After months and even years of rumors, NVIDIA introduced the GeForce RTX 4000 series graphics cards to gamers and content producers. While the GTC 2022 event witnessed a special “GeForce Beyond” presentation, details about the “Ada Lovelace” architecture were provided. This architecture takes its name from a mathematician in history.
NVIDIA CEO Jensen Huang introduced the RTX 4090, RTX 4080 16 GB and RTX 4080 12 GB models in the first place. There is no development yet about the RTX 4070. Or we can say that the company has prepared a card such as the RTX 4080 12 GB instead of the RTX 4070. NVIDIA will continue to rely on the RTX 3000 series for the lower performance segment for a while as we understand it.
Ada Lovelace Architecture
The engineers of the green team took advantage of the TSMC 4nm (N4) fabrication technology, which is an optimized version of the 5nm (N5) fabrication technology. The company thus managed to include 76 billion transistors and more than 18,000 shaders in its GPUs. None of the RTX 4000 series GPUs have NVLink. The cards will also continue to use the PCIe Gen4 x16 interface in the same way.
If you remember, Tensor and Ray Tracing cores were updated along with the Ampere architecture. NVIDIA is now transitioning to fourth-generation Tensor cores and third-generation RT cores with the Ada Lovelace architecture. Allegedly, it will offer up to twice the performance of artificial intelligence and up to twice the performance of Ray Tracing.
RT and Tensor Cores
Ada’s new fourth-generation Tensor cores use the FP8 Transformer Engine, first introduced with the Hopper H100 data center GPU, boosting throughput by up to 5X and up to 1.4 Tensor-petaFLOPS. On the other hand, RT cores come with the new Opacity Micromap (OMM) Engine and the new Displaced Micro-Mesh (DMM) Engine. The OMM Engine provides much better rendering of frequently used textures for leaves, particles and hedges. The DMM Engine, on the other hand, offers up to 20x less BVH storage and up to 10x faster Bounding Volume Hierarchy (BVH) rendering time, enabling real-time ray tracing of geometrically complex scenes.
Shader Execution Reordering
Advanced ray tracing requires the computation of a large number of rays hitting many different objects throughout a scene. Thus, different workloads are born for the cores. Shader Execution Reordering (SER) technology will make these previously inefficient workloads much more efficient by dynamically rearranging them. SER can increase shader performance by up to 3x for ray tracing and in-game framerates by up to 25%.
AV1 Codec Support
Graphics cards built on Ada architecture will have the eighth generation NVIDIA Encoder (NVENC) with AV1 encoding support. This will open up new possibilities for broadcasters and videographers. The AV1 codec is 40% more efficient than H.264. It will also allow users broadcasting in 1080p to increase their broadcast resolution to 1440p while operating at the same bitrate and quality.
DLSS 3
DLSS 3 delivers revolutionary breakthroughs in AI-powered graphics while greatly improving performance. Let’s start with the bad news, the next generation DLSS version will only be supported on RTX 4000 series graphics cards. The old generation RTX 3000 series will continue on its way with DLSS 2. NVIDIA says there is a 16-fold increase in performance between DLSS 3 and DLSS 1.
Yields of Architecture
In general, if we compare the Ampere and Ada Lovelace architecture, the following results appear.
- 2x more GPC (Graphic Processing Clusters).
- 50% more cores.
- 50% more L1 cache.
- 16x more L2 cache.
- The number of ROPs has doubled.
- 4th Gen Tensor and 3rd Gen RT Cores.
NVIDIA GeForce RTX 4090 Specifications
NVIDIA’s GeForce RTX 4090 has been long awaited, and it’s finally here. At the heart of the new flagship is the Ada Lovelace AD102 GPU. The GPU, which has a size of about 600 mm2, contains an enormous 76 billion transistors.
The AD102 GPU actually supports up to 144 SMs. The GeForce RTX 4090 combines 16,384 CUDA cores while using 128 of them. The new GPU will have 96MB of L2 cache and a total of 384 ROPs. However, these figures may be slightly lower due to the fact that the GPU used by the RTX 4090 is clipped.
