GPU Upgrade for Local Video: More VRAM or More Speed? (2026)

First separate "will not finish" from "too slow"

Insufficient VRAM shows up as a job that cannot complete, insufficient compute as a job that completes slowly, and the two symptoms point to two different specs, so confirm which one you have before upgrading. Write down your current workflow's model version, resolution, frame count and error message, then read the table:

The third row is the one people misread. When VRAM is just barely short, ComfyUI does not throw an error; it leaves the weights that do not fit in system RAM and streams them in chunks. The job completes, but every step waits on PCIe, which looks like a compute shortage and is really a VRAM shortage. A faster card does not fix it, because the bottleneck is transfer rather than math. See the explanation of diffusion-model offloading in does adding RAM help.

How much VRAM does a video model need? Wan 2.2 as the example

Wan 2.2's 5B version is 10 GB of fp16 weights, while the A14B version is two models, high-noise and low-noise, at 14.3 GB each in fp8 and 28.6 GB each in fp16, plus a 6.7 GB text encoder. These are the file sizes of ComfyUI's official repackaged builds (Comfy-Org repository), and they are the floor of the VRAM requirement:

Weights are only part of the peak. Peak VRAM during generation stacks up from three things:

1. Resident weights. The file sizes above. When ComfyUI runs A14B it loads the high-noise and low-noise models one after the other rather than together, so 24GB holds one 14.3 GB fp8 model plus runtime overhead, and the text encoder can release its VRAM once the prompt is encoded.
2. Intermediate tensors. Attention and latent volumes grow with frames times resolution. 720P for 5 seconds and 480P for 3 seconds use identical weights but differ several-fold in intermediate tensors, which is why lowering resolution rescues an OOM.
3. VAE decoding. The last step decodes dozens of latent frames into pixels at once, and its peak can exceed the generation stage. Weights are tiny here and tensors are huge, so offloading weights does not help. A run that blows up at the very end is usually this, and the fix is tiled decoding, fewer frames or lower resolution.

Read the official thresholds together with their assumptions. The ComfyUI docs say the 5B version should run on 8GB of VRAM with ComfyUI's native offloading (ComfyUI Wan 2.2 tutorial), while the Wan repository, using its own inference scripts, asks for at least 24GB for 5B, citing an RTX 4090, and at least 80GB for 14B on a single GPU (Wan2.2 repository). The three-fold gap is offloading: 8GB means it produces output, 24GB means nothing has to be shuttled. HunyuanVideo 1.5 states the same kind of figure for its 8.3B model, a 14 GB minimum with offloading enabled (HunyuanVideo 1.5 repository).

By VRAM tier, as of September 8, 2026:

16GB to 24GB is the best-value step in video generation, because it is the line where A14B fp8 goes from needing offload to not needing it. 24GB to 32GB buys shot length and headroom, and 32GB to 48GB buys fp16 precision; both of those steps cost far more.

VRAM sets how big a shot can be, compute sets how long you wait

Every step of a video diffusion model is a large matrix multiply, so wait time is set mostly by FP8 or BF16 tensor-core throughput, and memory bandwidth matters far less than it does for LLM decoding. This is the opposite of running language models: an LLM computes very little per token and is bottlenecked by the bandwidth of reading weights, while a video model runs attention over tens of thousands of spatiotemporal tokens per step and is bottlenecked by math.

So compare video cards on the compute column, and check that the precision lines up:

The RTX 3090 cell is the decisive one. Ampere has no FP8 tensor cores, so fp8_scaled weights load fine but have to be converted back to BF16 for the math, which means 142 TFLOPS in practice. The RTX 5070 Ti computes in FP8 directly at 352 TFLOPS, 2.5 times the 3090; the RTX 4090 is 4.6 times and the RTX 5090 is 5.9 times. With all weights resident and no offloading, wait time for the same shot shrinks roughly in that proportion. We have not run Wan 2.2 on these cards on our own bench, so the multiples above are compute ratios rather than measurements; the only official number Wan publishes is a 5-second 720P clip in under 9 minutes on a single consumer GPU.

Live specs and current prices:

The INT8 column shows INT8 TOPS for the 3090, and video models do not run INT8 inference, so use the BF16 figure from the previous table when reading the 3090. The RTX 4080 is the easy card to overlook: 16GB, 390 TFLOPS FP8, and an eBay median of $1,250 that sits $30 below the 5070 Ti. Its downsides are that it is a 2022 card, uses 12VHPWR, and carries the same used-market risk as a 3090.

Upgrading from 16GB: used RTX 3090 or RTX 5070 Ti?

Pick the RTX 3090 if you run out of VRAM, the RTX 5070 Ti if you wait too long, and the RTX 4090 if both. Live comparison of the three, plus the 5090:

Answers for three situations, at prices as of September 8, 2026:

• You have a 16GB card, A14B only runs with offloading, and output is slow because of the shuttling. Switch to an RTX 3090 at $1,545. The 24GB stops the fp8 model from offloading and the bulk of the wait disappears, even though the card computes slower than the 16GB Blackwell part you already own. There is no second option in this tier: the next 24GB card up is the $3,199 4090. Note that the 3090 costs $265 more than a 5070 Ti on eBay, so you are paying a small premium for the memory rather than trading sideways.
• You have a 16GB card, and 5B or a downscaled A14B finishes fine, just slowly. Switch to an RTX 5070 Ti at $1,280 or an RTX 5080 at $1,800, on the condition that VRAM stays sufficient. From an RTX 5060 Ti 16GB to a 5070 Ti, FP8 throughput goes from 190 to 352. Going from a 3090 to a 5070 Ti trades 8GB of memory for 2.5 times the compute and saves $265, which only pays off if you have confirmed the workflow fits in 16GB.
• It neither fits nor runs fast enough. The RTX 4090 at $3,199 brings 24GB and 661 TFLOPS FP8, roughly twice the price of a 3090 ($1,654 more) for 4.6 times the compute. The RTX 5090 at $6,599 adds another 8GB and 27% more compute for roughly 4.3 times the 3090's price, which is only worth it if you produce long 720P shots every day. The RTX 5090 D is China-market only at a Xianyu median of CNY 33,000 (about $4,600); it has the same 32GB but its AI throughput is capped in firmware to about 70% of the standard version, and at roughly 594 TFLOPS FP8 it is still close to a 4090.

