NVIDIA GeForce RTX 5090 evaluation: Pure AI extra for $2,000

A $2,000 video card for customers should not exist. The GeForce RTX 5090, just like the $1,599 RTX 4090 earlier than it, is extra a flex by NVIDIA than something actually significant for many players. NVIDIA CEO Jensen Huang stated as a lot when he revealed the GPU at CES 2025, assuming that it will be for hardcore gamers who’ve $10,000 rigs. Personally, I do not know anybody who truly matches that invoice, not except you rely parasocial relationships with streamers. (My very own setup does not even cross $5,000.)

However everyone knows why NVIDIA is hyping up the unattainable RTX 5090: It lets the corporate showcase benchmarks that AMD cannot contact, as soon as once more cementing itself because the supreme chief of the high-end video card market. It is not nearly gaming, both. The RTX 5090 can be being positioned as an AI workhorse because it’s powered by NVIDIA’s new Blackwell architecture, which leans on the corporate’s Tensor Cores for synthetic intelligence work greater than ever. Realistically, although, the $549 RTX 5070 is the GPU extra players will truly be capable of purchase.

I will admit, I went into this evaluation with a combination of pleasure and disgust. It is astonishing that NVIDIA was in a position to stuff 91 billion transistors and 21,760 CUDA cores within the RTX 5090, and I could not wait to see the way it carried out. Nonetheless, I discover it genuinely unhappy that NVIDIA retains pushing the bar increased for GPU costs, within the course of making the gaming world much more unequal. A $2,000 graphics card, in this economic system?!

However after hours of benchmarking and playtime, I spotted the RTX 5090 wasn’t a lot of a menace to gaming accessibility. Rich PC players have at all times overspent for graphics efficiency — I’ve seen individuals (unwisely) pay hundreds greater than shopper GPUs simply to get further VRAM from NVIDIA’s Quadro playing cards. However the rise of PC handhelds just like the Steam Deck, that are a direct offshoot of the Nintendo Change’s success, is a transparent signal that comfort issues greater than uncooked energy to mainstream gamers as we speak. I do not assume many Change 2 consumers are saving up for an RTX 5090.

For the few who can afford it, although, NVIDIA’s new flagship positive is a deal with.

{Hardware}: Leaning extra on AI

In some ways, the RTX 5000 GPU household is the convergence of NVIDIA’s decades-long GPU experience and its newfound function powering the AI hype practice. Certain, they will run video games quicker than earlier than, however what makes them distinctive is their potential to faucet into “neural rendering” AI for even higher efficiency. It is on the coronary heart of DLSS 4, the corporate’s newest AI upscaling expertise, which may now generate as much as three frames for each one which’s truly rendered by the RTX 5090.

That is how NVIDIA can declare this GPU is twice as quick because the RTX 4090, or that the RTX 5070 matches the pace of the 4090. Does it actually matter if these frames are “pretend” if you cannot inform, they usually result in smoother gameplay?

Earlier than I dive additional into the AI aspect of issues, although, let’s take a better have a look at the RTX 5090. As soon as once more, it options 21,760 CUDA cores, up from 16,384 cores on the 4090, in addition to 32GB of GDDR7 VRAM as a substitute of the 4090’s 24GB of GDDR6X. (I believed I used to be future-proofing my desktop after I outfitted it with 32GB of RAM years in the past, however now that video playing cards have caught up I am virtually satisfied to go as much as 64GB.) The 5090 additionally sports activities Fifth-gen Tensor cores with 3,352 of AI TOPs efficiency, whereas the 4090 had 1,321 AI TOPS with last-gen Tensor {hardware}.

