This year’s CES had major announcements from all of the big names in the industry with Intel and AMD both announcing new motherboard chipsets and AMD also announcing their upcoming GPUs. Nvidia had a lot to announce as well. The biggest of those was their new 50 series of GPUs based on their Blackwell architecture. That included four GPUs which they planned to launch in January and February. They have the RTX 5090 which is their flagship, the RTX 5080, RTX 5070 Ti, and the RTX 5070 all announced. They also announced Multi Frame Generation which along with the new Transformer based DLSS ray Reconstruction and Super Resolution models make DLSS 4 and Nvidia Reflex 2 as well. RTX Neural Shaders and Neural Faces were both announced as well. Today the performance benchmarks for the upcoming RTX 5090 lift and while I take a look at the new RTX 5090 Founders Edition we can also check out DLSS 4 alongside of seeing how the new flagship GPU performs. It’s a lot to cover, so let’s get to it!
Product Name: Nvidia RTX 5090 Founders Edition
Review Sample Provided by: Nvidia
Written by: Wes Compton
Amazon Affiliate Link: HERE
What is new?
Nvidia announced a lot with this one so let’s try to touch on as much as we can. They of course had a few hardware announcements for the 50-Series. They have announced four cards, the RTX 5090, the RTX 5080, the RTX 5070 Ti, and the RTX 5070. That is the new 50-series RTX family. The slides below include their focus on them. The 5090 for example is targeted at 4k 240 Hz and the other three are targeted at 2k or 1440p performance. They all have GDDR7 VRAM with the flagship RTX 5090 having 32 GB and a 512-bit interface. The RTX 5080 cuts that in half to 16 GB, the RTX 5070 Ti also has 16GB, and then the RTX 5070 has 12 GB.
The new GPUs are based on the Blackwell architecture and they do have new updated RT and Tensor cores with that being the 4th gen for the RT Cores and 5th gen for Tensor cores. The new tensor cores can now also handle floating point (FP) 4 along with FP8 and FP16. They also have introduced an AMP processer which is the AI Management Processor to help schedule AI tasks alongside of graphics rendering. They have increased the number of NV encoders and decoders, it now depends on the card model and isn’t a flat number of them across the entire generation of cards. They have also moved to PCIe Gen 5 and DisplayPort 2.1b including UHBR20. For pricing the RTX 5090 comes in at $1999, the RTX 5090 is half that (just like its memory) at $999. The RTX 5070 Ti is $749 and the RTX 5070 is $549.
Some of the main goals with Blackwell were to optimize the neural workloads and reduce the memory footprint. No big surprises there. Memory is one of the more expensive parts of the cards and anyone who has been paying attention knows that neural workloads and AI have been where Nvidia has been seeing the biggest improvements. Combining those things they have brought AI into shaders with neural shaders. The example of this they show is a hair being rendered with ray tracing and how using spheres rather than triangles helps use less data which means less VRAM and higher frame rates.
Nvidia announced Transformer which replaces CNN as the model they use with Super Resolution. They have improved the design significantly to get better detail when scaling this up. On top of that, they announced DLSS 4 which has improved on the frame generation that they introduced with DLSS 3. It is now Multi Frame Generation. Where before they were able to generate every other frame, they can now do x3 or x4. This gets interesting once they add in DLSS Super Resolution as well because that is already rendering ¾ of the image. With both, they are rendering 15 of 16 pixels using AI. Of course how well this works then depends a lot on how good the renders are. But it gives huge improvements in performance and with frame generation, we know that those improvements still happen even if you are PCU limited for example. The example they show has DLSS off at 27 FPS, turning on Super Resolution it goes to 71 FPS. DLSS 3.5 gets you to 140 FPS and DLSS 4 is 248 FPS. DLSS is already supported by a LOT of games and Nvidia is saying that DLSS 4 will have 75 games and apps supporting DLSS 4 at Day 0. Some of those will be by using the Nvidia App, which can override the DLSS settings on some games. It is also important to note that some of these features will go back and work with legacy cards as well. Specifically, DLAA is going to work back to the 20 series of cards, same with the improved DLSS Super Resolution. The new multi-frame generation however only works with 50 series cards.
They also introduced a new Nvidia Reflex. Reflex helps improve responsiveness to get lower latency through the entire pipeline. Were Reflex offered 50% faster responsiveness the new Reflex 2 gets you 75% by using frame warp. They say that is coming first to 50 series and will be available in games like Valorant soon.
