CogVideoX-2B · 13 GPUs measured first-party · text-to-video · Updated October 2026
What GPU Do You Need for CogVideoX-2B?
CogVideoX-2B on 13 GPUs, measured first-party: NVIDIA H100 80GB HBM3 leads at 1.17 frames/s, L4 trails at 0.179 frames/s, and it peaked at 20GB of VRAM.
Benchmarked weights: THUDM/CogVideoX-2b
NVIDIA H100 80GB HBM3
1.17 frames/s on CogVideoX-2B, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 1.17 frames/s on CogVideoX-2B
- 80GB, clears the CogVideoX-2B floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
NVIDIA GeForce RTX 5090
0.7 frames/s on CogVideoX-2B, fastest card you can buy at retail. Measured on our bench. 32GB of VRAM, $1,999 at launch.
- 0.7 frames/s on CogVideoX-2B
- 32GB, clears the CogVideoX-2B floor
- 575W board rating
NVIDIA GeForce RTX 3060
0.09 frames/s on CogVideoX-2B, lowest launch price that still fits. Measured on our bench. 12GB of VRAM, $329 at launch.
- 0.09 frames/s on CogVideoX-2B
- 12GB, clears the CogVideoX-2B floor
- 170W board rating
GeForce RTX 5070 Ti
0.29 frames/s on CogVideoX-2B, most speed per dollar. Measured on our bench. 16GB of VRAM, $749 at launch. That is 0.39 frames/s per $1,000 of launch price.
- 0.29 frames/s on CogVideoX-2B
- 16GB, clears the CogVideoX-2B floor
- 300W board rating
What GPU Do You Need for CogVideoX-2B?, frames/s by GPU
Efficiency: frames/s per 100W drawn
Power is the average pulled during the run, sampled at 1Hz. The fastest card is often not the one here, and for anything left running this is the number that shows up on the bill.
Value: frames/s per $1,000 of MSRP
Launch price, not street price, so it ages. A speed leaderboard always crowns the most expensive card; this is the counterweight.
CogVideoX-2B. Measured text-to-video speed by GPU
| GPU | frames/s | Peak VRAM | Avg power |
|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 1.17 | 34.2GB | 666.3 W |
| NVIDIA GeForce RTX 5090 | 0.7 | — | — |
| NVIDIA L40S | 0.6 | 34.1GB | 337.9 W |
| NVIDIA GeForce RTX 4090 | 0.52 | 19.5GB | 409.6 W |
| GeForce RTX 5080 | 0.36 | 9.4GB | 281.3 W |
| NVIDIA GeForce RTX 4080 | 0.31 | 9.3GB | 294.0 W |
| GeForce RTX 5070 Ti | 0.29 | 9.3GB | 220.1 W |
| NVIDIA GeForce RTX 3090 | 0.24 | 22.2GB | 348.0 W |
| NVIDIA GeForce RTX 4070 Super | 0.23 | 9.2GB | 193.3 W |
| NVIDIA L4 | 0.18 | 18.6GB | 71.6 W |
| GeForce RTX 5060 Ti | 0.16 | 9.2GB | 135.2 W |
| NVIDIA GeForce RTX 4060 Ti 16GB | 0.14 | 9.1GB | 144.8 W |
| NVIDIA GeForce RTX 3060 | 0.09 | 9.1GB | 155.2 W |
What the numbers show. Across 3 GPUs measured on our own bench, H100 80GB HBM3 is fastest at 1.17 frames/s. The slowest, L4, manages 0.18, so the spread is 6.5x from top to bottom. L4 is the most efficient, 0.18 frames/s at 72W. Per dollar of launch price, L40S gives the most (0.1 frames/s per $1,000).
About CogVideoX-2B. CogVideoX-2B: from zai-org, 1.7B parameters, on Hugging Face since August 2024, Apache 2.0 licence. 16,190 downloads in the last 30 days.
How it compares. RTX 4090: CogVideoX-2B 0.52 frames/s, LTX-Video (distilled) 6.7 (2B), Wan 2.1 1.3B 0.67 (1B), CogVideoX-5B 0.17 (6B), Wan 2.2 5B (720p) 0.43. 2 of 4 beat CogVideoX-2B here.
