332
119 shares, 332 points

We expect the Ada Lovelace architecture and GeForce RTX 40 series display cards to arrive by the end of the year, about which we have written and drawn a lot. NVIDIA will likely make its official announcement around September-October. We’ve known for a long time that the company has been introducing new architecture every two years, and that year has come.

The current Ampere architecture and the RTX 30 series display cards received great attention. The green team has released new GPU models from the top generation down over time. While the demand for hardware has increased considerably during the pandemic period, cryptocurrency mining has changed the color of things. In this process, in addition to all the hardware, especially the price of the display cards has increased incredibly. Fortunately, fiyadollarsar has recently been on the way to normalization and stocks are getting better day by day.

There have been a lot of rumors about NVIDIA’s Ada architecture lately. NVIDIA does not give a secret about this as usual. The company recently launched its Hopper architecture and its H100 GPU for data centers. The new Ada architecture will be used in the GeForce series kardollarsars, which will reach the hands of the players. Now let’s put together everything we know and expect about the new architecture and the upcoming onscreen cards.

Let’s talk about foundries first. We are almost certain that NVIDIA will be working with TSMC, but there is no official confirmation on this yet. Based on the island architecture, the GPUs will use either the Taiwanese giant’s 5nm (N5) or 4nm (N4) manufacturing technologies. It has been mentioned many times that 5nm production will be used so far, but there are some sources that think the company will adopt the 4nm process.

Hopper GPUs are built using a proprietary TSMC 4N fabrication technology. This shouldn’t be compared to the generic N4 (4nm) process that TSMC offers, and it uses a proprietary production fabrication technology for NVIDIA. As a result, there is a possibility that the company will adopt the same technique in its GeForce GPUs.

NVIDIA’s Hopper H100 hints at where the company could go with its Ada architecture. Transistor counts are currently the best estimation method. The A100 GPU in the past contained 56 billion transistors. The new generation Hopper H100 is home to 80 billion transistors.

The company had doubled the number of transistors with its current GA102 GPU. Let’s also note that this GPU is used in the RTX 3080 and RTX 3090 series. We’ve heard that the upcoming AD102 GPU will have a fairly large size. So we expect to see significant increases in the number of transistors and the amount of CUDA cores.

Judging by the leaked information so far, the flagship “AD102” GPU will be a beast. Again, we will see this silicon in high-end models such as the RTX 4080 and RTX 4090. Many more SMs and cores will be offered than current Ampere GPUs. Ultimately, there will be significant, satisfying increases in performance.

For example, the RTX 3080 was around 30% faster than the RTX 2080 Ti at launch. The RTX 3090, on the other hand, added another 15% performance gain when the GPU pushed its limits.

Let’s talk about the CPU side. We need powerful processors to get a good performance from the current Ampere display cards. Those who want to build a system with the RTX 4000 series display cards, based on the upcoming Ada architecture, will need very powerful chips.

GPU AD102 AD103 AD104 AD106 AD107
Production TSMC 4N? TSMC 4N? TSMC 4N? TSMC 4N? TSMC 4N?
Number of Transistors 60B? 40B? 30B? 20B? 15B?
SM / CU 144 84 60 36 24
CUDA
Czech.
18432 10752 7680 4608 3072
Tensor Check. 576 336 240 144 96
RT Czech. 144 84 60 36 24
Boost Frequency 1600-2000
MHz
1600-2000
MHz
1600-2000
MHz
1600-2000
MHz
1600-2000
MHz
Total L2 Cache 96MB 64MB 48MB 32MB 32MB
VRAM Speed ​​ 21-24 Gbps 21 -24 Gbps 16-21 Gbps 16-21 Gbps 14-21 Gbps
Bandwidth 384 256 192 128 128
ROP 128-196 ? 112? 96? 64? 48?
TMU 576 336 240 144 96
TFLOPS FP32 59-73.7 34.4-43 24.6-30.7 14.7-18.4 9.8-12.3
TFLOPS FP16 (Tensor) 472-590 275-344 197- 246 118-147 79-98
Band
Width
1008-1152GBps 672-768GBps 384-504GBps 256-336GBps 224-336GBps
TDP <600W <450W <300W <225W <150W

Island GPUs The most notable change regarding the SM is the number of SMs compared to the current Ampere generation. If we look at the top model, the AD102 could potentially carry 71% more SM than the GA102. Even assuming that there are no significant improvements in the architecture, a huge increase in performance is likely.

