RTX 5090 vs 4090 for Flux.2: Does 32GB Change It?

Flux.2 is 64GB at BF16, so both cards run the 4-bit build. The 5090 wins on bandwidth and on room left for ControlNet stacking after it loads.

Quick answer: Both cards run Flux.2 the same way — the 4-bit build with the text encoder offloaded, because its transformer is ~64GB at BF16 and ~32GB at FP8. The RTX 5090 is roughly 30-40% faster and has 8GB more left over once the model loads, which is what decides ControlNet stacking. The reason is narrow but real — Blackwell has native FP8 tensor cores that Flux.2 was explicitly tuned for, while the 4090 falls back to a software FP8 path on Ada. If Flux.2 is your daily driver, the 5090 is the right buy. If you’re a hybrid Flux.1 + occasional Flux.2 user, the 4090 still earns its keep.

Best for Flux.2

NVIDIA GeForce RTX 5090

32GB GDDR7

32GB GDDR7 and 1792 GB/s, ~9-11s per Flux.2 image at 1024px, with room left for ControlNet after the 4-bit model loads. The first consumer GPU built for the FP8 path NVIDIA recommends for Flux.2.

Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.

Who this guide is for

This is specifically for people deciding between the RTX 5090 and RTX 4090 with Flux.2 as the primary workload. If you want the broader head-to-head across all AI workloads (LLMs, SDXL, fine-tuning), my RTX 4090 vs RTX 5090 for AI comparison covers that — this article narrows the lens to Flux.2 only, because the per-workload gap is meaningfully different from the general-AI picture.

I’m also assuming you’ve already seen the best GPU for Flux.2 buyer guide and shortlisted these two flagships. If you’re still on Flux.1 and wondering whether the upgrade is worth the GPU jump, hold that question — I’ll get to it in the contrarian section.

Specs side-by-side

SpecRTX 4090RTX 5090
VRAM24GB GDDR6X32GB GDDR7
Memory bandwidth1,008 GB/s1,792 GB/s
TGP450W575W
ArchitectureAda LovelaceBlackwell
Compute capability8.910.0
FP8 tensor coresSoftware path (emulated via FP16/FP32)Native hardware FP8
Street price~$2,200~$4,900

One row in that table does most of the heavy lifting for Flux.2: the bandwidth jump from 1,008 to 1,792 GB/s. Flux.2 at 32B parameters is bandwidth-bound during the diffusion steps, so the 78% bandwidth uplift translates into a near-linear speedup on the parts of the pipeline that matter most.

The FP8 row is the one most people misread, in both directions. It is a real hardware difference — Ada (compute cap 8.9) supports FP8 only through emulation scheduled on the FP16/FP32 tensor units. Blackwell (compute cap 10.0) exposes FP8 directly as a first-class tensor op. On models that were trained or quantized with FP8 in mind — which is exactly what NVIDIA shipped for Flux.2 in May 2026 — Blackwell pulls ahead of Ada by more than the raw bandwidth gap alone would predict.

GPU VRAM Comparison (GB)
RTX 5090 32GB RTX 4090 24GB RTX 5080 16GB RTX 4070 Ti S 16GB RTX 5070 12GB RTX 4060 Ti 16GB RTX 4060 Ti 8G 8GB RTX 4060 8GB RTX 3060 12GB RX 7800 XT 16GB

Real Flux.2 gen times

Approximate ComfyUI figures for the 4-bit build with the text encoder offloaded, Euler sampler, no ControlNet unless noted, modelled from memory bandwidth (methodology) rather than measured, at 1024×1024 / 30 steps:

Flux.2 workloadRTX 4090RTX 50905090 advantage
Flux.2 Dev 4-bit, 1024×1024, 30 steps~14-16 s~9-11 s+35-40%
Flux.2 Schnell 4-bit, 1024×1024, 4 steps~2.8 s~1.9 s+35%
Flux.2 + 1 ControlNet (Depth)~18 s~12 s+33%
Flux.2 + 2 ControlNets (Depth + Canny)tight on 24GB, ~24 s with swapping~15 s, comfortable+37% (plus no swap risk)
Flux.2 LoRA training, rank 32, batch 1~22GB used, ~3.1 it/s~24GB used, ~4.3 it/s+38%
Flux.2 LoRA training, rank 64, batch 2OOM at 24GB~28GB used, ~3.6 it/s5090 only