The answer differs for language models: LLMs follow bandwidth, where the 3090 and 5070 Ti are nearly tied, so 24GB almost always wins. See 3090 vs 5070 Ti for local LLMs. Do not carry that conclusion over to video, because the 3090's missing FP8 barely affects single-stream LLM inference and costs 2.5 times in video generation.

Are the 48GB modded card and the pro card worth it?

The RTX 4090 48GB mod is the only cheap way past 24GB and suits people who specifically need fp16 precision or the whole A14B set resident; nobody else needs it. It is a China-market card with no US eBay listings, at a Xianyu median of CNY 26,600 as of September 8, 2026 from only three listings, CNY 3,700 above the stock 24GB 4090 at CNY 22,900.

What it can do: keep both A14B fp8 models and the text encoder resident (28.6 + 6.7 GB) so switching models needs no reload, or run a single fp16 model of 28.6 GB without offloading. What it cannot do: go faster. Compute is still the 4090's 661 TFLOPS, and doubling memory does not change that. The risk is the modification itself. Most are Founders cards rebuilt onto a dual-slot blower PCB with clamshell memory, with no NVIDIA warranty, loud blower noise, and memory chips and PCB quality that depend on the shop. Have the seller record a VRAM stress test that fills all 48GB before you pay, and run one again on arrival.

Above that sits the 96GB of the RTX PRO 6000 Blackwell at an eBay median of $16,985, enough to hold both A14B fp16 models and close to the 80GB figure in Wan's own scripts. At roughly 2.6 times a 5090 it is a studio purchase, outside the scope of a personal upgrade.

The rest of the build: PSU, connectors, case

Moving to a 3090 or above usually means replacing the PSU, the power connector and the case, and that money belongs in the upgrade budget. Video generation is a sustained full load for tens of minutes to a whole night, not the up-and-down load of a game, so size the PSU at rated power plus 30%:

The 3090's GDDR6X chips sit on the back of the board and run hot under long loads, so ask any used seller for a 30-minute VRAM stress test. The case has to give a triple-slot card room to draw air, or it will throttle during an overnight queue. Push the 16-pin connector on a 4090 or 5090 all the way in and do not bend the cable sharply at the plug. Rebuilding the rest of the machine usually costs a few hundred dollars more, so the real gap between a 5070 Ti and a 4090 is not $1,919 but that plus this line item.

Price your own waiting time

Only waiting that recurs every week justifies paying for compute; a job you run occasionally can live with offloading. The method is simple:

1. Log how many shots you actually generate in a week and how long each one waits, then total the hours.
2. Estimate the share you would save from the compute ratio. A 3090 to a 4090 cuts waiting to roughly a quarter or a fifth; a 5060 Ti 16GB to a 5070 Ti roughly halves it.
3. Multiply the hours saved by your hourly rate or the extra work you could take on, and compare against the price gap.

Example: 20 five-second 720P shots a day, assuming 8 minutes each on a 3090. That is 160 minutes a day, about 13 hours across a five-day week. The compute ratio puts a 4090 near 3 hours, saving 10 hours a week. The card gap is $1,654, and a PSU and case bring the real outlay to roughly $2,000. At $30 an hour that pays back in about seven weeks. Someone who renders 10 shots a week saves 1 hour a week instead, and the same $2,000 takes over a year. The 8 minutes per shot is an assumption; substitute your own log.

Also separate single-run latency from overnight throughput. While you are iterating on a shot you watch each result, so latency decides the experience and the faster card wins. When you queue candidates overnight and only care how many are ready in the morning, more VRAM and a larger batch without offloading matter more.

Price trends and when to buy

The 30-day trend for four cards updates live; the used 3090 sits a couple hundred dollars above the 5070 Ti, so your bottleneck rather than the price should decide. Trends:

The 3090 is a 2020 card whose price is set by remaining supply, and it falls slowly, which is why 24GB now carries a $265 premium over a newer 16GB card rather than trading at parity. The 5070 Ti and 5080 are current cards, so their used prices track retail. The 4090 has held around $3,200 since discontinuation, because nothing below a 5090 replaces the combination of 24GB and FP8. If you are watching for a deal, watch the listing count on the 3090: a larger sample makes the median more trustworthy, as explained under listing-price methodology in our methodology.

Before buying, write your reason as a single sentence: "I want A14B fp8 to run without offloading at 720P for 5 seconds," or "I want to cut an 8-minute wait per shot down to 3 minutes." If you can say where the resource goes, you know whether to buy VRAM or compute. If you cannot, use the card you have to finish one usable video first, then come back to this table.

Prices are eBay US used-listing medians as of September 8, 2026; the RTX 4090 48GB mod and RTX 5090 D have no US listings, so their figures are Xianyu medians. Model file sizes and VRAM thresholds come from the official ComfyUI, Wan and HunyuanVideo documentation. We have not run Wan 2.2 on our own test bench, so the speed multiples here are compute ratios rather than measurements.