5090 vs 4090 for AI: Reading a Generational Comparison Correctly

5090 vs 4090 for AI in the LynxBenchAI 26Q3 results: the RTX 5090 leads in Training and Inference, the RTX 4090 in Compute. How to read the gap.

5090 vs 4090 for AI: Reading a Generational Comparison Correctly
Written by TechnoLynx Published on 24 Aug 2026

In the LynxBenchAI 26Q3 results, the RTX 5090 scores higher than the RTX 4090 overall (GT 597.5 against 450.0) and in Training and Inference, while the RTX 4090 scores higher in Compute. So the 5090 vs 4090 answer from this release is a generational lead in two categories and a deficit in the third, and which card the evidence favours for you depends on which of those categories your work resembles.

In the 26Q3 results, the categories point this way:

  • Training-like work (training loops in FP32 and BF16): the RTX 5090 scores 1.53x the RTX 4090 on the LynxBenchAI 26Q3 Training Index.
  • Inference-like work (forward passes in FP16, INT8 and FP8): the RTX 5090 scores 1.49x the RTX 4090 on the 26Q3 Inference Index.
  • Compute-like work (raw matrix kernels, especially sparse matrix multiplication): the RTX 4090 scores higher, with the RTX 5090 at 0.76x on the 26Q3 Compute Index.

Where Each Card Scores Higher Across the 26Q3 Categories

Score (26Q3) RTX 4090 (native Linux) RTX 5090 (native Linux) RTX 5090 (WSL) RTX 5090 native ÷ RTX 4090
GT (overall) 450.0 597.5 545.7 GT is ordinal, so no ratio is read from it
Training 335.9 515.3 470.3 515.3 ÷ 335.9 = 1.53x
Inference 659.4 985.8 891.8 985.8 ÷ 659.4 = 1.49x
Compute 367.9 279.6 261.4 279.6 ÷ 367.9 = 0.76x

Each figure above is read from the published 26Q3 results. The RTX 4090 result page, the RTX 5090 result page and the RTX 5090 under WSL result page carry each device’s 26Q3 result, and the side-by-side RTX 5090 vs RTX 4090 comparison page holds the pair’s entry.

The WSL row is worth a second look. The same RTX 5090 run under WSL scores lower than its native Linux run in every category score, yet it still sits above the RTX 4090 in Training and Inference and below it in Compute. LynxBenchAI binds every result to the device plus the backend, driver, framework and runtime on the machine that produced it, so “the RTX 5090” is two different executors in this table, and no score belongs to the silicon alone.

How Far You Can Rely on a 26Q3 Score for Your Own Work

Each 26Q3 score reflects the model catalogue that release runs on every device, not your application. Each test is one timed measurement window after a discarded warm-up, not a median across repeated trials. GT is an ordinal aggregate, so a gap between two GT values is not a percentage.

Scores are comparable only within one release name, so 26Q3 numbers are read against 26Q3 numbers only. LynxBenchAI benchmarks the catalogue with stock PyTorch.

Reading the Three Category Scores as Separate Statements About Each Card

The comparison becomes three separate statements about two executors, each readable on its own: how well each handles training-style work, inference-style work and raw compute kernels. GT is the overall figure aggregated from them, and the category scores show where that figure comes from. Compute goes the other way from Training and Inference.

The per-test scores locate the Compute deficit. On Dense MatMul the RTX 5090 scores higher at every precision tested (FP32 357.9 against 239.9, FP16 421.6 against 295.4, BF16 427.5 against 283.8, FP64 10.1 against 9.5). On Sparse MatMul the RTX 4090 scores higher, at FP32 598.5 against 301.7 and at FP64 666.2 against 211.9. The 26Q3 data shows where the gap is, not why. Each test is a single timed window, so a gap like this is where a diagnosis starts, not where it ends.

Inside the categories where the RTX 5090 leads, the size of the lead varies from model to model. In 26Q3 Inference, MobileNetV3 at FP16 scores 2622.5 on the RTX 5090 against 1519.0 on the RTX 4090 (2622.5 ÷ 1519.0 = 1.73x), while DistilGPT2 at FP16 scores 1176.8 against 957.0 (1176.8 ÷ 957.0 = 1.23x). In Training, GraphGPS at BF16 scores 280.7 against 151.2 (1.86x), and MobileNetV3 at BF16 scores 313.4 against 268.3 (1.17x).

A single generational headline would fold all of that into one number and hide the case that matters most: a device that is strong at one class of work and weaker at another.

Why the Spec-Sheet Jump Does Not Carry into Every Category

On paper the generational jump is large. The RTX 5090 has 170 streaming multiprocessors against the RTX 4090’s 128, about 33% more cores. It has 680 fifth-generation Tensor Cores against 512 fourth-generation ones, which Puget Systems translates into dense FP16/BF16 throughput rising about 27%, from 165.2 to 209.5 TFLOPS. Memory goes from 24 GB to 32 GB, and bandwidth rises from 1.01 TB/s to 1.79 TB/s with the move from GDDR6X to GDDR7.

