For gaming alone, usually not at today's prices. For AI and heavy creative work, often yes, because 16 GB is a hard wall that no amount of speed gets around. The GeForce RTX 5090 has twice the memory of the RTX 5080 (32 GB vs 16 GB), almost twice the memory bandwidth, and in Tom's Hardware's June 2026 tests it was about 43% faster at native 4K. At US launch prices it cost twice as much. At late-2026 US street prices it costs roughly three and a half to five times as much, and that changes the maths for most gamers.
How do the two cards compare on paper and in tests?
| Measure | RTX 5090 | RTX 5080 | Difference |
|---|---|---|---|
| GPU / CUDA cores | GB202, 21,760 | GB203, 10,752 | 2.0x cores |
| Memory | 32 GB GDDR7, 512-bit | 16 GB GDDR7, 256-bit | 2x capacity |
| Memory bandwidth | 1,792 GB/s | 960 GB/s | 1.87x |
| TGP / NVIDIA required system power | 575 W / 1,000 W | 360 W / 850 W | +215 W at the card |
| 1440p Ultra raster, native (fps) | 167.3 | 128.3 | +30% |
| 4K Ultra raster, native (fps) | 110.8 | 77.4 | +43% |
| 4K ray tracing, native (fps) | 64.1 | 42.1 | +52% |
| Blender 5.2.0 Open Data median (Cycles OptiX) | 16,080 | 9,148 | 1.76x |
| gpt-oss-20b, llama.cpp token generation (tok/s) | 282.5 | 204.9 | +38% |
| Qwen3 32B at 4-bit, llama.cpp (tok/s) | 50.9 | Doesn't fit | Only the 5090 runs it |
| US launch MSRP | US$1,999 | US$999 | 2.0x |
Source: NVIDIA spec pages and Tom's Hardware review spec tables (fetched 6 Nov 2026); Tom's Hardware GPU Benchmarks Hierarchy (updated 24 Jun 2026; Ryzen 7 9800X3D, native resolution, no DLSS or frame generation); Blender Open Data (fetched 6 Nov 2026, n=220 and n=387); llama.cpp gpt-oss-20b guide (discussion #15396, Aug 2025 builds); hardware-corner GPU ranking (Q4_K_XL, 16K context, Dec 2025). Percentages are our arithmetic from those figures.
Does the 5090's extra VRAM matter for gaming?
Not much yet. The 5090 is faster in games because it has twice the cores and nearly twice the bandwidth, not because of the extra 16 GB. Hardware Unboxed found Cyberpunk 2077 Phantom Liberty wanting about 16.5 GB at 4K with ray tracing and frame generation, which is right at the 5080's limit, and it recommends 16 GB or more for several titles at 4K with ray tracing. No test we found shows a current game needing more than 24 GB at stock settings. So 16 GB is borderline at 4K with path tracing, and 32 GB is headroom for the future rather than a measured need today.
The speed gap also depends on resolution. At 1080p the 5090 is only about 22% ahead in Tom's data, because even a Ryzen 7 9800X3D can't feed it fast enough. At 4K it pulls ahead by 43% in raster and 52% in ray tracing. If you play at 1440p, you're paying for a lot of performance your CPU and monitor may not use. Our RTX 5090 bottlenecks post explains why.
Both cards get DLSS 4 Multi Frame Generation and the DLSS 4.5 5x and 6x modes, because those are RTX 50 features. A 5080 with frame generation can show very high frame rates, but responsiveness still follows the frame rate it actually renders.
Where does 32 GB actually change the result?
In AI and large creative scenes, the difference is about what can run at all, not just how fast it runs.
- Local language models: Qwen3 32B at 4-bit is a 19.8 GB file and Gemma 3 27B at 4-bit is 16.6 GB (Hugging Face file sizes). Neither loads on a 16 GB card. The 5080 tops out around Mistral Small 24B at 4-bit or gpt-oss-20b. See why 32 GB still frustrates AI developers for where even the 5090 runs out.
- Fine-tuning: Unsloth's published minimums put a 32B QLoRA job at 26 GB and an 8B 16-bit LoRA at 22 GB. Both fit a 5090, and neither fits a 5080. A 16 GB card handles QLoRA up to about 14B.
- Rendering: Blender renders a scene on one GPU's memory. When the scene doesn't fit, Blender falls back to system memory, with a performance cost. A 5080 caps you at 16 GB of scene per GPU.
- Speed when both fit: on models that fit in 16 GB, the 5090 was about 38% faster on gpt-oss-20b and about 54% faster on Qwen3 8B in the llama.cpp results above. That's less than the 1.87x bandwidth gap, because small models don't saturate the 5090's memory system.
