Is a 32GB RTX 4080 SUPER enough for FLUX, SDXL and 14B-32B LLMs?

Short answer: For most image work and for LLMs up to about 32B at 4-bit, yes. The RTX 4080 SUPER 32G is the RTX 4080 SUPER GPU with 32 GB of memory; CloudGPU lists it at 32 GB in /gpu-prices.json (read 2026-11-04), while NVIDIA’s own RTX 4080 SUPER has 16 GB (NVIDIA RTX 4080 family page, read 2026-11-04). SDXL (about 7 GB) and the FP8 FLUX.1-dev checkpoint (17.25 GB file) fit with room to spare. The full 16-bit FLUX.1-dev pipeline (about 31.4 GB loaded in BF16) is at the limit. Qwen3 14B at FP8 (16.3 GB of weights) and Qwen3 32B at 4-bit (about 19-20 GB) fit with room for context; Qwen3 32B at FP8 (34.3 GB) and any 70B model do not.

The card

RTX 4080 SUPER 32G RTX 4090 RTX 5090
Architecture Ada Lovelace Ada Lovelace Blackwell
VRAM 32 GB (CloudGPU listing) 24 GB GDDR6X 32 GB GDDR7
Memory interface 256-bit (NVIDIA’s RTX 4080 SUPER) 384-bit 512-bit
CUDA cores 10,240 16,384 21,760
AI TOPS (NVIDIA) 836 1,321 3,352
Compute capability 8.9 (NVIDIA lists the RTX 4080 at 8.9) 8.9 12.0

Sources: NVIDIA’s RTX 4080 family, RTX 4090 and RTX 5090 pages and the CUDA GPUs list, read 2026-11-04. NVIDIA does not sell a 32 GB RTX 4080 SUPER, so the VRAM figure is CloudGPU’s listing, not an NVIDIA spec. Check what your instance reports:

nvidia-smi --query-gpu=name,memory.total,compute_cap --format=csv

What this means in practice: 32 GB decides what fits; the narrower memory bus and fewer CUDA cores mean you should expect fewer images per minute and fewer tokens per second than on an RTX 4090 or RTX 5090 running the same model. Time your own workload before you commit to a long run.

Image generation: what fits in 32 GB

Figures are file sizes and published measurements, not CloudGPU benchmarks. VRAM need depends on model, precision and workflow.

Workload What it needs On 32 GB
SDXL about 7 GB FP16 weights fits, with room for ControlNet, LoRAs and an upscaler
FLUX.1-dev FP8 checkpoint 17.25 GB file (Comfy-Org); about 12-16 GB of weights fits, with room for LoRAs and larger batches
FLUX.1-dev full 16-bit about 31.4 GB loaded in BF16 at the limit; let ComfyUI offload the T5 text encoder, or set weight_dtype to fp8
Wan 2.2 TI2V-5B ComfyUI: “should fit well on 8GB vram” with native offloading fits
Wan 2.2 14B (T2V / I2V) two 14.29 GB FP8 experts plus a 6.74 GB FP8 text encoder relies on offloading; see the 48 GB options

Sources: How much VRAM do I need for FLUX and SDXL?, the Hugging Face Diffusers blog (FLUX.1-dev about 31.4 GB in BF16), the Comfy-Org/Flux-1-Dev and Comfy-Org/Wan-2.2 file listings and the ComfyUI Wan 2.2 tutorial, read 2026-11-04.

Ada Lovelace GPUs run FP8 checkpoints, so the FP8 FLUX file is the practical default on this card. For a full walk-through, see how to run FLUX.1-dev in ComfyUI on a cloud GPU.

LLMs: what fits in 32 GB

Weights are Hugging Face file sizes (read 2026-11-04). KV cache per token is from each model’s config.json, as worked out in How context length adds to VRAM use; GB there means GiB.

Model and precision Weights KV cache (FP16) On 32 GB
Qwen3 14B, FP8 (Qwen/Qwen3-14B-FP8) 16.3 GB 160 KiB per token, about 5 GB at 32k fits, with the native 32k context
Qwen3 14B, BF16 about 29.5 GB as above tight; little room for context
Qwen3 32B, 4-bit AWQ (Qwen/Qwen3-32B-AWQ) or Q4_K_M about 19-20 GB 256 KiB per token, about 8 GB at 32k fits; 16k-32k context is comfortable
Qwen3 32B, FP8 (Qwen/Qwen3-32B-FP8) 34.3 GB does not fit
Llama 3.3 70B, Q4_K_M about 40-43 GB does not fit

On the 32B rows, the KV-cache post’s 32 GB table (written for the RTX 5090, which has the same 32 GB) gives a ceiling of about 45k tokens for one sequence at Q4_K_M, above Qwen3’s native 32k. Plan below the ceiling: runtimes reserve extra buffers, and parallel requests multiply the cache.