A few things worth flagging in those numbers:

  • The 35-40% gap is bigger than the 25-30% you’d see on Flux.1. The main driver is bandwidth — 1,792 GB/s against 1,008 — and Flux.2 is bandwidth-bound during the diffusion steps. Note what is not driving it: neither card runs an FP8 checkpoint of Flux.2, because FP8 weights are ~32GB before the encoder. Both are on the 4-bit build, and Blackwell’s tensor units help the compute around it rather than directly.
  • The ControlNet stacking row is where 24GB starts to bite. The 4-bit transformer is already ~16GB, so two ControlNets on top push the 4090 into swapping. The 5090’s 32GB has another 8GB of headroom that keeps the whole pipeline resident.
  • Rank-64 LoRA training simply doesn’t fit on a 4090 at batch size 2 with Flux.2. You can drop to rank 32 and batch 1, but if you’re serious about LoRA training on Flux.2, the 5090 unlocks training configurations the 4090 just won’t run.
Check NVIDIA GeForce RTX 4090 on AmazonBuy on Shopee SG

Which should YOU buy?

  • You generate Flux.2 daily, stack ControlNets, train LoRAs: RTX 5090. The 32GB headroom is what the stacking actually needs, and the per-image speedup pays back fast if you iterate a lot.
  • You’re a Flux.2 hobbyist, single-ControlNet workflows, no training: RTX 4090 is honestly fine. 14-16s per image is iterable, and the $2,700 saving is most of a build, not a rounding error.
  • You do mostly Flux.1 and only occasional Flux.2: RTX 4090. Flux.1’s FP16 weights fit comfortably in 24GB and the bandwidth gap matters less on the smaller 12B model. The 5090 premium isn’t worth it for this profile — see my Flux.1 GPU guide for the broader Flux.1-first picture.
  • You’re considering Flux.2 vs SD 3.5 and aren’t sure yet: Flux.2 vs SD 3.5 hardware covers which model’s hardware footprint fits which budget — relevant if you haven’t fully committed to Flux.2 yet.
  • You bought a 5090 already and want sanity-check: my 6-month RTX 5090 retrospective covers what real users actually got from the upgrade across image-gen, LLMs, and training. Flux.2 is one of the workloads where the upgrade has aged the best.

The contrarian read: if you only run Flux.1, the 5090 isn’t worth it

Worth being direct about, because it is the most common version of this question. If your workflow is Flux.1 Dev (12B parameters, FP16-native), an RTX 4090 already runs it in ~6-8s per image with full 24GB headroom for ControlNets and IP-Adapter. The 5090 will cut that to ~4-5s, but you’re spending $2,700 to save a few seconds per image on a workload that doesn’t stress either VRAM or FP8. That’s a bad upgrade, and at this price it is not a close call.

The 5090’s case rests almost entirely on workloads that need the capacity — Flux.2 with ControlNet stacks, 30B+ LLMs, and rank-64+ LoRA training. If those aren’t on your roadmap, save the $2,700.

Common mistakes when picking between these two

  1. Buying the 5090 “to future-proof” for Flux.1 workflows. As above — if Flux.1 is your real workload, the 4090 already over-serves it. You’re paying $2,700 for headroom you won’t use until your model choice changes.
  2. Underestimating the PSU upgrade. The 5090 wants 575W TGP and NVIDIA officially recommends a 1000W PSU. If you’re on an 850W unit (typical 4090 build), add ~$150-200 for a proper PSU upgrade. That’s real money on top of the GPU premium.
  3. Expecting FP8 support to let either card load Flux.2 at FP8. It does not: FP8 weights are ~32GB before the text encoder. Native FP8 matters for compute throughput on models whose weights do fit — the 4090 emulates it through FP16 tensor units while the 5090 runs it natively. That’s why the 35-40% gap is bigger than raw bandwidth predicts. Don’t compare these GPUs on Flux.2 the same way you’d compare them on FP16 Flux.1.
  4. Ignoring the 32GB headroom on multi-ControlNet workflows. A 4090 can technically run Flux.2 + 2 ControlNets, but with the text encoder loaded it starts swapping to system RAM, and gen time balloons from 16s to ~24s. The 5090 keeps the whole pipeline resident in VRAM. If your workflow is “two ControlNets + IP-Adapter,” that’s the 5090’s strongest case.