The 26Q3 per-test spread runs from 0.76x in Compute up to 1.86x on one training test, while every spec delta above is an increase. Different workloads are limited by different resources: Puget Systems points out that prompt processing is compute-bound while token generation is bound by memory bandwidth.

Tensor TFLOPS on a GeForce Spec Sheet Need Careful Reading

The headline tensor figures need two qualifiers. Tom’s Hardware quotes the RTX 4090 at 661 TFLOPS of FP16 tensor compute with sparsity, and NVIDIA’s own Ada whitepaper rates the RTX 4090 at roughly 165 TFLOPS for FP16 matrix math with FP32 accumulation, half of what the same die manages with FP16 accumulation. FP32 accumulation is the mode mixed-precision training uses, and NVIDIA’s whitepapers rate GeForce cards at half speed in that mode while the professional cards on the same chips run both modes at the same rate. A measurement repository dated October 2026 found three stock GeForce cards ran 32-bit sums at 0.51 to 0.59 of their 16-bit speed per clock cycle, against 0.85 to 1.06 for the professional cards, though its author notes these are matrix multiplications on rented hosts, with two cards run once.

When you compare tensor TFLOPS across the two generations, check whether a figure assumes sparsity and which accumulation mode it uses before treating it as the number your training job will see.

Blackwell Software Maturity Moves Early RTX 5090 Results

AI performance is a property of hardware and software together, which is why identical GPUs often perform differently. A 2026 study of consumer LLM inference found the RTX 4090 consistently delivered the lowest time-to-first-token (8 ms against the RTX 5090’s 18 ms) despite the RTX 5090’s higher throughput, and suggested software maturity, with TensorRT-LLM still being optimised for Blackwell, rather than hardware as the cause. Puget Systems’ launch review saw RTX 5090 prompt processing in Windows llama.cpp land near the RTX 4080 SUPER and suspected a driver issue; a later Puget article reports that the issue was resolved in llama.cpp build 5203. The same launch review found the RTX 5090 often unable to keep up with the RTX 4090 in Topaz Video AI.

A published RTX 5090 result therefore describes an executor at a point in its software life, and the 26Q3 rows carry the stack they ran on for exactly that reason.

Does 32 GB Against 24 GB Change Which Models You Can Run?

Yes, and it is a separate question from speed. Capacity decides whether a model fits on the card at all. Category scores describe how fast a card runs the work it can hold. Even the RTX 5090’s 32 GB is a hard limit: models such as Llama-3.3-70B exceed it and must either be quantised heavily or offloaded to system RAM over PCIe, and that study reports CPU offloading cutting throughput by over 90% compared with running fully on the GPU. The same study found that quantising a 70B model to NVFP4 needs peak memory far beyond 32 GB, so even running such a model on the card depends on pre-quantised weights or slow CPU-side quantisation.

Precision changes the arithmetic of fit. NVIDIA reports that FP4 with its low-VRAM mode brings the full FLUX pipeline down to 11.1 GB, small enough for GPUs with less than 24 GB. Every 26Q3 catalogue test produced a non-zero score on both cards. Check first whether the model fits at the precision you will actually run, and only then which category score applies.

What Has to Be True Before an RTX 4090 Result and an RTX 5090 Result Can Be Fairly Compared?

A fair comparison needs every condition below to hold.

  1. Same release name. A 26Q3 result is read against 26Q3 results only.
  2. Same model, same precision. The model architecture and numerical precision are held constant on both devices.
  3. Workload scaled to each device’s saturation. LynxBenchAI raises the batch size on each device until throughput saturates, so neither card is measured at a load that happens to flatter it.
  4. A recorded, comparable software stack. Backend, driver, framework and runtime are bound to each result.
  5. Bounded, identical optimisation. The same prepared artefacts run on both cards under the same bounded optimisation effort, and a precision that produces no result is shown as 0.0.
  6. A precision both devices can execute. FP4 compute arrives with Blackwell’s fifth-generation Tensor Cores on the RTX 50 series, and both cards produced results at every 26Q3 inference precision: FP16, INT8 and FP8.

Condition 5 is what turns a same-vendor generational comparison into something you can check rather than a marketing claim: the same artefacts, under the same constraints, ran on both cards. Condition 6 is where NVIDIA’s own Flux.dev comparison breaks. Putting an RTX 5090 at FP4 against an RTX 4090 at FP8 compares two precisions as much as two cards; the A100 vs H100 and FP8 comparison raises the same question.

“Up to 2x”: Why NVIDIA’s Generational Claim Answers a Different Question

NVIDIA’s launch announcement states that Blackwell architecture innovations and DLSS 4 let the RTX 5090 outperform the RTX 4090 by up to 2x. That is a vendor claim, not a test result, and it bundles a rendering feature exclusive to the new generation into the comparison. Several technology journalists criticised RTX 50 claims for relying on Multi Frame Generation. Tom’s Hardware, reading NVIDIA’s own preview, estimated about 27% more performance in Far Cry 6 and about 43% in A Plague Tale: Requiem where the new AI features were not involved, and noted that the 2x Flux.dev jump came with the RTX 4090 at FP8 and the RTX 5090 at FP4.

