When money is no object: Asus GeForce Rog Strix RTX 3090 tested
29.9.2020
Translation: machine translated
The NVIDIA GeForce RTX 3090 is a beast. It takes the lead as the fastest consumer graphics card. However, it is currently only five per cent faster than the RTX 3080 in games, but costs twice as much.
Jensen Huang presented the RTX 3090 as a "BFGPU" (Big Ferocious GPU) and advertised it with 8K gaming. With its 28 billion transistors on 628 mm2, the graphics processor of the RTX 3090 is truly suitable for large formats. By comparison, the NV10, Nvidia's first graphics chip in the GeForce series, had 17 million transistors in 1999. In the last twenty years, the number of transistors on GeForce GPUs has therefore multiplied by a factor of 1647.
The card in detail
The Ampere architecture of the RTX 3090 replaces the Turing architecture of its predecessor, the RTX 2080 Ti. Although Ampere is a further development of Turing, which was a mini-revolution, there are still various changes. Nvidia uses 8 nm production from Samsung. This brings greater transistor density and performance efficiency. The new streaming multiprocessor (SM) doubles the theoretical maximum FP32 processing power. This is why the TFLOPs of the 30-series cards are so much higher. The RT and Tensor cores have also been upgraded, bringing performance improvements in certain areas. Other important changes include the switch to GDDR6X memory and PCIe 4.0, which significantly increases memory and PCIe bandwidth.
The shader cores of the Ampere architecture are said to be 2.7 times faster, the ray tracing cores 1.7 times faster and the Tensor cores up to 2.7 times faster than those of the Turing architecture.
On the GA-102 chip of the GeForce RTX 3090, 82 of 84 shader clusters are activated, resulting in a total of 10,496 CUDA cores. In terms of memory, the GeForce RTX 3090 takes a big step forward and offers more than twice as much memory as the RTX 2080 Ti. The new flagship is equipped with 24 GB GDDR6X memory from Micron. This means that the RTX 3090 delivers a bandwidth of almost 1 TB/s. The memory is 19.5 Gbps fast. In conjunction with the 384-bit bus interface, this results in a bandwidth of 936 GB/s.
Compared to the RTX 2080 Ti, the RTX 3090 has more of pretty much everything: CUDA cores, RT cores, texture units, ROPS, cache, memory and memory bandwidth. Tensor cores are the exception, as Ampere halves the number of cores per SM. Each tensor core works two to four times faster, depending on the workload. In terms of clock speed, the RTX 3090 goes one better: 1.7 GHz in Boost compared to 1.54 GHz for the RTX 2080 Ti. Asus provided me with the GeForce Rog Strix RTX 3090 for the test. Here, the boost clock is even 1890 MHz
Despite the new process, which results in better energy efficiency, the power consumption has increased significantly due to the enormous amount of additional transistors and processing blocks: 350 watts, which is 100 more than the RTX 2080 Ti.
More performance naturally also means more waste heat. Nvidia has given the Founders Edition a new, special cooler. The test will show how the Strix performs in the classic design with three axial fans.
The design of the Strix is kept simple - this was different in previous generations. The only RGB colour is on the front with the Rog logo. Otherwise, the card is grey-black. The backplate is also rather plain, apart from the cut-outs at the rear end. In contrast to the Founders Edition, the Strix does not require a 12-pin connector, instead three standard 8-pin connectors are required.
With a length of just under 32 centimetres, the card is not for compact systems. At 5.78 centimetres or three slots thick, the Strix is a real piece of furniture.
In terms of connectivity, two HDMI 2.1 and three DisplayPort 1.4 ports are available. USB-C is missing. Asus recommends 750 W for the power supply unit.
I tested the card on my DimasTech Easy V3.0 benchtable with the following components:
Synthetic gaming benchmarks, volume and temperatures
In the synthetic benchmarks from 3D Mark, I only give the graphics score. The benchmark is very inconsistent with the overall score, which is why the result is distorted. I run the benchmarks three times each and select the best result. Here are the results of the Time Spy and Fire Strike benchmarks:
In the four benchmarks, the RTX 3090 beats the RTX 2080 Ti by around 45 per cent. Compared to the lower model, the RTX 3080, it is only 15 per cent. The card remains relatively quiet at 44 dB. For the sake of completeness, I will also provide you with the values of the RTX 2080 Super, but will not discuss them.
The RTX 3090 reached a maximum temperature of 62° Celsius in Time Spy. The average was 59° Celsius. Compared to the Asus TUF RTX 3080, these are very good values. The card got warmer despite a lower TDP. Since I did the test on the open testbench, higher values are to be expected in a case. Here is the heat development on the back of the PCB during the two-minute Time Spy demo:
As far as the noise development of the two cards is concerned: If I turn the three axial fans up to 100 per cent, they generate 59 dB. This is rather loud and can be annoying. Fortunately, the fans never run at 100 per cent when gaming and even if they do: If the card is in the case, the sound is still somewhat muffled. If I run the Time Spy benchmark with the fans at 100 per cent, I get 307 points in the Graphics Score 20. This is actually a slightly worse result than the first time I ran the benchmark with the fans on standard. And this despite the RTX 3090 reaching a maximum temperature of 53° and an average of 48° Celsius at 100 per cent fan speed. It is not clear whether the temperature difference of 11° Celsius on average has no influence on performance or whether the result speaks in favour of the inaccuracy of the benchmark.
In order to be able to make a better statement about the cooling performance of the cards, I test them normalised by volume. I lower the percentage performance of the fans step by step until I measure 40 dB at a distance of 30 centimetres from the fans. This means that the Strix fans are running at 44 per cent of their maximum power. People perceive 40 dB as quiet. If I run Time Spy like this, I achieve a score of 20,322 points. This means I achieve almost the same result as when running the fans at 100 per cent. Whether an average of 48° at 100 per cent fan speed or 62° Celsius at 44 per cent fan speed: within this fourteen-degree range during the test, the card's clock rate is fairly unimpressed.