This will be true for other components as well, not just graphics. Tensor cores provide functionality in many applications. A full AD102 GPU die running at close to 2 GHz can have up to 590 TFLOPS of deep learning/AI computing power in FP16. The GA102 in the RTX 3090 Ti delivers approximately 321 TFLOPS of FP16 performance. This difference may vary according to core counts and clock speeds, but a theoretical 84% increase is mentioned. Also, similar performance gains should apply on the ray tracing side.

It can, however, implement a new generation of NVIDIA RT and Tensor cores. With the Ampere architecture, performance per RT core was increased by about 75% compared to Turing. The truth is, we don’t expect such big differences, but a 25-50% performance increase can be made in the RT and Tensor cores.

GPUs based on Ampere architecture, as you know, go through Samsung Foundry’s 8N process. Even switching to one of NVIDIA’s TSMC 4N (or 5N) technologies is enough on its own. Thus, more cores can be added to the graphics processors and their frequency speeds can increase significantly. In the end, it is not difficult to achieve performance increases that will satisfy us and be sufficient for years.

You can see some SM and CUDA values ​​in the table above. However, NVIDIA may disable some of the chips to increase efficiency. For example, the Hopper H100 has 144 potential SMs. However, only 132 SMs are enabled on the SXM5 model, while 114 SMs are enabled on the PCIe 5.0 card.

We are not sure of the ROP values ​​on Ada GPUs as we don’t know how it is configured yet. With Ampere, NVIDIA connected ROPs to GPCs, namely Graphics Processing Clusters. Each GPC contains a certain number of SMs (Streaming Multiprocessors) that can be disabled in pairs. Even if we know the number of SMs, we do not know how they are divided into GPCs. Let’s take the AD102 with

144 SM as an example. This GPU can be built with 12 GPCs of 12 SMs each, 8 GPCs of 18 SMs each, or 9 GPCs of 16 SMs each. There are other possibilities, but we’ll see what the green team’s engineers choose over time.

Hopper H100, which will be used in data centers, has 8 GPC clusters, each consisting of 18 SMs. This seems like a reasonable configuration for the AD102 as well. Not only will HBM3 memories be used, and the focus will be on games, not deep learning.

On the other hand, rumors stating that AD102 will be a huge chip may be true. NVIDIA does not manufacture separate silicon for the consumer and professional markets, as is believed with A-series chips like the RTX A6000. The RTX A6000 uses the same GA102 chip as the RTX 3080 and 3090 Ti. In addition, a few extra features are open only in drivers. Ray tracing is essential for the professional market as well as the gaming world. Therefore, using more RT cores will be a boon for 3D rendering professionals.

Various Ada GPUs will also be used for inference platforms running AI (artificial intelligence) and ML (machine learning) algorithms, again meaning more Tensor cores and computation can be used. As a result, NVIDIA will measure everything thoroughly before determining the amount of SM it will use in its GPUs. In the flagship model, we can see that 144 SMs are fully active, and finally 18432 CUDA cores together.

Micron recently announced that it is working on the production of GDDR6X memory operating at speeds of up to 24 Gbps. The RTX 3090 Ti, which is currently the most powerful card, uses memory that can reach 21 Gbps. NVIDIA is also the only company currently using GDDR6X. This immediately raises the question of when to use 24 Gbps GDDR6X memory.

Micron will likely start producing 24Gbps memory for the RTX 40 series soon. Lower-end GPUs are more likely to stick with GDDR6 rather than GDDR6X. In GDDR6 memory, the peak is at 18 Gbps.

Graphics processors need proportionally scaled bandwidth and computing power to deliver the promised amount of performance. For example, the RTX 3090 Ti offers 12% more processing power than the 3090. Higher clocked memory can provide 8% more bandwidth. A hypothetical RTX 4090 may have around 80% more computing power than the RTX 3090 Ti. However, when it comes to memory, we can only estimate that it will offer 14% more bandwidth.

A great way to reduce memory bandwidth needs, it’s based on a technique that has been known and used for decades. If you add more cache to a chip, the GPU accesses data faster, which has a positive impact on performance. In other words, the GPU at the heart of the graphics card has to draw less data from the GDDR6/GDDR6X memories.

AMD’s Infinity Cache technology allowed RDNA 2 chips to basically do more with less raw bandwidth. Some leaked information shows that NVIDIA will follow a similar approach, significantly increasing the amount of cache.