Cost on a rented GPU. 1 minute of 24fps video of CogVideoX-2B: $0.16 on a RTX 3060 ($0.036/hr, 4.5 hours), $0.73 on a H100 80GB HBM3 ($2.14/hr, 21 min, 4.5x the cost).
CogVideoX-2B: cost per 1 minute of 24fps video on rented GPUs
| GPU | Cheapest rate | Speed (frames/s) | Cost per 1 minute of 24fps video |
|---|---|---|---|
| NVIDIA GeForce RTX 3060 | $0.036/hr | 0.09 | $0.16 |
| NVIDIA GeForce RTX 3090 | $0.12/hr | 0.24 | $0.20 |
| GeForce RTX 5070 Ti | $0.15/hr | 0.29 | $0.20 |
| NVIDIA GeForce RTX 5090 | $0.39/hr | 0.7 | $0.22 |
| GeForce RTX 5080 | $0.21/hr | 0.36 | $0.23 |
| NVIDIA GeForce RTX 4090 | $0.34/hr | 0.52 | $0.26 |
| NVIDIA GeForce RTX 4080 | $0.20/hr | 0.31 | $0.26 |
| GeForce RTX 5060 Ti | $0.14/hr | 0.16 | $0.35 |
| NVIDIA L40S | $0.79/hr | 0.6 | $0.53 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 1.17 | $0.73 |
| NVIDIA L4 | $0.44/hr | 0.18 | $0.98 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for CogVideoX-2B. 1-10 frames/s: 1 (H100 80GB HBM3); under 1 frames/s: 12 (RTX 5090, RTX 4090, RTX 5080). 10 frames/s turns out a minute of 24fps video in under 2.5 minutes.
VRAM for CogVideoX-2B. Measured peak 9.1GB, so 12GB is the smallest common card size; smallest card it ran on: RTX 4070 Super (12GB).
Power on CogVideoX-2B. Most efficient: L4, 72W, 0.16 kWh per 1 minute of 24fps video. Hungriest: H100 80GB HBM3, 666W, 0.23 kWh. At $0.15/kWh: $0.024 per 1 minute of 24fps video.
Our verdict
Fastest on CogVideoX-2B: NVIDIA H100 80GB HBM3, 1.17 frames/s. Best desktop card: NVIDIA GeForce RTX 5090, 0.7 frames/s. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 ($329, 0.09 frames/s). Cheapest to rent per job: NVIDIA GeForce RTX 3060, $0.16 per 1 minute of 24fps video.
FAQ
- What GPU do I need to run CogVideoX-2B?
- About 10GB. Cheapest consumer card that ran it: NVIDIA GeForce RTX 3060 (12GB, 0.09 frames/s).
- How fast is CogVideoX-2B on the NVIDIA GeForce RTX 5090?
- 0.7 frames/s, 60% of the NVIDIA H100 80GB HBM3.
- How much does it cost to run CogVideoX-2B in the cloud?
- $0.16 per 1 minute of 24fps video on a NVIDIA GeForce RTX 3060 at $0.036/hr, cheapest of 11 rentable cards we measured.
- Can I run CogVideoX-2B on a 12GB, 16GB or 24GB card?
- It used 9.1GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the RTX 4090 or the RTX 3090 faster for CogVideoX-2B?
- The RTX 4090: 0.52 vs 0.24 frames/s, 115% faster on our bench.
- Should I buy or rent a RTX 5090 for CogVideoX-2B?
- Its $1,999 launch price buys 5,139 rented hours at $0.39/hr, enough for about 8,993x 1 minute of 24fps video of CogVideoX-2B. Buy only if you'll run more than that.
How we test
CogVideoX-2B in diffusers, bf16, one warmup clip and three timed clips from the same prompt, measured as frames of finished video per second, with CPU offload only on cards below the model's full-GPU size. Text-to-video is compute-bound and memory-hungry; cards that need CPU offload lose far more speed than their raw compute suggests.