The RTX 5090 and RTX 5080 are the two cards I get asked about most for local AI, and in November 2026 the choice is stranger than ever. At launch the 5090 cost twice as much as the 5080. Today the cheapest new 5090 is around $6,900 while a 5080 is around $2,000, so you’re paying more than three times as much for double the VRAM. I run two RTX 5080s myself, so I’ll be honest about where 16 GB is enough, where it hurts, and whether a second 5080 is a smarter buy than one 5090.
- RTX 5090: 32 GB GDDR7, 1,792 GB/s, 21,760 CUDA cores, 575 W. Roughly 1.8x a 5080 for LLM token speed and 1.6-1.9x for image and video generation (estimates from bandwidth and core counts). Runs 27B-32B LLMs fully on the GPU and big video models without heavy offloading.
- RTX 5080: 16 GB GDDR7, 960 GB/s, 10,752 CUDA cores, 360 W. Great for 8B-20B LLMs, MoE models with CPU offload, and image models in FP8 or GGUF. Video needs GGUF and patience.
- Prices (US, Nov 2026): 5090 from about $6,900 (MSRP $1,999); 5080 about $2,000 (MSRP $999).
- My verdict: for most people the 5080 (or renting a 5090 at about $0.50-$1.20/hour) is the better deal right now. Buy a 5090 only if 32 GB on one card earns you money every week.
- Two 5080s (about $4,000) give 32 GB for LLMs split across cards, but image and video models still see 16 GB per GPU.
Specs Side by Side
| Spec | RTX 5090 | RTX 5080 | 5090 advantage |
|---|---|---|---|
| Architecture | Blackwell (GB202) | Blackwell (GB203) | – |
| CUDA cores | 21,760 | 10,752 | 2.0x |
| VRAM | 32 GB GDDR7 | 16 GB GDDR7 | 2.0x |
| Memory bus | 512-bit | 256-bit | 2.0x |
| Memory bandwidth | 1,792 GB/s | 960 GB/s | 1.87x |
| AI TOPS (FP4, NVIDIA figure) | 3,352 | 1,801 | 1.86x |
| Total graphics power | 575 W | 360 W | +215 W |
| Recommended PSU | 1,000 W | 850 W | – |
| Power connector | 1x 16-pin (12V-2×6) | 1x 16-pin (12V-2×6) | – |
| Launch MSRP | $1,999 | $999 | 2.0x |
Both support FP4 and FP8 tensor maths, which newer ComfyUI and inference engines use to fit bigger models in less memory. Both need recent PyTorch builds with CUDA 12.8 or newer; I cover the exact setup in ComfyUI on RTX 50 series.
November 2026 Prices
| RTX 5090 | RTX 5080 | |
|---|---|---|
| Launch MSRP | $1,999 | $999 |
| Cheapest new (US, early Nov 2026) | about $6,900 (Newegg from $6,919) | about $2,000 (premium models up to about $2,400) |
| Price per GB of VRAM | about $216 | about $125 |
| Cloud rental | about $0.51-$1.19 per hour | rarely offered; 4090s rent for $0.35-$0.89 |
Why the gap? AI data centres are buying most of the world’s new memory, and the 5090 is the cheapest way to get 32 GB of very fast memory on one card, so AI labs and resellers compete with gamers for it. Reports in October 2026 even showed 5090s being stripped for AI servers in China. Board partners raised prices again in early November. The full story is in why GPU prices are so high in 2026.
Local LLMs: What Fits and How Fast
For LLMs, VRAM decides what fits and memory bandwidth decides speed. Each generated token reads the whole active model from memory, so the 5090’s 1.87x bandwidth translates almost directly into about 1.8x faster generation for any model that fits on both cards.