	RTX 5090	RTX 5080	RTX 5070 Ti	RTX 5070	RTX 4090
Structure	Blackwell	Blackwell	Blackwell	Blackwell	Lovelace
CUDA cores	21,760	10,752	8,960	6,144	16,384
AI TOPS	3,352	1,801	1,406	988	1,321
Tensor cores	Fifth Gen	Fifth Gen	Fifth Gen	Fifth Gen	4th Gen
RT cores	4th Gen	4th Gen	4th Gen	4th Gen	third Gen
VRAM	32 GB GDDR7	16 GB GDDR7	16 GB GDDR7	12 GB GDDR7	24 GB GDDR6X
Reminiscence bandwidth	1,792 GB/sec	960 GB/sec	896 GB/sec	672 GB/sec	1,008 GB/sec
TGP	575W	360W	300W	250W	450W

I examined the RTX 5090 Founder’s Version GPU (offered by NVIDIA), which is dramatically slimmer than its 4090 counterpart. The 5090 has a glossy two-slot case that may truly slot in small type issue programs. The three-slot 4090, in the meantime, was so huge it felt prefer it was going to tear my PCIe slot out of my motherboard. NVIDIA additionally added one other cooling fan this time round, as a substitute of simply counting on a vapor chamber and a single fan. The 5090’s fundamental PCB sits within the middle of the cardboard, and it is linked to different PCB modules on the PCIe slot and rear ports (three DisplayPort 2.1b and an HDMI 2.1b connection).

Devindra Hardawar for Engadget

DLSS 4: The true star of the present

Whereas multi-frame era is the defining function for the RTX 50 playing cards, there are a number of different DLSS 4 options that ought to assist video games look dramatically higher. Better of all, these capabilities are additionally trickling right down to earlier RTX GPUs. RTX 40 playing cards will probably be extra environment friendly with their single-frame era, whereas RTX 30 and 20 playing cards may also see an improve from AI transformer fashions used for ray reconstruction (resulting in extra steady ray tracing), Tremendous Decision (increased high quality textures) and Deep Studying Anti-Aliasing (DLAA).

These transformer fashions also needs to repair some rendering artifacts current in earlier variations of DLSS. At NVIDIA’s Editor’s Day earlier this month, the corporate confirmed off how the up to date model of Ray Reconstruction made a chainlink fence in Alan Wake 2 seem utterly sharp and clear. An earlier model of the function made the identical fence look muddy, virtually as if it was out of focus. In Horizon Forbidden West, the brand new model of Tremendous Decision revealed extra element from the feel of Aloy’s bag.

DLSS 4 will probably be supported in 75 video games and apps at launch, together with Indiana Jones and the Nice Circle and Cyberpunk 2077, in line with NVIDIA. For titles that haven’t but been up to date with new DLSS menu choices, you’ll additionally be capable of power assist for the newest options within the NVIDIA app.

In use: An absolute powerhouse, with pretend frames and with out

I might virtually hear my motherboard breathe a sigh of reduction after I unplugged the RTX 4090 and swapped within the slimmer 5090. Set up was a cinch, although I nonetheless wanted to plug in 4 PSU connectors to fulfill its demand for 575 watts of energy and a 1,000W PSU. If you happen to’re fortunate sufficient to have a brand new PSU with a 600W PCIe Gen 5 cable, that may also work (and likewise keep away from tons of cable muddle).

I examined the RTX 5090 on my dwelling rig powered by an AMD Ryzen 9 7900X and 32GB of RAM, alongside a 1,000W Corsair PSU. I additionally used Alienware’s 32-inch 4K QD-OLED 4K 240Hz monitor to get probably the most out of the 5090, and actually, you wouldn’t wish to run this GPU on something much less.

As soon as I began benchmarking, it didn’t take lengthy for the RTX 5090 to impress me. Within the 3DMark Metal Nomad check, which is a demanding DX12 demo, it scored 14,239 factors, properly above the 9,250 factors I noticed on the RTX 4090. Equally, the 5090 hit 15,416 factors within the 3DMark Speedway benchmark, in comparison with the 4090’s 10,600 factors. These are notable generation-over-generation positive aspects with out the usage of body era or any DLSS sorcery — it’s simply the uncooked energy you see with extra CUDA and RT cores.

