NVIDIA RTX 5080 vs RTX 5090 Buying Guide

NVIDIA shipped the RTX 5090 and the RTX 5080 weeks apart on the same Blackwell generation, and the gap between them is wider than a normal step up the lineup. Both are GDDR7 cards on TSMC's 4N process, both carry DLSS 4 with multi frame generation, and neither supports NVLink. What separates them is that the 5090 is built on the full GB202 die while the 5080 uses the smaller GB203,^[6] and that difference cascades into nearly double the CUDA cores, double the VRAM, and a price that is also roughly double.^[2][3] The buying question is rarely about raw speed, because the 5090 is the faster card by a large margin in almost every workload. The question is whether that margin is worth the money for the resolution and the workload you actually run. This article walks through the specs, the gaming and AI numbers, the power bill, and the real street price, and then gives a verdict for each buyer type.

What is actually different between the two dies

The headline spec sheet reads like two different product categories rather than a tier and a flagship. The RTX 5080 has 10,752 CUDA cores, 336 fifth-generation Tensor cores, 84 ray tracing cores, 16 GB of GDDR7 on a 256-bit bus, and a 360 W board power rating. The RTX 5090 has 21,760 CUDA cores, 680 Tensor cores, 170 ray tracing cores, 32 GB of GDDR7 on a 512-bit bus, and a 575 W board power rating.^[2][5] The 5090 has about twice the compute and twice the memory of the 5080, which is a bigger jump than the step from a 4080 to a 4090 in the previous generation.

One detail that surprises people is the memory speed. The 5080's GDDR7 runs at 30 Gbps, which is slightly faster per pin than the 5090's 28 Gbps. But the 5090's bus is twice as wide, so its total bandwidth works out to about 1,792 GB/s against the 5080's roughly 960 GB/s.^[2][3] That 87 percent bandwidth lead is the number that matters most for AI inference, which is unpacked in the AI section below. Both cards use the new 12V-2x6 power connector and a PCIe Gen 5 interface, so a modern mid-range case and a current power supply can physically accommodate either.^[5]

Gaming performance at 4K and 1440p

At 4K the 5090 is a different machine. Independent benchmarking puts it anywhere from 30 percent to about 69 percent faster than the 5080 at 4K, depending on the title, with most games landing in the 45 to 55 percent band.^[1] In Dragon's Dogma 2 at 4K the 5090 averaged 133 FPS against a figure that gave the 5080 a 57 percent deficit, and in Dying Light 2 the 5090's 4K lead was around 56 percent. Ray tracing widens the gap further, with the 5090 about 57 percent ahead of the 5080 at 4K in the Dragon's Dogma 2 maximum RT preset.^[1] If you are playing at 4K, the 5090 is the unambiguous performance pick and the 5080 leaves a visible amount of frame rate on the table in the most demanding titles.

At 1440p the story tightens. The two cards are commonly within 10 percent of each other at 1440p because the GPU stops being the only bottleneck and the CPU starts to share the load.^[1] In FFXIV at 1440p the 5080 held 217 FPS against the 5090's 317 FPS, and in Dragon's Dogma 2 at 1440p the 5080 delivered about 71 percent of the 5090's performance at roughly half the price. For a 1440p high refresh display, the 5080 is the rational buy. You pay a premium for the 5090 at that resolution, but the extra cores and bandwidth are partly wasted because the frame rate ceiling is set by the CPU and the display rather than the GPU.

Where the 32 GB of VRAM and the bandwidth actually matter

This is the section that decides the purchase for anyone doing local AI rather than pure gaming. LLM inference at the decode stage is memory bound, not compute bound, so the token generation rate tracks memory bandwidth and available VRAM far more closely than it tracks CUDA core count.^[3] The 5090's 1,792 GB/s and 32 GB of VRAM translate almost directly into higher tokens per second and the ability to fit larger models in memory. Practical testing puts the 5080 in the 40 to 60 tokens per second range on quantised 7B to 13B parameter models, which is a strong number for a $999 class card.^[3] The 5090's 32 GB becomes the deciding factor the moment you want to run 30B or larger models, or to do full precision fine tuning, because those workloads simply will not fit in 16 GB.