I can't say anything about the card's power consumption. I don't have the appropriate tools to provide exact information.
Application: Puget Systems Photoshop Benchmark
The Photoshop benchmark uses the following reference workstation as the basis for calculating the scores:
- Intel Core i9 9900K 8 Core
- NVIDIA GeForce RTX 2080 8GB
- 64GB RAM
- Samsung 960 Pro 1TB
The results of the reference workstation can be used to estimate how well other systems perform. My test benchmark with the GeForce RTX 3090 achieved the following results:
The testbench with 3090s performs better in the test than the testbench with 2080 Ti and 3080, but the difference is relatively small at 17 and 37 points respectively. The performance of Photoshop depends more on the CPU than the GPU. The clock rate also plays a greater role than the number of Cuda Cores. This explains the narrow result.
Application: Puget Systems Premiere Benchmark
In contrast to the Photoshop benchmark, the test benchmark here does not compete against a reference workstation. With the Puget Systems Premiere Benchmark, the score is calculated relative to the existing frame rate of the test videos. If the test video has an FPS of 29.97 and the system renders it at 29.97 FPS, this means 100 points. If it is only 14.98 FPS, 50 points are awarded.
The benchmark runs media in the formats 4K H.264 with 150 Mbps in 8 bit (59.94 FPS), 4K ProRes 422 16 bit (59.94 FPS) and 4K RED (59.94 FPS). He tested live playback in Adobe Premiere Pro and the export. A value of 100 is the maximum for live playback, as Premiere cannot play back the media faster than specified. For export, on the other hand, over 100 points are feasible, as rendering is not limited to the FPS of the media.
This is how the RTX 3090 compares to the 2080 Ti and 3080:
In Premiere Benmark, the difference to the predecessor 2080 Ti is 35 per cent greater than in Photoshop. Compared to the 3080, the performance increase of 6.5 per cent is rather small. This is because Premiere utilises the card's Cuda Cores and therefore benefits from a much greater performance boost.
Blender
In the bmw27 Blender benchmark, the RTX 3090 beats its predecessor by far. Compared to the 2080 Ti, the performance has almost doubled. Compared to the 3080, the 3090 still renders the test scene 15 per cent faster.
The games
The games paint a similar picture to the one that many reviewers have already drawn with the Founders Edition: The RTX 3090 doesn't deliver much more performance than the RTX 3080. In 1080p and 1440p, the difference is 7 per cent. In 2160p it is only around 5 per cent. Other reviewers have certified that the 3090 delivers 10 per cent more performance than the 3080. In my case, "Strange Brigade" and "Red Dead Redemption 2" in particular ensure that the difference is only 5 per cent. The RTX 3080 even performs better than the RTX 3090 in these games, although it is possible that the games were not yet optimised for the RTX 3090 at the time of testing.
At 1080p resolution, the 3090 delivers around 15 per cent more FPS compared to the 2080 Ti. This value increases with increasing resolution. At 1440p it is 23 per cent and at 2160p even 36 per cent.
In case you're wondering why the RTX 2080 Ti is missing the percentiles: I recorded the FPS differently when I tested it back then. Unfortunately, I had to return the 2080 Ti to Asus, which is why I couldn't do any night tests.
And 8K?
Of course, I also tested 8K gaming - or rather gaming in 7680x4320 resolution. As the VRAM requirements are particularly high at 4320p, I chose "Gears 5" from my test suite. Compared to other games, the title doesn't consume a lot of graphics memory. At 2160p it's around 5.5 GB and at 4320p around 9.8 GB. The game should therefore also be playable with a 2080 Ti or a 3080. Unfortunately, I currently lack both cards for a comparison, as I had to return them after testing.
I achieve an average of 29 FPS. At the 99th percentile it's 23 and at the 99.9th percentile 14 FPS. This means that "Gears 5" is theoretically playable, but as it is a shooter, I would like to see at least 60 FPS. Future titles will be even more performance-hungry. Nvidia's promise of 8K gaming is therefore still premature with the 3090.
Conclusion: Not worth the extra charge over the 3080 according to my test method
With the RTX 3090, it is possible to play games at over 60 FPS in 2160p resolution without any problems. However, the card is only suitable for gaming in 8K to a limited extent. Anyway: Does anyone really need that these days?
The leap in performance compared to its predecessor, the 2080 Ti, is enormous. This doesn't just apply to games: the increase in performance is also huge in Premiere and Blender. Compared to the RTX 3080, however, the difference in performance is limited. This applies to games and the applications tested. However, this does not mean that the card does not have its justification. In my test suite, no programme or game simply exploits the card's specifications to the full. The surcharge from a 3080 to a 3090 definitely doesn't pay off here.
If you always want the best, there is no way around the RTX 3090. However, if you don't use any applications that require the 24 GB memory of the 3090, the increase in performance is (still) too small. This does not mean that more memory will not be necessary in the future. It's not just applications that are getting hungrier and hungrier in this respect, games also need more and more memory. You are definitely equipped for the future with the RTX 3090.
In which case should you buy an RTX 3080 now? As always, it depends on what you want. If you have the money, always want the best of the best and have the patience to wait for a card - they are in very short supply - then go for it. In any other case, I would go for an RTX 3080 or wait and see what the other cards in the RTX 30 series or AMD with the Radeon 6000 series have to offer.
Kevin Hofer
Senior Editor
kevin.hofer@digitecgalaxus.chFrom big data to big brother, Cyborgs to Sci-Fi. All aspects of technology and society fascinate me.
