Sana 1.6B · 24 GPUs measured first-party · text-to-image · Updated October 2026

What GPU Do You Need for Sana 1.6B?

We ran Sana 1.6B at 1024px on 24 GPUs, from budget cards to datacenter parts, and measured image speed, power draw and peak VRAM on every one. The fastest was NVIDIA GeForce RTX 5090 at 51.52 images/min.

Benchmarked weights: Efficient-Large-Model/Sana_1600M_1024px_diffusers

Fastest we measured

NVIDIA GeForce RTX 5090

51.52 images/min on Sana 1.6B, the ceiling. Measured on our bench. 32GB of VRAM, $1,999 at launch.

Pros

• 51.52 images/min on Sana 1.6B
• 32GB, clears the Sana 1.6B floor

Cons

• 575W board rating

Best consumer card

NVIDIA GeForce RTX 4090

41.47 images/min on Sana 1.6B, fastest card you can buy at retail. Measured on our bench. 24GB of VRAM, $1,599 at launch.

Pros

• 41.47 images/min on Sana 1.6B
• 24GB, clears the Sana 1.6B floor

Cons

• 450W board rating

Cheapest card that runs it

NVIDIA GeForce RTX 3060

7.4 images/min on Sana 1.6B, lowest launch price that still fits. Measured on our bench. 12GB of VRAM, $329 at launch.

Pros

• 7.4 images/min on Sana 1.6B
• 12GB, clears the Sana 1.6B floor

Cons

• 170W board rating

Best value

GeForce RTX 5070

19.71 images/min on Sana 1.6B, most speed per dollar. Measured on our bench. 12GB of VRAM, $549 at launch. That is 35.9 images/min per $1,000 of launch price.

Pros

• 19.71 images/min on Sana 1.6B
• 12GB, clears the Sana 1.6B floor

Cons

• 250W board rating

What the numbers show. Across 24 GPUs measured on our own bench, RTX 5090 is fastest at 51.5 images/min. L4 is the most efficient, 13.8 images/min at 72W (0.192 images/min per watt). Per dollar of launch price, RTX 5060 Ti gives the most (38.4 images/min per $1,000). The fastest card with 16GB or less is RTX 5080 at 30.0 images/min. The measured peak was ~12GB, so it runs on 12GB cards and up. That floor is a property of the model, so it applies to every GPU, measured or not.

How it compares. L40S: Sana 1.6B 50.81 images/min, FLUX.2 klein 4B 50.29 (4B), Stable Diffusion 1.5 58.54, Swin2SR 4x Upscaler 34.14, PixArt-Sigma XL 26.95. 1 of 4 beat Sana 1.6B here.

Cost on a rented GPU. 1,000 images of Sana 1.6B: $0.081 on a RTX 3060 ($0.036/hr, 2.3 hours), $0.13 on a RTX 5090 ($0.39/hr, 19 min, 1.6x the cost).

Speed tiers for Sana 1.6B. 30+ images/min: 4 (RTX 5090, RTX 4090, RTX 5080); 6-30 images/min: 19 (RTX 4080, RTX 4080 Super, RTX 5070 Ti); under 6 images/min: 1 (T4). 30 images/min means two seconds or less per picture.

Time per image. Sana 1.6B at 1024px and 20 steps: 1.2s per image on the RTX 5090, 53.1s on the T4. A batch of 100 takes 2 min on the fastest card and 88 min on the slowest.

VRAM for Sana 1.6B. Measured peak 10.6GB, so 12GB is the smallest common card size; smallest card it ran on: RTX 4070 Ti (12GB).

Power on Sana 1.6B. Most efficient: L4, 72W, 86.8 Wh per 1,000 images. Hungriest: RTX 5090, 549W, 0.18 kWh. At $0.15/kWh: $0.013 per 1,000 images.

Our verdict

Fastest on Sana 1.6B: NVIDIA GeForce RTX 5090, 51.52 images/min. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 ($329, 7.4 images/min). Cheapest to rent per job: NVIDIA GeForce RTX 3060, $0.081 per 1,000 images.

FAQ

What GPU do I need to run Sana 1.6B?

About 11GB. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 (12GB, 7.4 images/min).

How fast is Sana 1.6B on the NVIDIA GeForce RTX 5090?

51.52 images/min at 549W, faster than every datacenter card we ran it on.

How much does it cost to run Sana 1.6B in the cloud?

$0.081 per 1,000 images on a NVIDIA GeForce RTX 3060 at $0.036/hr, cheapest of 21 rentable cards we measured.

Can I run Sana 1.6B on a 12GB, 16GB or 24GB card?

It used 10.6GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.

Is the RTX 4090 or the RTX 3090 faster for Sana 1.6B?

The RTX 4090: 41.47 vs 20.6 images/min, 101% faster on our bench.

Should I buy or rent a RTX 5090 for Sana 1.6B?

Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 15,885x 1,000 images of Sana 1.6B. Buy only if you'll run more than that.

How we test

Sana 1.6B at 1024px in diffusers, bf16, native precision with no offload, timed over three generations after a warmup, with power and VRAM sampled throughout. Image speed tracks tensor compute and architecture generation more than memory bandwidth, so the order here differs from our LLM boards.