AI music and sound-effect generation · 6 models · 4 GPUs measured first-party · Updated October 2026
Best GPU for AI music and sound-effect generation
Which graphics card to use for AI music and sound-effect generation, from first-party measurements of ACE-Step 1.5, EzAudio XL, MOSS-SoundEffect v2.0 and more on 4 GPUs.
NVIDIA H100 80GB HBM3
27.05 x realtime on ACE-Step 1.5, the ceiling. Measured on our bench. 80GB of VRAM, $30,000 at launch.
- 27.05 x realtime on ACE-Step 1.5
- 80GB, clears the ACE-Step 1.5 floor
- Rentable by the hour rather than bought
- 700W board rating
- Datacenter or workstation hardware, not a retail purchase
Audio generation models write music from a text description or produce sound effects (rain, footsteps, explosions) for video and games. Open models like ACE-Step can now produce full songs in seconds on a datacenter card, which makes the GPU question about throughput, not feasibility.
We measured 6 models for AI music and sound-effect generation on 4 GPUs. Speed is times faster than realtime (10x = a 30 second track in 3 seconds). Every number below is a first-party run on our own harness; cards absent from a model's chart have not been run on it yet.
ACE-Step 1.5: x realtime by GPU
EzAudio XL: x realtime by GPU
MOSS-SoundEffect v2.0: x realtime by GPU
ACE-Step v1 3.5B: x realtime by GPU
MiDashengLM-Gen: x realtime by GPU
DiffRhythm 2: x realtime by GPU
Which models fit which card, for AI music and sound-effect generation
| Model | VRAM used | 8GB card | 12GB card | 16GB card | 24GB card | 32GB card | Licence |
|---|---|---|---|---|---|---|---|
| ACE-Step 1.5 | 7.7GB | Yes | Yes | Yes | Yes | Yes | MIT |
| EzAudio XL | 9.9GB | No | Yes | Yes | Yes | Yes | MIT |
| MOSS-SoundEffect v2.0 | 13.8GB | No | No | Yes | Yes | Yes | Apache-2.0 |
| ACE-Step v1 3.5B | 9GB | No | Yes | Yes | Yes | Yes | Apache-2.0 |
| MiDashengLM-Gen | 12.1GB | No | No | Yes | Yes | Yes | Apache-2.0 |
| DiffRhythm 2 | 8.8GB | No | Yes | Yes | Yes | Yes | Apache-2.0 |
From the lowest VRAM peak we measured for each model, plus 5% headroom. 'No' means it did not fit in that much memory at our settings, not that no setting ever could.
Every GPU x every AI music and sound-effect generation model (x realtime)
| GPU | ACE-Step 1.5 | EzAudio | MOSS-SFX v2 | ACE-Step v1 | MiDashengLM-Gen | DiffRhythm 2 |
|---|---|---|---|---|---|---|
| NVIDIA H100 80GB HBM3 | 27.05 | 1.58 | 2.8 | 14.83 | 1.43 | 5.63 |
| NVIDIA L40S | 20.44 | 1.34 | 1.23 | 13.78 | 1.07 | 5.43 |
| NVIDIA A10G | 18.65 | 0.71 | 0.53 | — | — | — |
| NVIDIA L4 | 12.84 | 0.38 | 0.37 | 8.26 | 0.71 | — |
— = not measured on that card yet.
What the numbers show.
ACE-Step 1.5: fastest on the NVIDIA H100 80GB HBM3 at 27.05 x realtime, 2.11x the slowest card we measured (NVIDIA L4); it used about 7.7GB of VRAM.
EzAudio XL: fastest on the NVIDIA H100 80GB HBM3 at 1.58 x realtime, 4.16x the slowest card we measured (NVIDIA L4); it used about 9.9GB of VRAM.
MOSS-SoundEffect v2.0: fastest on the NVIDIA H100 80GB HBM3 at 2.8 x realtime, 7.57x the slowest card we measured (NVIDIA L4); it used about 13.8GB of VRAM.
