The RTX 4090 is a massive performance upgrade - but nowhere near four-times of the RTX 3090 Ti.
Let's start with the obvious. The RTX 4090 is a big card, as in, I had to remove the hard drive array inside my E-ATX case to fit it, and remove the other PCIe card as well. It comes with an anti-sag bracket for a reason, it's heavy as hell, more than 2kgs, and out-weighs the RTX 3090 by a lot. The 3.5 slot card is also.. big, 15 cm wide, 7.5 cm tall, and 34 cm long.
The cards are made with a PCIe 5.0 PSU connector, and an adapter with 4x8-pins is supplied, however, due to thick cabling, it is pushed up against the side of the glass panel of my case, and I can't realistically see a scenario where adapters are used in SFF builds. It also really challenges your cable management skills, especially if you like me, don't have any. Oh, and a 1000 watt PSU is recommended due to the card using 447.3 watts at full load. If you combine it with one of the new generation CPUs, you will draw some 700 watts from the CPU/GPU alone. But at least it isn't 800 watts as the early rumours suggested. However, while the exact same Gigabyte version of the RTX 3090 Ti ran idle at 37 degrees, this hits 47.
Gigabyte has done what you would expect of them, dual bios, slight overclock, massive cooling, and RGB built into the fans for some subtle, but really effectful RGB lightning - I actually didn't turn it off, which I usually do. Add massive fans, a metal backplate and heat pipes galore, and you have a card that despite hours of torture, didn't go past 68 degrees. There is also a nice four-year warranty, and the Windforce cooling, consisting of 10 heat pipes and three 110mm fans with graphene nano lubricant are supported by a massive screen-cooling along with a vapour chamber.
My only real complaint, besides the price, is that the anti-sag bracket comes with a manual, and requires you to fiddle with motherboard screws, and an engineering degree certainly helps with mounting the damn thing. It is however, also a much more aesthetically pleasing solution than the small jack-stand that came with the RTX 30 series.
This is an ad:
The new Ada Lovelace architecture is quite the leap performance wise. While I won't bore you with the 30 page tech deep dive that was a part of the review guide, I do need to point out some very interesting technical advances. Expected to run at 2235 MHz base clock, with a boost to 2535, the test sample actually ran slightly higher at 2655 Mhz, and I hadn't even done anything to overclock it, however, it seems there isn't a lot of headroom to tinker with it.
The AD102 GPU chip used has been combined with 16384 CUDA cores, and they do the brunt of the computing work. The 4th gen Tensor cores do double the performance in the AI department, while the 3rd gen RT cores do double the performance in the Ray-Tracing department. To be more specific, it now has two new hardware components, the "Opacity Micromap Engine" that speeds up the Ray-Tracing itself, and the "Displaced Micro-Mesh Engine" that generates "Displaced Micro-Triangles" on-the-fly to create additional geometry. Basically, it inserts extra frames that have been calculated from scratch, such as dynamic shadows. It also has "shader execution reordering" - which is a fancy way of saying that it can better organise its own workload for shaders, improving performance.
However, one of the main stables of the 4000 series is the new DLSS 3, which was not available fully for testing at the time of writing. It combines many things, including a new Optical Flow Accelerator, and AI frame generation, which isn't built into older cards, and therefore DLSS 3 only works on the RTX 40 series. DLSS 3 basically combines a few new tricks of frame generation with the "old" super resolution upscaling and integrates the Reflex system, which at least on the 30 series worked very well. The main goal is to go 100+ FPS in modern titles in 4K, which is great, but the quadruple performance promised is with a caveat, as it requires use of both DLSS and Ray-Tracing at the same time, and compared with the same titles using either DLSS 2 or no DLSS, running natively, quadruple performance is possible, but reading through all the material makes you realise it's in very specific conditions only. However, given the raw computing power, it should still yield a massive boost in the 80-130/140% range.
This is an ad:
One of the critiques of DLSS 3 is that there are so many fundamental changes that old cards won't benefit from it, however, that is not completely true as the fine prints read that developers can integrate Nvidia Reflex into their games, which is integrated into DLSS 3, and thus partially use DLSS 3 and at minimum get reduced input lag - which is always welcome.
Before the benchmarks, the raw numbers; 128 SMs, 128 RT Cores, 512 Tensor cores, 512 Texture Units and 176 ROPS. Almost double the Texture Fill Rate, more than double the Pixel Fill Rate. It's not just more of everything, it's newer and faster hardware all around. L2 cache has gone from 6144KB to 73728KB. No, it's not a typo. The old RTX 3090 Ti had 28.3 billion transistors, this has 76.8 billion.
The test-system is based on an AIO cooled Ryzen 7900X, 32GB DDR5 CL40 memory, and everything running from PCIe 4.0 NVMe and an Asrock X670E Taichi motherboard. First a run had been made with an identical setup using an RTX 3090 from Gigabyte. Some titles didn't work with the beta drivers during the test period, so Red Dead Redemption 2 and Metro Exodus Enhanced, titles almost made for this card, aren't included in the benchmarks. Far Cry 6 was a bit of the same, the RTX 4090 performing worse in all but 4K where there was a 42.6% increase, at 107 FPS.
And now for the performance numbers - as compared to the RTX 3090
Time Spy: 29071
Time Spy Extreme: 16379
Fire Strike: 37445
Fire Strike Extreme: 37445
Fire Strike Ultra: 11430
Port Royale: 24564
Best case, 64% increase over the RTX 3090, 87% increase in Ray-Tracing.
So, should you buy one? Well if you have the cash to spare and want the best on the market, and maybe you, like me, look forward to the DLSS 3 implementation in Cyberpunk 2077, or you play Microsoft Flight Simulator - then yes. Has Gigabyte done a fabulous job at cooling this card? Yes, they have. But we are still left with a few issues. The card has grown so big, not helped by the massive power adapter cable, that not all cases can even fit it. It does require your PSU to be very, very good, and 1000 watts is extremely highly recommended. And we are still way off a general two-to-four-times leap in performance as promised, however, it is still superior to prior releases and their performance uplifts, and we are now truly a 4K/60FPS generation no matter the game. Well, almost.