None	3DMark TimeSpy Excessive	Port Royal (Ray Tracing)	Cyberpunk (4K RT Overdrive DLSS)	Blender
NVIDIA RTX 5090	19,525	36,003/166fos	246fps (4X body gen)	14,903
NVIDIA RTX 4090	16,464	25,405/117fps	135fps	12,335
NVIDIA RTX 4080 Tremendous	13,168	18,435/85fps	80fps	8,867
NVIDIA RTX 4070 Ti Tremendous	11,366	15,586/72fps	75fps	7,342

As soon as I began gaming and let DLSS 4 do its magic, my jaw nearly hit the ground. However I suppose that’s only a pure response to seeing a PC hit 250fps on common in Cyberpunk 2077 whereas enjoying in 4K with maxed-out ray tracing overdrive settings and 4x body era. As compared, the 4090 hit 135fps with the identical settings and single body era.

Now I do know most of these frames aren’t technically actual, however it’s additionally the primary time I’ve seen any recreation fill out the Alienware monitor’s 4K 240hz refresh price. And most significantly, Cyberpunk merely regarded wonderful as I rode my bike down rain-slicked metropolis streets and soaked within the reflections and sensible lighting from strong ray tracing.

Like Cypher in The Matrix (removed from the perfect function mannequin, I do know), after struggling by years of low 4K framerates, I couldn’t assist however really feel like “ignorance is bliss” in the case of body era. I didn’t see any artifacts or stuttering. There wasn’t something that took away from my expertise of enjoying Cyberpunk. And the sport genuinely regarded higher than I’d ever seen it earlier than.

And for those who’re the type of one that might by no means stay with “pretend frames,” the RTX 5090 can be the one card I’ve seen that may get near 60fps in Cyberpunk natively in 4K with maxed out graphics and no DLSS. I hit 54fps on common in my testing, whereas the 4090 chugged alongside at 42fps in native 4K. You possibly can additionally compromise a bit and activate 2x or 3x body era to get a strong fps enhance, if the thought of 4x body era simply makes you’re feeling soiled.

And for those who can’t inform, I shortly received over any pretend body trepidation. After I used the NVIDIA app to activate 4x body era in Dragon Quest: The Veilguard, I as soon as once more noticed a mean framerate of round 240fps in 4K with maxed out graphics. I’ve already spent over 25 hours within the recreation, however operating by a couple of missions at that framerate nonetheless felt revelatory. Fight sequences have been clearer and simpler to comply with, probably thanks to higher Ray Reconstruction and Tremendous Decision, and I might additionally make out much more element in my character’s ornate costumes. On the 4090, I sometimes noticed round 120fps with customary body era.

The 5090’s DLSS 4 efficiency makes me desirous to see how the cheaper RTX 5070 and 5070 Ti playing cards carry out. If a $550 card can truly get near what I noticed on the $1,599 4090, even when it’s counting on huge quantities of body era, that’s nonetheless a significant accomplishment. It will even be nice information for anybody who invested in a 4K 120Hz display, which is hard to fill with different mid-range GPUs.

Exterior of gaming, the RTX 5090 additionally managed to transform a minute-long 4K clip into 1080p utilizing the NVENC H.264 encoder in simply 23 seconds. That’s the quickest conversion I’ve seen but. As compared, the RTX 4090 took 28 seconds. Add up these seconds on a a lot bigger mission, and the 5090 might probably prevent hours of repeated rendering time. Naturally, it additionally noticed the quickest Blender benchmark rating we’ve ever seen, reaching 14,903 factors. The RTX 4090, the earlier chief in our benchmarks, hit 12,335 factors.