If your largest job fits in 16 GB, the 5080 gives you most of the experience for a fraction of the cost. If it doesn't, the 5090 isn't the faster option; it's the only GeForce option.
What does the extra power cost to run?
The 5090's rated board power is 575 W against the 5080's 360 W. PC Games Hardware measured peaks of 598 W and 393 W in games, and NVIDIA asks for a 1,000 W power supply for the 5090 against 850 W for the 5080. The 5090 also idles higher: PCGH measured 27–31 W at the desktop, against 12–15 W for the 5080.
At SP Group's Q4 2026 tariff of 31.16 cents per kWh including GST, the card's rated power works out like this. Four hours a day for a year at 575 W is about 840 kWh, or roughly S$262. The same hours at 360 W is about 526 kWh, or roughly S$164. That's an illustrative upper bound: real gaming draw is lower than TGP, and tariffs reset each quarter. For a gamer, the running-cost gap is about S$100 a year. For a 24/7 AI box it's several times that. Our 575 W power post covers heat and PSU sizing in detail.
Is the price premium justified?
At MSRP, the answer was close to neutral for 4K gaming. US$1,999 against US$999 bought about 43% more native 4K performance, so the 5090 cost about US$18 per 4K frame against about US$13 for the 5080 (our arithmetic from Tom's data).
Street prices have moved the line. Tom's Hardware's US price tracker listed the 5090 at US$8,477 and the 5080 at US$1,699 on 26 Oct 2026. That's about US$77 per 4K frame against about US$22. Asking-price data from eBay in November 2026 put them at about US$6,542 and US$1,865, a ratio of about 3.5x. These prices change weekly and differ by country, and NVIDIA has not changed either MSRP. NVIDIA has said memory supply is constrained, and Tom's Hardware has pointed to AI demand as part of why the 5090 is scarce.
So for gaming, the 5080 is the more sensible card unless you're chasing 4K at high refresh and price doesn't matter. For AI or creative work that needs more than 16 GB, compare the 5090 with other 32 GB-plus options rather than with the 5080. Our 5070 Ti vs 5080 vs 5090 value guide breaks this down by resolution, and RTX 5090 vs RTX 4090 covers the upgrade case.
Which one should you buy?
- Choose the RTX 5080 for 1440p gaming, for 4K with DLSS, and for AI models and scenes that fit in 16 GB.
- Choose the RTX 5090 for native 4K at high refresh rates, for 24B–32B language models, fine-tuning above 14B, and large GPU-rendered scenes.
- Consider a professional card if you need 32 GB with ECC and certified drivers more than raw speed. The RTX PRO 4500 Blackwell has 32 GB but is much slower in rendering: its Blender 5.2.0 median is 9,186, close to the 5080's.
Buying in Singapore
RTX 5090 and RTX 5080 desktop cards are not in stock at SourceIT at the time of writing; ask us about availability and a quotation. For a complete compact system, the ASUS ROG NUC 16 with an RTX 5080 Laptop GPU (16 GB) is S$5,650.03 (listed 6 Nov 2026). Laptop GPUs are slower than the desktop cards with the same name. If you need 32 GB with professional drivers, the RTX PRO 4500 Blackwell is S$8,880. Prices change, so ask for a current quotation. Browse the ASUS ROG NUC range, then use Request for Quotation on the product page, call (65) 6978 3502, email Sales@sourceit.com.sg or use our contact page.
Frequently asked questions
How much faster is the RTX 5090 than the RTX 5080? In Tom's Hardware's June 2026 tests at native resolution, about 30% at 1440p, 43% at 4K and 52% in 4K ray tracing. In Blender's Open Data benchmark the 5090's median is about 1.76x the 5080's.
Is 16 GB of VRAM enough for gaming in 2026? For most games at 1440p and 4K, yes. A few titles come close to 16 GB at 4K with ray tracing and frame generation, so it's borderline at the very top settings.
Can an RTX 5080 run local AI models? Yes, up to about gpt-oss-20b or Mistral Small 24B at 4-bit. Models like Qwen3 32B and Gemma 3 27B at 4-bit don't fit in 16 GB.
How much more power does the RTX 5090 use? Its rated board power is 575 W against 360 W for the 5080, and NVIDIA recommends a 1,000 W power supply against 850 W.
Is a 24 GB RTX 5080 coming? Reports of an RTX 5080 24 GB or "Super" card were still rumours in October 2026. Today the 5090 is the only RTX 50 desktop card with more than 16 GB.
