RTX 5090 cloud GPU benchmark: how it compares with the RTX 4090 for AI

Short answer: In the independent tests below, the RTX 5090 generates LLM tokens about 1.3 to 1.6 times as fast as the RTX 4090. It runs SDXL and FLUX.1 about 1.3 to 1.6 times as fast at matched precision. It reads prompts (prefill) at about the same speed, and in one test it was slower. It also has 32 GB of VRAM against the 4090’s 24 GB. We found no independent, like-for-like video generation comparison, so this page makes no video speed claim for the 5090. It gives our own measured Wan 2.2 times on an RTX 4090 instead.

Every number below is taken from the source named next to it, or was measured by us on a CloudGPU RTX 4090 instance on 2026-11-04. The speedup column is our arithmetic on those published figures.

Specifications

RTX 4090 RTX 5090
VRAM 24 GB GDDR6X 32 GB GDDR7
Memory bandwidth 1,008 GB/s 1,792 GB/s
Tensor-core precision goes down to FP8 FP4

Source: NVIDIA product pages. A higher memory bandwidth mostly helps workloads that read all the weights for every step, such as LLM token generation.

Image generation (SDXL, SD 1.5, FLUX.1)

Test Settings RTX 4090 RTX 5090 5090 vs 4090 Source
SDXL FP16, UL Procyon 1024×1024, 100 steps, batch 1 7.461 s/image 5.223 s/image 1.43× StorageReview, 2025-01-23
SD 1.5 FP16, UL Procyon 512×512, 100 steps, batch 4 1.188 s/image 0.763 s/image 1.56× StorageReview, 2025-01-23
SD 1.5 INT8, UL Procyon 512×512, 100 steps, batch 4 0.503 s/image 0.394 s/image 1.28× StorageReview, 2025-01-23
SDXL Base 1.0, SD.Next 512×512, 50 steps, batch 16 67.19 it/s 85.75 it/s 1.28× Puget Systems, 2025-02-21
SDXL Base 1.0, SD.Next 512×512, 50 steps, batch 1 16.78 it/s 16.16 it/s 0.96× Puget Systems, 2025-02-21
FLUX.1-dev, TensorRT, FP8 on both cards 30 steps (resolution not stated) 10,620 ms 6,681 ms 1.59× NVIDIA Technical Blog, 2025-05-14
FLUX.1-dev, TensorRT, 5090 at FP4 vs 4090 at FP8 30 steps (resolution not stated) 10,620 ms 3,853 ms 2.76× NVIDIA Technical Blog, 2025-05-14

Notes:

  • Puget Systems attributes the small-batch SDXL result, where the 5090 is slightly slower, to a CPU bottleneck. A single small image does not keep either card busy.
  • The 2.76× FLUX.1 figure compares the 5090 at FP4 with the 4090 at FP8. The 4090 has no FP4 path, so part of that gain comes from the lower precision rather than the card alone. At the same FP8 precision the figure is 1.59×.

LLM inference

Test Settings RTX 4090 RTX 5090 5090 vs 4090 Source
Token generation, llama.cpp b4493 Phi-3-mini-4k-instruct Q4_K_M 245 tok/s 316 tok/s 1.29× Puget Systems, 2025-02-21
Prompt processing, llama.cpp b4493 Phi-3-mini-4k-instruct Q4_K_M 13,697 tok/s 9,967 tok/s 0.73× Puget Systems, 2025-02-21
UL Procyon AI Text Generation, Mistral TensorRT 183.3 tok/s 255.9 tok/s 1.40× StorageReview, 2025-01-23
UL Procyon AI Text Generation, Llama 3 TensorRT 150.0 tok/s 214.3 tok/s 1.43× StorageReview, 2025-01-23
UL Procyon AI Text Generation, Llama 2 TensorRT 92.9 tok/s 134.5 tok/s 1.45× StorageReview, 2025-01-23
Token generation (tg128, flash attention on), llama-bench Llama 2 7B Q4_0 188.96 tok/s 300.40 tok/s 1.59× llama.cpp CUDA scoreboard, community results*
Prompt processing (pp512, flash attention on), llama-bench Llama 2 7B Q4_0 14,770.63 tok/s 14,970.15 tok/s 1.01× llama.cpp CUDA scoreboard, community results*

* The scoreboard rows were submitted by different users, on different machines and llama.cpp builds. Treat them as indicative rather than a controlled test.

