX 4000 series cost?
There is still a lot of time before the premiere and it is difficult to predict the cost of new graphics cards. Assuming an optimistic scenario and that Encore systems will replace their RTX 3000 series counterparts, we can expect the manufacturer’s suggested prices to be repeated in the case of the Ampere family:
- Nvidia GeForce RTX 3060 – $329 / $1,599
- Nvidia GeForce RTX 3060Ti – $ 399 / PLN 1899
- Nvidia GeForce RTX 3070 – $499 / $2,359
- Nvidia GeForce RTX 3070Ti – $ 599 / PLN 2,800
- Nvidia GeForce RTX 3080 – $699 / 3279 AU
- Nvidia GeForce RTX 3080Ti – 1199 USD / 5600 PLN
- Nvidia GeForce RTX 3090 – $1,499 / PLN 7,029
Starting with the RTX 3060, which should start at $329, Nvidia is noticeably cheaper than the competing Radeon RX 6000 series.
Hopefully, this trend will continue in the next generation despite the performance jump. In any case, we saw a similar performance jump in the premiere of Ampere series cards, and these cards “though in theory” are still competitively priced.
The problem that most of us are aware of is the lack of inventory, which leads to the price of graphics cards going up.
Retailers began to dramatically inflate prices to unprecedented levels.
It is worth noting that the price of the RTX 3070 should be $ 499 (2,359 PLN), as it is often available in stores for more than twice that amount. Worst of all, people are actually buying chips at these prices.
Nvidia could see that gamers were willing to pay a lot more than the suggested price .
This may prompt the manufacturer to raise the price of the entire future RTX 4000 series.
Besides, similar rumors are reaching us from the competition regarding the Radeon RX 7000 .
Let’s hope that doesn’t happen. I believe that after the situation has calmed down with a shortage in the market, we all need to return to “normal life”.
What do you expect from the RTX 4000?
As mentioned earlier, according to Greymon55 , the new cards will be based on Ada Lovelace’s Monolithic Cores.
It will be manufactured using 5nm lithography at TSMC’s facilities.
The custom at Nvidia is to name the following exact architectures after famous researchers in the field of mathematics or physics.
This time it was a 19th-century British mathematician, known primarily for her publications on Charles Babbage’s Mechanical Computer.
The work contains the first published algorithm written with the intent of running it on an analytical machine.
improve the performance
Despite the huge performance gains offered by the RTX 3000 series over the 2000 series, the RTX 40 series is expected to deliver similar performance.
The rumor comes from another Twitter user this time around, TtLexington . This confirms previous rumors of 66.4 TFLOPs for the Lovelace AD102-based chip (38.7 TFLOPs for the GA102).
These numbers may also indicate an increase in the number of SMs to 144 , so that the number of shaders ( CUDA cores ) could reach 18,432 . Over 75% off the RTX 3090 $1499.
The same source says that cards from the ADA family will have a much higher TDP of around 450W , and even as low as 500W. It is not known if this rumor is about the GA102 complete system for professional applications or the system intended for consumer solutions.
However, this change means that we will again see such an increase.
This power consumption will certainly affect the amount of heat released from the system. Are we facing another revolution in cooling systems?
There are also rumors that the new RTX 4000 series may offer clock speeds between 2.2GHz and even 2.5GHz in Boost mode .
This would be a nice increase over what Ampere cards are currently achieving, which range between 1.7 – 1.9GHz. Comparing the 2.5GHz achieved by competing AMD systems based on the RDNA2 architecture, it might seem like not much.
However, it should be remembered that with a powerful processor, the clock does not play such a big role, which is why AMD tempted ZEN2 and ZEN3 processors.
However, Ada-series cards will compete with RDNA3, let both platforms compete with each other, and it will give us at least a shadow of a chance for lower prices.
to summarise:
- We expect the premiere at the end of 2022
- We expect a twofold increase in productivity
- We assume that the TPD will increase to 500 watts
If you’re looking for information on where to buy cards in the Amper family, we’ve put together a list of stores with RTX 3090, RTX 3080 TI, and RTX 3070 available.
.