The RTX 4090 Founders Edition appears to have a standard clock speed of 2.23 GHz and an increased clock speed of 2.52 GHz. NVIDIA says its labs have overclocked Ada GPUs to over 3 GHz. The reference design cannot reach these speeds. However, we expect the factory overclocked specially cooled models to reach speeds close to 3.0 GHz.
As for the memory specifications, the AD102 GPU will be accompanied by 24 GDDR6X memories with a capacity of 24GB running on a 384-bit bus interface and running at 21 Gbps. This results in 1 TB/s bandwidth, which is the same as the GeForce RTX 3090 Ti on paper.
The TBP (total card power) of the graphics card is listed as 450W, which means the TGP (total graphics power) could be lower. However, it seems likely that custom designs with massive cooling will consume over 500W.
| RTX 4090 | RTX 4080 16GB | NVIDIA GEFORCE RTX 4080 12GB | RTX 3090 Ti | RTX 3080 | |
|---|---|---|---|---|---|
| GPU | AD102-300 | AD103-300 | AD104-400 | Ampere GA102-225 | Ampere GA102-200 |
| Production technology | TSMC 4N | TSMC 4N | TSMC 4N | Samsung 8nm | Samsung 8nm |
| Mold Size | ~600mm2 | ~450mm2 | ~450mm2 | 628.4mm2 | 628.4mm2 |
| Transistor | ~75 billion | ? | ? | 28 billion | 28 billion |
| CUDA nuclei | 16384 | 9728 | 7680 | 10240 | 8704 |
| TMU / ROP | ? | ? | ? | 320 / 112 | 272 / 96 |
| Tensor / RT Cores | ? | ? | ? | 320 / 80 | 272 / 68 |
| Base Clock | 2230MHz | 2210MHz | 2310MHz | 1365MHz | 1440MHz |
| Boost Clock | 2520MHz | 2510MHz | 2610MHz | 1665MHz | 1710MHz |
| FP32 Calculation | 82.6 TFLOPs | ~50 TFLOPs | ~40 TFLOPs | 34 TFLOPs | 30 TFLOPS |
| Memory | 24GB GDDR6X | 16GB GDDR6X | 12GB GDDR6X | 12GB GDDR6X | 10GB GDDR6X |
| Data Bus | 384-bit | 256-bit | 192-bit | 384-bit | 320-bit |
| Memory Speed | 21Gbps | 23Gbps | 21Gbps | 19Gbps | 19Gbps |
| Band width | 1008GB/s | 736GB/s | 504GB/s | 912Gbps | 760Gbps |
| TBP | 450W | 320W | 285W | 350W | 320W |
| Max. TGP | 660W | 516W | 366W | – | – |
| List price | 1599$ | 1199$ | $899 | 1199$ | 699$ |
| Release date | 12 October 2022 | November 2022 | November 2022 | June 3, 2021 | September 2020 |
NVIDIA GeForce RTX 4090 Performance
According to NVIDIA’s claims, next-generation graphics cards will offer up to two to four times higher performance, depending on the usage scenario.
NVIDIA says they reach over 300 FPS at 1440p resolution in competitive games. There are still no 1440p monitors with a 360 Hz refresh rate on the market, but now manufacturers must have rolled up their sleeves because hardware that can take advantage of it is coming to the market.
In the RTX 3090 Ti benchmark table offered by NVIDIA, we see games such as Microsoft Flight Simulator, Warhammer 40,000: Darktide and Cyberpunk 2077. Tests were conducted on the i9-12900K processor, 32 GB of RAM and Windows 11 operating system. It was running in DLSS Performance mode.
NVIDIA GeForce RTX 4090 Release Date and Price
The NVIDIA GeForce RTX 4090 Founders Edition will be available on October 12, priced at $1599. As you know, the graphics cards of production partners such as MSI, ASUS and Gigabyte have different price tags depending on the model. Therefore, priced
The RTX 3090 Ti was launched long after the RTX 3000 series was introduced. That’s why we don’t expect the RTX 4090 Ti or the possible Titan model anytime soon.