The 5090’s gain shows mainly in token generation, which fits its higher memory bandwidth. Prompt processing gains little. In Puget’s early-2025 llama.cpp run it was slower on the 5090, and newer builds may change that result. We found no independent vLLM comparison between the two cards.

Video generation: our RTX 4090 measurements

We found no independent, like-for-like Wan or HunyuanVideo comparison between the two cards, so we don’t give a 5090 speedup for video. What we can give is what we measured ourselves on an RTX 4090 instance.

Model (ComfyUI) Output Settings Time per clip Peak VRAM
Wan 2.2 TI2V-5B, FP16 weights 1280×704, 121 frames (5 s at 24 fps) text to video, 20 steps, uni_pc, CFG 5 175 s first run, 169 s second run 21.7 GB
Wan 2.2 T2V-A14B, FP8 scaled weights (high-noise and low-noise experts) 832×480, 81 frames (5 s at 16 fps) text to video, 20 steps (10 per expert), euler, CFG 3.5 361 s first run, 354 s second run 21.0 GB

How we measured:

  • One RTX 4090 24 GB on CloudGPU, driver 580.105.08, PyTorch 2.14.1 with CUDA 13.0, ComfyUI’s native nodes, measured on 2026-11-04.
  • The text encoder was umt5_xxl_fp8_e4m3fn_scaled, and we used ComfyUI’s default memory management, with no extra speed-up nodes such as TeaCache or Sage Attention.
  • Time per clip is wall-clock time from queueing the prompt to the finished MP4.
  • The first run is the first generation after loading the model. The second run used a new seed, with the model already in memory.
  • Model download and install time is not included.

At current rates, one 5B clip at about 3 minutes costs about $0.019 of RTX 4090 time. One 14B clip at about 6 minutes costs about $0.039. The RTX 4090 rate is $0.39/hour and the RTX 5090 rate is $0.69/hour (prices as of 2026-11-07 17:03 (UTC+8), live on /pricing).

Both workflows peaked at about 21 GB, close to the 4090’s 24 GB. The 5090’s extra 8 GB leaves room for longer clips, higher resolution, or keeping more of the model on the card instead of offloading it.

Which one should you rent?

  • LLM serving and chat, where token generation speed matters: the 5090 is the faster card in every token generation test above. Its 32 GB also fits larger quantised models or longer contexts.
  • Image generation in batches: the 5090 was 1.3 to 1.6 times as fast in the tests above. For single small images the two cards were about the same.
  • Wan 2.2 video in ComfyUI: the 4090 ran both the 5B and 14B FP8 workflows in our test, at the times above. If your clips need more than 24 GB, rent the 5090, a 48 GB card or an 80 GB card.
  • Prompt-heavy work (long documents, RAG): prefill speed was similar on both cards, so compare the rates on /pricing.

Billing is per minute, with a 1-minute minimum, so you can run the same job on both cards and keep the faster one. Which cards you can start right now is shown on /deploy.

If a PyTorch build fails on the 5090 with “no kernel image is available”, see how to fix CUDA “no kernel image” on the RTX 5090. The 5090 needs CUDA 12.8 or newer.

FAQ

Is the RTX 5090 faster than the RTX 4090 for AI?

For LLM token generation and batched image generation, yes. Published tests show about 1.3 to 1.6 times the speed. For LLM prompt processing and single small images, the published results are about equal, and in some tests the 5090 is slower.

Is the RTX 5090 faster for Wan video generation?

We found no independent like-for-like test, so we don’t state a figure. On an RTX 4090 we measured 169–175 s per 5-second 720p Wan 2.2 TI2V-5B clip and 354–361 s per 5-second 480p Wan 2.2 14B FP8 clip in ComfyUI.

Does Wan 2.2 14B fit on a 24 GB RTX 4090?

In ComfyUI with the FP8 scaled weights, at 832×480 and 81 frames, it did in our test, peaking at 21.0 GB. The official Wan2.2 script asks for 80 GB for the 14B models.

How much VRAM does the RTX 5090 have?

32 GB of GDDR7, against 24 GB of GDDR6X on the RTX 4090.

Do I need a special PyTorch build for the RTX 5090?

Yes. You need a build for CUDA 12.8 or newer, because older builds have no kernels for the 5090’s architecture.

Sources