Below I have the specifications for the RTX 5090 as well as the last two xx90 cards. We can see that the number of GPCs hasn’t changed from the 4090 to the 5090 but they did increase the SMs from 128 to 170. The CUDA core SM count is still the same but there are now 21760 CUDA cores to the previous 16384, a 32% increase. We still have 4 Tensor SMs but the new 5th gen design has 680 cores now. For clock speed, the RTX 5090 is set lower than the RTX 4090 with a clock speed of 2407 MHz. There are 32%more RT cores which is a 32$ increase but the RT FLOPS has jumped up 66% with the new RT design. The memory has increased from 24 GB up to 32GB and it now uses GDDR7. They have also increased the memory interface up to 512-bit from 384-bit which the 3090 and 4090 both had. The memory data rate has gone from 21 Gbps up to 28 Gbps and the bandwidth increased with that and the larger pipeline from 1008 GB/sec up to 1792 GB/sec a 77% increase. The cache has increased in side with the L1 cache going from 16384 KB up to 21760 KB and the L2 increasing as well from 73728 KB up to 98304 KB. The TGP also had a big jump going from 450 watts on the RTX 4090 to 575 watts here on the RTX 5090. This is an even bigger gap when you include the RTX 3090 which was 350 watts. The manufacturing process is still the TSMC 4nm 4N process so no changes there I mentioned it earlier but all of the 50 series cards moved to PCIe gen 5 and the RTX 5090 is no different. For pricing, it has an MSRP of $1999 which is $400 more than the RTX 4090 but is similar to the RTX 3090 Ti which also launched at $1999 back in 2022.
Specifications |
GeForce RTX 3090
|
GeForce RTX 4090
|
GeForce RTX 5090
|
GPU Codename |
GA102 |
AD102 |
GB202 |
GPU Architecture |
NVIDIA Ampere |
NVIDIA Ada Lovelace |
NVIDIA Blackwell
|
GPCs |
7 |
11 |
11 |
TPCs |
41 |
64 |
85 |
SMs |
82 |
128 |
170 |
CUDA Cores / SM |
128 |
128 |
128 |
CUDA Cores / GPU |
10496 |
16384 |
21760 |
Tensor Cores / SM |
4 (3rd Gen) |
4 (4th Gen) |
4 (5th Gen) |
Tensor Cores / GPU |
328 (3rd Gen) |
512 (4th Gen) |
680 (5th Gen) |
GPU Boost Clock (MHz) |
1695 |
2520 |
2407 |
RT Cores |
82 (2nd Gen) |
128 (3rd Gen) |
170 (4th Gen)
|
RT TFLOPS |
69.5 |
191 |
317.5
|
Frame Buer Memory Size and Type |
24 GB GDDR6X
|
24 GB GDDR6X
|
32 GB GDDR7
|
Memory Interface |
384-bit |
384-bit |
512-bit |
Memory Clock (Data Rate) |
19.5 Gbps |
21 Gbps |
28 Gbps |
Memory Bandwidth |
936 GB/sec |
1008 GB/sec |
1792 GB/sec |
ROPs |
112 |
176 |
176 |
Pixel Fill-rate (Gigapixels/sec) |
189.8 |
443.5 |
423.6 |
Texture Units |
328 |
512 |
680 |
Texel Fill-rate (Gigatexels/sec) |
555.96 |
1290.2 |
1636.76 |
L1 Data Cache/Shared Memory |
10496 KB |
16384 KB |
21760 KB |
L2 Cache Size |
6144 KB |
73728 KB |
98304 KB |
TGP (Total Graphics Power) |
350 W |
450 W |
575 W |
Manufacturing Process |
Samsung 8 nm 8N NVIDIA Custom Process |
TSMC 4nm 4N NVIDIA Custom Process |
TSMC 4nm 4N NVIDIA Custom Process |
PCI Express Interface |
Gen 4 |
Gen 4 |
Gen 5 |
Launch MSRP |
$1499 |
$1599 |
$1999 |
Before getting into testing I did also run GPUz to double-check that our clock speeds matched up with the specifications. The RTX 5090 Founders Editon had its boost clock at 2407 MHz which matches the specifications. For testing all testing was done using the Nvidia 571.86 driver which was provided to press ahead of the launch.