Commands

Ollama (Ollama template, or a Raw GPU with Ollama installed). The qwen3:32b tag is a 4-bit build of about 20 GB in the Ollama library (read 2026-11-04):

export OLLAMA_FLASH_ATTENTION=1
export OLLAMA_KV_CACHE_TYPE=q8_0        # 8-bit KV cache: about twice the context in the same memory
export OLLAMA_CONTEXT_LENGTH=16384
ollama serve &
ollama run qwen3:32b

vLLM (Raw GPU). Install it as in how to serve an OpenAI-compatible API with vLLM, then:

# Qwen3 32B at 4-bit, context capped so the KV cache fits
vllm serve Qwen/Qwen3-32B-AWQ --max-model-len 16384 --gpu-memory-utilization 0.90

# Qwen3 14B at FP8; vLLM runs FP8 W8A8 on compute capability 8.9 and newer
vllm serve Qwen/Qwen3-14B-FP8 --max-model-len 32768

The vLLM docs state that “FP8 computation is supported on NVIDIA GPUs with compute capability >= 8.9 (Ada Lovelace, Hopper, Blackwell)” (vLLM FP8 W8A8 docs, read 2026-11-04).

Watch peak memory while you test your real prompt length and batch size:

nvidia-smi --query-gpu=memory.used,memory.total,utilization.gpu --format=csv -l 1

If you run out of memory, lower the context first, then quantise the KV cache; see how to fix CUDA out of memory.

What a test run costs

At 2026-11-04 19:05 (UTC+8) the RTX 4080 SUPER 32G was $0.33/hour. Billing is per minute, rounded up, with a 1-minute minimum, so a 37-minute test costs $0.204 and $10 of balance covers about 30 hours. The instance is billed while models download, so download once and keep the instance until you are done. Destroying it stops billing; current rates are on /pricing.

How to rent one

  1. On /deploy, choose a template (ComfyUI, Ollama, JupyterLab + Ollama, LLaMA-Factory) or Raw GPU, then the RTX 4080 SUPER 32G. Which cards can be started right now is shown on /deploy. /deploy also shows the vCPU and memory that come with each card.
  2. Download the weights you need; check the file sizes first, because instance storage is limited (how to speed up model downloads).
  3. Copy your outputs off before you destroy the instance: destroying it deletes everything stored on it.

Frequently asked questions

Can a 32GB RTX 4080 SUPER run FLUX.1-dev?

Yes. The FP8 checkpoint (17.25 GB file) fits with room for LoRAs. The full 16-bit pipeline is about 31.4 GB loaded in BF16, which is at the limit of 32 GB, so let ComfyUI offload the text encoder or use fp8 weight_dtype.

Can it run a 32B LLM?

Yes at 4-bit: Qwen3 32B AWQ or Q4_K_M weights are about 19-20 GB, which leaves room for a 16k-32k context. The FP8 version of Qwen3 32B is 34.3 GB and does not fit.

Can it run a 70B model?

Not on one card. A 4-bit Q4_K_M 70B is about 40-43 GB of weights. Look at a 48 GB card such as the RTX 4090 48G or RTX 4090D 48G; see which 48GB GPU to rent for 70B models.

Is it as fast as an RTX 4090 or RTX 5090?

No. It has the same 32 GB as an RTX 5090, but NVIDIA’s RTX 4080 SUPER has 10,240 CUDA cores and a 256-bit memory bus, against 16,384 and 384-bit on the RTX 4090 and 21,760 and 512-bit on the RTX 5090. It is a card for fitting a model in 32 GB, not for speed; time your own workload.

Does it support FP8?

Yes. It is an Ada Lovelace GPU (compute capability 8.9), and vLLM runs FP8 W8A8 on compute capability 8.9 and newer. ComfyUI’s FP8 FLUX checkpoints also run on it.

Sources (read 2026-11-04, Asia/Shanghai)

The Deploy page lists the cards you can start. Rates on this page are read from CloudGPU’s live price list when the page is rendered; /pricing has the current rate for every card.

Last updated: 2026-11-04 (Asia/Shanghai).