| Model (Q4_K_M GGUF unless noted) | Approx. size | RTX 5080 16 GB | RTX 5090 32 GB |
|---|---|---|---|
| 8B-9B (Qwen3, Llama, MiMo-V2.6 9B) | 5-6 GB | Fits with long context | Fits, about 1.8x faster |
| 14B | 9 GB | Fits, moderate context | Fits with long context |
| 20B-24B | 12-14 GB | Fits with short context | Fits comfortably |
| 27B (Qwen3.8-27B, Gemma 4 31B at Q3/Q4) | 16-19 GB | Needs Q3 or partial CPU offload | Fits fully |
| 32B dense | 19-20 GB | Partial offload, slow | Fits with 16-32K context |
| Big MoE (120B-class with few active params) | 60+ GB | Runs with experts in system RAM | Runs with experts in RAM, faster |
| 70B dense | 40+ GB | Mostly CPU, very slow | Partial offload, usable |
The practical difference: on a 5080 I plan around 14B-24B dense models or MoE models with RAM offload. On a 5090 the 27B-32B tier, which is where a lot of the best open coding and reasoning models sit, runs entirely on the GPU. If that tier is your daily driver, the 5090 is a real upgrade, not just a faster card. My dedicated guide on running Qwen3.8-27B on a 16 GB GPU shows what the compromise looks like on a 5080.
Image and Video Generation
Diffusion models are more compute-bound than LLMs, so the 5090’s doubled core count matters as much as its memory. Expect roughly 1.6x to 1.9x faster generation per image when the model fits on both cards (estimate based on core counts, clocks and bandwidth; real numbers vary by model and attention backend).
| Workload | RTX 5080 16 GB | RTX 5090 32 GB |
|---|---|---|
| SDXL, FLUX-class and Qwen-Image in FP8 | Good, with text encoders offloaded | Excellent, everything stays loaded |
| Full BF16 image models (20B+ parameters) | GGUF Q8/Q6 or FP8 needed | Often fits in BF16 |
| Wan 2.2 / LTX video at 720p | GGUF quants and block swapping, slow | FP8, far fewer workarounds |
| Long or high-res video, upscaling (SeedVR2) | Tight, frequent out-of-memory | Comfortable |
| Batching (4+ images at once) | Limited | Easy |
On 16 GB, my go-to tricks are GGUF models (ComfyUI-GGUF guide), SageAttention, and putting the text encoder on a second GPU. They work, but each one is time spent fiddling. With 32 GB most workflows load as the author shared them.
LoRA Training and Fine-Tuning
- Image LoRAs (SDXL, FLUX-class, Qwen-Image): both cards can train with FP8/quantised base models. The 5090 allows higher resolution, bigger batches and full-precision bases.
- Video LoRAs (Wan 2.2): realistic on a 5090; on a 5080 you’ll rely on heavy offloading and low resolutions.
- LLM fine-tunes (QLoRA with Unsloth): a 5080 handles up to about 14B in 4-bit; a 5090 handles roughly 27B-32B in 4-bit.
Two RTX 5080s vs One RTX 5090
This is my own setup, so here’s the honest trade-off.
| 2x RTX 5080 | 1x RTX 5090 | |
|---|---|---|
| Price (Nov 2026) | about $4,000 | about $6,900 |
| Total VRAM | 32 GB (2 x 16 GB) | 32 GB on one card |
| LLMs | llama.cpp, Ollama and LM Studio split layers across both cards; works well | Simpler, a bit faster for single requests |
| Image and video models | Each model must fit in 16 GB; you can put text encoder or VAE on GPU 2 | Whole pipeline on one GPU |
| Parallel jobs | Two ComfyUI instances at once, one per card | One job at a time |
| Power | about 720 W for GPUs | about 575 W |
| Case and board | Needs two x16-size slots with spacing, more airflow | One very large card |
Two 5080s make sense if you mostly run LLMs or want two image jobs at once. One 5090 is better if your main work is big image or video models that need all 32 GB in one place.
Power, PSU and Cooling
- RTX 5090: use a quality 1,000 W or larger ATX 3.1 PSU with a native 12V-2×6 cable. Push the connector fully home and avoid tight bends near the plug. Many owners power-limit to 70-80% with almost no loss in AI throughput.
- RTX 5080: an 850 W PSU is enough for one card. For two, I’d use 1,200 W or more.
- Case: 5090 coolers are often 3.5-4 slots and over 330 mm long. Check clearance before buying.