AMD uses high L3 cache solutions, 128MB on Navi 21 GPU, 96MB on Navi 22, 32MB on Navi 23 and 16MB on Navi 24. Surprisingly, even the smaller 16MB cache works wonders for the memory subsystem. The Radeon RX 6500 XT fell short of the class overall, but it basically keeps up with the kardollars with almost twice the memory bandwidth.

NVIDIA RTX 40 GPUs will be equipped with up to 96MB of L2 cache, up to 90MB more compared to the Ampere-based RTX 30 series. Incoming data confirms that “Ada” GPUs will have 16MB of cache per 64-bit memory bus, with 512KB per 32-bit memory, significant increases over the Ampere architecture.

According to this information, the flagship Ada Lovelace GPU, codenamed AD102, will have 96 MB cache. The AD103 and AD104 will be supported with 64MB cache, while the AD106 and AD107 will be supported with 48MB and 32MB cache respectively.

Now let’s get to the thing you might not like: power consumption. Igor from Igor’s Lab said that the RTX 4000 series display cards will set a record with 600W TBP (Typical Card Power) levels. This sounds surprising at first glance, but don’t be surprised if you see total power consumption in excess of 600W with specially cooled models. Igor also supports these claims with some arguments.

NVIDIA graphics cardsyellow have hovered near 250W for many years. We’ve seen a quick jump to 350W TDP on the RTX 3090 graphics card with Ampere. The company also gave some signals while announcing the Hopper H100 features.

Despite a smaller production technology, the TDP of the H100 in the SXM variant has been increased to 700W compared to 400W for the A100 SXM modules. That means 75% more power for upgrades that range from 50% to 500% depending on workload. Overall, performance is expected to be two to three times faster than the NVIDIA A100, but power consumption has increased considerably.

After the technology giant, the RTX 3090 Ti graphics card was released. The GeForce RTX 3090 Ti reference model comes with a 450W limit, taking consumer-grade power consumption to another dimension. In certain special cooled display cards, this value can reach up to 550W.

Long story short, it is highly likely that we will see 600W and even higher levels. The reference cardollarsar’s maximum power handling will likely be 600W, but there are already rumors that some next-gen third-party overclocked kardollarsars may come with dual 16-pin power connectors.

It was said that the “Ada Lovelace” architecture, which will be released at the end of this year, will come with a PCIe Gen5 interface. However, it seems that this will not be the case. Judging by some reliable sources, the next generation AD102 GPU will rely on PCIe Gen4 connectivity. Frankly, this is not a shortcoming, because the PCIe 4.0 connection is sufficient for the display cards and it will not cause any bottlenecks.

The RTX 40 series will most likely come with a newly popular 12+4-pin (12VHPWR) power connector. We have mentioned all the details of the Intel ATX12V v3.0 standard in this article.

ATX 3.0/PCIe 5.0 power connector will be able to power an external graphics card from 150W to 600W. Power supplies conforming to the ATX 3.0 specification will continue to be released throughout 2022. The new standard includes a new 12-pin 12VHPWR connector that can deliver up to 600W of power. In addition, the new connector is also called PCIe 5.0 in the industry. At this point, we can say that the new standard can make cable management easier because instead of several 8-pin power connectors, a single 12/16 pin is sufficient.

There is no definite information about the nomenclature of the next generation NVIDIA GPUs. We always use the nomenclature “RTX 40”, following the path established by the last few generations. The name “RTX” is certain to remain the same, but we may see a change in the naming of the series, although it is unlikely.

To give a succinct answer, the RTX 40 series will be quite expensive. NVIDIA launched the Ampere family with a series of financial models, and these decisions were deemed to be completely wrong for the time of the pandemic. Actual sales prices skyrocketed and black marketers continued to hoard.

Cryptocurrency miners did not hesitate to pay several times the suggested list price for graphic cardsyellow. Currently, GPU prices are high, but the good news is that GPU prices are falling.

The price of chips will increase with the RTX 40 series, probably due to the increase in the price of the chip, the production cost of the series and many reasons. The RTX 3090 Ti debuted at a high price of $ 1999. Maybe NVIDIA will resurrect the Titan brand for the Ada architecture.

On the other hand, the company is not expected to shift all GPU production directly to Ada architecture. We will see RTX 30 series GPUs being produced for quite some time, especially as there are no other GPUs or CPUs competing for Samsung Foundry’s 8N production. NVIDIA will continue to lead the market by utilizing all the available capacity it can get from TSMC and, if necessary, lowering the price of the RTX 30.