Final verdict

Workload patternBetter buyWhy
Daily Flux.2 + ControlNet stacks + LoRA trainingRTX 509032GB headroom after a 16GB model + 1792 GB/s bandwidth
Single-ControlNet Flux.2, no trainingRTX 409014-16s per image is fine, save $2,700
Mixed Flux.1 / occasional Flux.2RTX 4090Flux.1 doesn’t stress either card; 5090 premium unjustified
Pure LLM use case (no image gen)See 4090 vs 5090 general AI guideDifferent bandwidth-vs-VRAM tradeoff
Best for Flux.2 Workloads

NVIDIA GeForce RTX 5090

32GB GDDR7

32GB GDDR7 + 1792 GB/s bandwidth, and 16GB still free once Flux.2 4-bit loads. The first consumer card built for the FP8 path Flux.2 was tuned for — 35-40% faster than the 4090 per image, with 8GB of ControlNet/LoRA headroom on top.

Affiliate links — we may earn a commission at no extra cost to you. Amazon ships globally; Shopee SG covers Singapore & ASEAN.

For Flux.2 specifically, the 5090 is one of the rare cases where a flagship-to-flagship upgrade actually earns its 25% price premium — but only if Flux.2 is your real workload, not just one of several models in your rotation.

Frequently asked questions

How much faster is the RTX 5090 than the RTX 4090 for Flux.2?

Roughly 30-40% faster per image on Flux.2 Dev at 1024×1024, both cards running the 4-bit build. The gap is bigger than on Flux.1 or SDXL mainly because Flux.2 is bandwidth-bound and the 5090 has 1,792 GB/s against 1,008, helped by Blackwell’s native tensor cores, while the 4090 runs FP8 through a software path on Ada. Real-world generation times come in around 9-11s on the 5090 versus 14-16s on the 4090 at 30 steps.

Can the RTX 4090 still run Flux.2?

Yes, for single-ControlNet workflows. 24GB holds the 4-bit build’s ~16GB with the text encoder offloaded and room for one ControlNet — FP8 weights would be ~32GB and fit neither card. Gen times around 14-16s per image are perfectly workable. Where the 4090 starts to struggle is when you stack two ControlNets plus IP-Adapter, or when you try LoRA training above rank 32 — the 24GB ceiling becomes the limiter, not the compute.

Is the RTX 5090 worth it if I’m still on Flux.1?

Probably not. Flux.1 Dev is a 12B-parameter FP16 model that the 4090 already runs in roughly 6-8 seconds with full headroom for ControlNets. The 5090 would shave that to around 4-5 seconds, but you’re paying $2,700 plus a likely PSU upgrade for marginal time savings on a workload that doesn’t stress either bandwidth or FP8. The 5090’s value case is built around Flux.2, larger LLMs, and rank-64+ LoRA training.

Do I need a new PSU to upgrade from a 4090 to a 5090?

Most likely. The 5090’s 575W TGP pushes NVIDIA’s official PSU recommendation to 1000W, while typical 4090 builds run an 850W unit. If you’re upgrading the GPU only and skipping the PSU, expect throttling under sustained Flux.2 batches. Budget another $150-200 for a quality 1000W or 1200W PSU on top of the $4,900 GPU cost.

What about Flux.2 LoRA training — can the 4090 do it?

The 4090 can train Flux.2 LoRAs at rank 32, batch size 1, using roughly 22GB of its 24GB. Rank 64 or batch size 2 will out-of-memory. The 5090’s 32GB handles rank 64 at batch 2 with about 28GB used and runs the iterations roughly 38% faster on top. If serious Flux.2 LoRA training is on your roadmap, the 5090 unlocks configurations the 4090 simply will not run.

Affiliate Disclosure: This article may contain affiliate links. If you purchase through these links, we may earn a commission at no extra cost to you. Learn more