Rent a 5090 Instead?
At about $0.51-$1.19 per hour on Vast.ai and RunPod, a $6,900 5090 takes roughly 5,800-7,800 rented hours to break even. That’s years of evening use. For occasional big video jobs or training runs, renting is clearly cheaper; I compare hosts in the cheapest cloud GPU for AI.
Which Should You Buy?
- Buy the RTX 5080 if you run 8B-24B LLMs, generate images in FP8/GGUF, and can live with some video workarounds. It’s the sensible card at today’s prices.
- Buy the RTX 5090 if you need 27B-32B models fully on GPU, generate a lot of video, train video LoRAs, or bill clients for GPU time.
- Wait if you can: rumours point to a 24 GB RTX 5080-class refresh in Q1 2027, though it has slipped several times.
- Look elsewhere if you only need memory capacity for LLMs: a used RTX 3090 (24 GB) or a 32 GB Radeon AI PRO R9700 costs far less. See is a used RTX 3090 still worth it.
Common Problems on RTX 50 Cards
“CUDA error: no kernel image is available”
Your PyTorch build is too old for Blackwell. Install a build for CUDA 12.8 or newer (for example the cu128 or cu130 wheels).
Model runs slowly even though it should fit
Check Task Manager or nvidia-smi: if dedicated memory is full, the model is spilling into shared system memory. Set CUDA – Sysmem Fallback Policy to Prefer No Sysmem Fallback in the NVIDIA Control Panel to get a clear out-of-memory error instead. My VRAM checking guide shows how.
Black screen or crash under AI load
Update to the latest NVIDIA driver, reseat the 16-pin cable, and try a 75-80% power limit. Persistent crashes at stock settings are worth an RMA.
Second GPU not used in ComfyUI
ComfyUI uses one GPU per instance. Start a second instance with --cuda-device 1 --port 8189, or use multi-GPU custom nodes to place the text encoder on the other card.
Frequently Asked Questions
Is the RTX 5090 worth it over the RTX 5080 for AI?
Only if you need 32 GB on one card, for 27B-32B LLMs, big video models or video LoRA training. At about $6,900 vs $2,000 in November 2026, most people get better value from a 5080 or from renting a 5090 by the hour.
How much faster is the RTX 5090 than the RTX 5080 for AI?
Roughly 1.8x for LLM token generation (from 1.87x memory bandwidth) and about 1.6-1.9x for image and video generation when the model fits on both cards. These are estimates, real results vary by model and software.
Is 16 GB of VRAM enough for AI in 2026?
It’s enough for 8B-24B LLMs, MoE models with RAM offload and most image models in FP8 or GGUF. It is tight for 27B-32B dense models and for video generation.
Can I combine two RTX 5080s to get 32 GB?
For LLMs yes, because llama.cpp, Ollama and LM Studio split layers across GPUs. For image and video models no, each model must fit in one card’s 16 GB, though you can move parts like the text encoder to the second GPU.
Why is the RTX 5090 so expensive right now?
AI demand and a global memory shortage. The 5090 is the cheapest single card with 32 GB of fast GDDR7, so AI buyers compete with gamers, and board partners raised prices several times in 2026.
What PSU do I need for an RTX 5090?
NVIDIA recommends 1,000 W. Use a quality ATX 3.1 unit with a native 12V-2×6 cable, and make sure the connector is fully seated.
Should I wait for an RTX 5080 Super?
Rumours point to a 24 GB 5080-class card in Q1 2027, but it has been delayed several times and prices are rising. Wait only if your current GPU still does the job.
Final Thoughts
At MSRP, the RTX 5090 was the easy recommendation for serious local AI. At more than three times the 5080’s street price, it’s a business purchase, not a hobby one. My own two 5080s handle LLMs with layer splitting and run two ComfyUI jobs side by side; the only time I miss a 5090 is for long video. If that sounds like your work, rent a 5090 for those jobs and keep the change.
More hardware help: best GPU for local LLMs, how much VRAM do I need for AI and best workstation GPU for AI.