The company’s CFO, Colette Kress, made some statements at the Technology, Media and Telecom Conference. Kress said the COVID-19 outbreak and the supply chain problems it has exacerbated have given the company the opportunity to sell both the RTX 30 and RTX 20 series cardsollars. As a matter of fact, it seems likely that the RTX 30 series Ampere display cards will remain on the market for a long time.

We are doing this to provide more supply to our players and we may see something like this going on in the future. It’s been successful with Ampere and we’ll see as we progress.

We always point to September for the launch of the RTX 40 series “Ada” GPUs. But the first Ada kardollars will be just the tip of the yellow iceberg. NVIDIA released the RTX 3080 and RTX 3090 models in September 2020. The RTX 3070 arrived a month late, then the RTX 3060 Ti was introduced again a month later. The RTX 3060 didn’t come out until late February 2021. The GPU manufacturer then refreshed the series with the RTX 3080 Ti and RTX 3070 Ti in June 2021. We’ve been waiting for the budget-friendly RTX 3050 for a long time, and it only hit the market in January 2022. Finally, the flagship changed hands and the RTX 3090 Ti was launched at the end of March 2022.

We expect a similar approach for the next generation, but there may of course be some differences in company strategies. Budget-oriented AD106 and AD107-based GPUs will likely sag by 2023. NVIDIA may also consider some affordable RTX graphics cards to replace the GTX 1660 and GTX 1650 series.

About a year after initial launch, Ada models will inevitably receive a refresh. These will arrive on the shelves as always with the addition of “Ti” or “SUPER”. As a final reminder, in late April, Ada Lovelace GPUs are currently undergoing tesdollarser testing by NVIDIA and its manufacturing partners.

NVIDIA has been a dominant player in the graphics card market for a long time. The GPU giant controls roughly 80-85% of the total GPU market. Pioneering technologies such as ray tracing and DLSS undoubtedly strengthened the green team’s hand.

However, the importance of AI and computing for scientific research and other computational workloads continues to grow. While GPUs play an important role in these processes, many other companies, including Intel, are looking for ways to enter the industry. Except for the discontinued Larrabee project, Intel hasn’t attempted a dedicated graphics card since the late 90’s. This time, the blue giant took a serious step under the name Intel Arc and headed for the mobile market first.

Intel has included three separate series for graphics processors to be used in laptop computers, namely Arc 3, Arc 5 and Arc 7. Very soon we will have the opportunity to use desktop screen cards. As a result, an important name enters the graphics market. Also, this name is a technology giant like “Intel”.

Arc Alchemist display cardsyellow will rival NVIDIA and AMD cardsollars up to a certain size. However, we do not expect a competitor to high-end display cards like the RTX 4080/4090 in the first place. In Batdollarsemage and later generations, we can see that Intel has more say in all segments.

AMD will also not stand still and has confirmed that it will announce the RDNA 3 architecture in 2022. We expect the red team to migrate to TSMC’s N5 technology, meaning they will utilize the same or very similar manufacturing technique as NVIDIA.

There is no doubt that NVIDIA currently offers far superior ray tracing performance compared to AMD’s RX 6000 series cardollars. However, AMD should be aware of its shortcomings in this regard and it is time to make an important breakthrough with the new Radeon RX 7000 series. In the meantime, let’s note that Intel will offer ray tracing support from the first generation, and that the cardollarsar will carry the DirectX 12 Ultimate logo.

For the past two years, we have struggled with GPU lows and high prices. It seems that the graphics market will reshape with the year 2022. We hope that with the new architectures, it will be easier to reach kardollarsar and fiyadollarsar will continue to retreat.


Like it? Share with your friends!

332
119 shares, 332 points

What's Your Reaction?

hate hate
1226
hate
confused confused
736
confused
fail fail
368
fail
fun fun
245
fun
geeky geeky
122
geeky
love love
981
love
lol lol
1104
lol
omg omg
736
omg
win win
368
win
Michael Lewis

Emperor

0 Comments

Your email address will not be published. Required fields are marked *

Choose A Format
Personality quiz
Series of questions that intends to reveal something about the personality
Trivia quiz
Series of questions with right and wrong answers that intends to check knowledge
Poll
Voting to make decisions or determine opinions
Story
Formatted Text with Embeds and Visuals
List
The Classic Internet Listicles
Countdown
The Classic Internet Countdowns
Open List
Submit your own item and vote up for the best submission
Ranked List
Upvote or downvote to decide the best list item
Meme
Upload your own images to make custom memes
Video
Youtube, Vimeo or Vine Embeds
Audio
Soundcloud or Mixcloud Embeds
Image
Photo or GIF
Gif
GIF format