The honest framing is this. For gaming and for the majority of local AI use, which is small to medium quantised models, the 5080's 16 GB is enough and the 5090's extra VRAM is headroom you will not touch. For a researcher, a studio, or a developer who regularly loads 30B plus models or trains on full precision, the 5090 is the only sensible card in this comparison, and the 5080's 16 GB becomes a hard ceiling that forces you to offload to CPU memory, which collapses the speed advantage. Neither card supports NVLink, so multi GPU tensor parallelism is off the table for both, and that limits how far you can scale a single rig with these two parts alone.^[2]

Power, thermals, and the supply question

The 5090's 575 W board rating against the 5080's 360 W is a 60 percent increase in power draw, and it is the reason the 5090 needs a larger and more expensive power supply. A quality 850 W unit is the practical minimum for a 5080 build, while a 5090 build should start at 1,000 W and many reviewers recommend 1,200 W to leave headroom for transient spikes, which are the reason NVIDIA pushed the 12V-2x6 connector. Thermally the 5090 is also the harder card to cool, and a high end air cooler or a small custom loop is a more comfortable match than it is for the 5080, which behaves like a normal high end card. If your case is compact or your power supply is already committed, the 5080 is the less disruptive choice and will run cooler and quieter in the same chassis.^[5]

The price that matters is the street price, not the MSRP

The official list prices are $999 for the 5080 and $1,999 for the 5090, a clean 2x split that mirrors the spec split.^[3] The street prices have not been that clean. Since launch, the 5090 has regularly traded well above MSRP, with reviews and retailer listings noting scalped pricing that pushes it toward the $2,500 and higher range in many markets, while the 5080 has settled closer to its $999 list, typically in the $1,000 to $1,100 band when stock is normal.^[1][4] That gap between list and street changes the value maths more than the spec sheet does. At $999 the 5080 offers excellent value at 1440p and for mid sized AI models. At $1,999 the 5090 is a premium flagship that earns its keep at 4K and for large AI workloads. At $2,500 or more, the 5090 stops being a buying decision and starts being a pricing problem, and the rational move is to wait for stock to normalise or to buy the 5080 and accept the lower ceiling.

Verdict by buyer

For the 1440p gamer, buy the RTX 5080. It is within about 10 percent of the 5090 at that resolution, runs cooler, needs a smaller power supply, and costs roughly half as much. For the 4K gamer who plays the most demanding titles with ray tracing and wants the highest frame rate the platform offers, buy the RTX 5090, and only at or near MSRP. For the local AI developer, the decision is set by model size. If your work stays at 13B and below, the 5080's 16 GB and 960 GB/s are enough. If you run 30B plus models or do full precision fine tuning, the 5090's 32 GB and 1,792 GB/s are the minimum that keeps the workload on the GPU at all. If you are doing both gaming and AI, the 5090 is the more future proof card, but the 5080 is the more sensible one, because the extra 5090 spend buys you headroom that is only useful at the top end of each workload. The on device AI trend that both of these cards feed is also visible on the phone side, where Google's Tensor G5 in the Pixel 10 makes the same bandwidth versus capacity trade off at a much smaller scale, and that trade off is broken down in our Tensor G5 analysis for the Pixel 10, which is worth reading if you are thinking about where AI compute is heading across form factors.

Frequently asked questions

How much faster is the RTX 5090 than the RTX 5080 in games?

At 4K the 5090 is commonly 45 to 55 percent faster, with a measured range of about 30 to 69 percent depending on the title. At 1440p the two cards are usually within 10 percent of each other because the CPU and the display, not the GPU, set the ceiling.^[1][7]

Which card has more VRAM and does it matter for AI?

The RTX 5090 has 32 GB of GDDR7 against the 5080's 16 GB. For AI it matters a lot, because LLM decode is memory bound. The 32 GB lets the 5090 run 30B and larger models fully in GPU memory, while the 5080 is limited to smaller or more aggressively quantised models that fit in 16 GB.^[3]

Do I need a bigger power supply for the RTX 5090?

Yes. The 5090 has a 575 W board rating against the 5080's 360 W. A 1,000 W quality supply is the sensible minimum for a 5090 build, with 1,200 W recommended for headroom on transient spikes. The 5080 is comfortable on an 850 W unit. Both cards use the 12V-2x6 connector.^[5]

Which is the better value at the launch prices?

At the official $999 and $1,999 list prices, the 5080 is the better value for most buyers, because it delivers close to the 5090 at 1440p and for mid sized AI models at half the cost. The 5090 earns its premium only at 4K and for large AI workloads. The value equation flips against the 5090 whenever street pricing pushes it well above MSRP, which it regularly does.^[1][4]

Can I use both cards together for more AI performance?

Not in the way multi GPU setups used to work. Neither the RTX 5090 nor the RTX 5080 supports NVLink, so there is no high bandwidth GPU to GPU link for tensor parallelism. You can run two cards in a system for separate tasks or for data parallelism, but you cannot treat them as one larger GPU the way a linked pair of 2080s or 3090s could be treated.^[2]