NVIDIA’s own FLUX table lets you separate the two effects. FLUX.1-dev took 10,620.37 ms on the RTX 4090 at FP8, 6,680.93 ms on the RTX 5090 at FP8 and 3,852.75 ms on the RTX 5090 at FP4. At matched FP8 that is 10,620.37 ÷ 6,680.93 = 1.59x; the RTX 5090 at FP4 against the RTX 4090 at FP8 is 10,620.37 ÷ 3,852.75 = 2.76x. The headline ratio mixes a 1.59x gain at matched precision with a change of precision.

Multi-GPU results answer yet another question. In CloudRift’s vLLM serving benchmark, two RTX 5090s reached 1,230 tokens/s against 467 for two RTX 4090s on a model needing two cards, which the author attributes partly to PCIe Gen 5 against Gen 4 inter-GPU bandwidth. That benchmark targets maximum throughput under high concurrency, and the 26Q3 rows above are single-device results.

A same-release LynxBenchAI comparison is a different kind of claim from any of these: identical catalogue, identical precisions, identical prepared artefacts. Runs submit to a public leaderboard, so a claim about either card can be checked against runs other people produced with the same instrument under the same release.

When Is an RTX 5090 Upgrade over an Existing RTX 4090 Actually Justified by the Published Categories?

The upgrade is backed by evidence when the 26Q3 category that resembles your work favours the RTX 5090, your model fits and runs at a precision both cards share, and your software stack already supports Blackwell. It is only assumed when the case rests on a spec delta or a vendor multiplier.

Your situation What the evidence supports
Work resembles the Training category (FP32/BF16 training loops) 26Q3 scores the RTX 5090 at 1.53x on the Training Index, with per-test gains from 1.17x to 1.86x
Work resembles the Inference category (FP16, INT8, FP8) 26Q3 scores the RTX 5090 at 1.49x on the Inference Index, with model-to-model spread
Work is dominated by sparse matrix kernels 26Q3 Compute favours the RTX 4090 (0.76x for the RTX 5090), driven by Sparse MatMul
Model does not fit in 24 GB but fits in 32 GB Capacity, not a category score, is the deciding fact
Case rests on FP4 throughput FP4 is not among the 26Q3 test precisions, so these results neither confirm nor refute it
Case rests on “up to 2x” NVIDIA attributes that figure partly to DLSS 4
Latency-critical serving on TensorRT-LLM One study measured lower TTFT on the RTX 4090 (8 ms against 18 ms); check your stack’s current version

Puget Systems, in its RTX 5090 AI review, said that until software support improved it was hard to say definitively the average 40-series owner would benefit significantly from an upgrade, and that this would almost certainly change with time. Treat that as a point-in-time assessment, which is how any early verdict on a new architecture should be read.

Measuring Your Own Workload on Both Cards When It Falls Outside the Catalogue

Measure it rather than extrapolate. A category score is evidence about the catalogue that produced it, and nothing guarantees your workload scales the way the catalogue did on either card. The 0.76x to 1.86x spread inside 26Q3 between these two cards shows how far a borrowed ratio can miss.

LynxBenchAI’s Personal Edition measures your card and software stack against the 26Q3 catalogue; it installs with pip install lynxbench-ai on Python 3.11 or later, on Linux or Windows via WSL2, needs roughly 15 GB of free disk, downloads about 2 GB of model data once and takes 15 to 30 minutes, and it is free for non-commercial use. For the workload itself, apply the fairness checklist above to your own runs on both cards: same model, same precision, batch raised to saturation on each, the same recorded stack and the same optimisation effort. TechnoLynx’s GPU Performance Engineering scopes a 4090-to-5090 decision for a specific workload, including tuning and profiling a production pipeline. Every device with a published 26Q3 result is in the AI hardware directory of all LynxBenchAI results.

FAQ

How Much Better Is a 5090 Compared to a 4090?

In the 26Q3 results the RTX 5090 scores 1.53x the RTX 4090 on the Training Index and 1.49x on the Inference Index, and 0.76x on the Compute Index. GT ranks it higher (597.5 against 450.0), but GT is ordinal, so that gap is not a percentage.

Is a 4090 Better than a 5090 at Anything?

In 26Q3 Compute, yes: the RTX 4090 scores 367.9 against 279.6, driven by Sparse MatMul (FP32 598.5 against 301.7). One study also measured lower time-to-first-token on the RTX 4090 (8 ms against 18 ms) under TensorRT-LLM.

Is the GeForce RTX 5090 Good for AI?

The RTX 5090 posts the higher 26Q3 Training and Inference scores of the two cards (515.3 and 985.8 against 335.9 and 659.4) and carries 32 GB of GDDR7. Whether it suits your work depends on your category, your model’s fit and your stack’s Blackwell support.