All through benchmarks and prolonged gaming periods, the RTX 5090 sometimes reached round 70 levels Celsius with audible, however not annoying, fan noise. The cardboard additionally shortly cooled right down to idle temperatures between 34C and 39C when it wasn’t underneath load. Aiming to push the boundaries of NVIDIA’s cooling setup, I additionally ran a number of stress check periods in 3DMark, which entails looping a benchmark 20 occasions. It by no means crashed, and achieved over 97 % accuracy in many of the assessments. There was only one Metal Nomad session the place it scored 95.9 % and failed 3DMark’s 97 % threshold. That might simply be resulting from early driver points, however it’s nonetheless value noting.

The one time I actually received the RTX 5090 cooking was throughout an exploration of the Speedway benchmark, the place I might transfer the digital camera across the ray traced scene and have a look at completely different objects and characters. The cardboard hit 79C virtually instantly and stayed there till I stop the demo. Throughout that session, in addition to typical gaming, the 5090 drew between 500W and 550W of energy.

Trying forward: AI NPCs and neural shaders

On prime of DLSS, NVIDIA can be planning to faucet into its RTX playing cards to energy AI NPCs in games like PUBG and ZooPunk. Based mostly on what I noticed at NVIDIA’s Editor’s Day, although, I’m extra fearful than excited. The corporate’s Ace expertise can let NPCs generate textual content, voices and even have conversational voice chats, however each instance I noticed was robotic and disturbing. The AI Ally in PUBG makes numerous sense on paper — who wouldn’t need a pc companion that would allow you to battle and discover ammo? However within the demo I noticed, it wasn’t a lot of a conversationalist, it couldn’t discover weapons when requested and it additionally took approach too lengthy to hop right into a car throughout a harmful firefight.

As I wrote final week, “I am personally bored with being bought on AI fantasies, after we know the important thing to nice writing and performances is to provide human expertise the time and sources to refine their craft.“ And on a sure degree, I believe I will at all times really feel just like the director Hayao Miyazaki, who described an early instance of an AI CG creature as, “an affront to life itself.”

NVIDIA’s Neural Shaders are an try and carry AI proper into texture shaders, one thing the corporate says wasn’t potential on earlier GPUs. These could be carried out in quite a lot of methods: RTX Neural Supplies, for instance, can use AI to render complicated supplies like silk and porcelain, which regularly have nuanced and reflective textures. RTX Neural Texture Compression, however, can retailer complicated textures whereas saving as much as 7 occasions the VRAM used from typical block compression. For ray tracing, there’s RTX Neural Radiance Cache, which is skilled on stay gameplay to assist simulate path-traced oblique lighting.

Very similar to NVIDIA’s early ray tracing demos, it’s unclear how lengthy it’ll take for us to see these options in precise video games. However from the glimpses to this point, NVIDIA is clearly considering of recent methods to deploy its AI Tensor Cores. RTX Neural Faces, for instance, makes use of quite a lot of strategies to make faces appear extra sensible, and fewer like plastic 3D fashions. There’s additionally RTX Mega Geometry, which may help builders make as much as “100x extra ray traced triangles,” in line with NVIDIA. Demos present it getting used to assemble a big constructing in addition to an unlimited dragon.

Wrap-up: The brand new unattainable GPU king

The $2,000 GeForce RTX 5090 will not be meant for mere mortals, that a lot is evident. Nevertheless it factors to an attention-grabbing new path for NVIDIA, one the place AI options can seemingly result in exponential efficiency positive aspects. Whereas I hate that it’s pushing GPU costs to new heights, there’s no denying that NVIDIA has crafted an absolute beast. However, like most individuals, I’m extra excited to see how the $549 RTX 5070 fares. Certain, it’s additionally going to lean into body era, however a minimum of you received’t need to spend $2,000 to profit from your 4K monitor.

This text initially appeared on Engadget at https://www.engadget.com/gaming/laptop/nvidia-geforce-rtx-5090-review-pure-ai-excess-for-2000-140053371.html?src=rss

Trending Merchandise

Add to compare