ACE-Step v1 3.5B: fastest on the NVIDIA H100 80GB HBM3 at 14.83 x realtime, 1.8x the slowest card we measured (NVIDIA L4); it used about 9GB of VRAM.
Which model to pick. On the same card, the NVIDIA H100 80GB HBM3, ACE-Step 1.5 runs at 27.05 x realtime in about 7.7GB; ACE-Step v1 3.5B runs at 14.83 x realtime in about 9GB; DiffRhythm 2 runs at 5.63 x realtime in about 8.8GB; MOSS-SoundEffect v2.0 runs at 2.8 x realtime in about 13.8GB; EzAudio XL runs at 1.58 x realtime in about 9.9GB; MiDashengLM-Gen runs at 1.43 x realtime in about 12.1GB. ACE-Step 1.5 gets through the work 18.9x as fast as MiDashengLM-Gen, so the model you choose moves the speed as much as the card does. Check the licence column in the fit table before you build on one: Apache-2.0, MIT are not the same deal.
How it compares. H100 80GB HBM3: ACE-Step 1.5 27.05 x realtime, ACE-Step v1 3.5B 14.83, DiffRhythm 2 5.63, MOSS-SoundEffect v2.0 2.8, MusicGen Small 2.47. ACE-Step 1.5 beats all 4 here.
Cost on a rented GPU. 1 hour of audio of ACE-Step 1.5: $0.034 on a L4 ($0.44/hr, 5 min), $0.079 on a H100 80GB HBM3 ($2.14/hr, 2 min, 2.3x the cost).
ACE-Step 1.5: cost per 1 hour of audio on rented GPUs
| GPU | Cheapest rate | Speed (x realtime) | Cost per 1 hour of audio |
|---|---|---|---|
| NVIDIA L4 | $0.44/hr | 12.84 | $0.034 |
| NVIDIA L40S | $0.79/hr | 20.44 | $0.039 |
| NVIDIA H100 80GB HBM3 | $2.14/hr | 27.05 | $0.079 |
Cheapest hourly rate we track on RunPod and Vast.ai, divided by the measured speed. Startup time and storage are extra.
Speed tiers for ACE-Step 1.5. 10-100 x realtime: 4 (H100 80GB HBM3, L40S, A10G). 1x is the speed of playback.
VRAM for ACE-Step 1.5. Measured peak 7.4GB, so 8GB is the smallest common card size; smallest card it ran on: A10G (24GB).
Power on ACE-Step 1.5. Most efficient: L4, 56W, 4.3 Wh per 1 hour of audio. Hungriest: H100 80GB HBM3, 179W, 6.6 Wh.
Our verdict
For AI music and sound-effect generation, the NVIDIA H100 80GB HBM3 is the fastest card we measured. We have not measured a consumer card on this job yet; the picks above are datacenter and workstation hardware. Check the fit table before buying: VRAM, not speed, is what rules a card out.
FAQ
- What is the fastest GPU for AI music and sound-effect generation?
- In our runs, the NVIDIA H100 80GB HBM3 at 27.05 x realtime on ACE-Step 1.5. We measured 6 models on 4 GPUs for this page.
- How much VRAM do I need for AI music and sound-effect generation?
- The lightest model here, ACE-Step 1.5, used about 7.7GB. The table above shows which models fit 8, 12, 16, 24 and 32GB cards, from measured peaks.
- Are these numbers measured or estimated?
- Measured. Every number on this page is a first-party run on our own harness, with power and VRAM sampled during the run. Cards we have not run yet are simply absent, not filled in.
- What GPU do I need to run ACE-Step 1.5?
- About 8GB. Smallest card that ran it: NVIDIA A10G (24GB).
- How much does it cost to run ACE-Step 1.5 in the cloud?
- $0.034 per 1 hour of audio on a NVIDIA L4 at $0.44/hr, cheapest of 3 rentable cards we measured.
- Can I run ACE-Step 1.5 on a 12GB, 16GB or 24GB card?
- It used 7.4GB at the precision we tested. 12GB: yes; 16GB: yes; 24GB: yes.
- Is the H100 80GB HBM3 or the L40S faster for ACE-Step 1.5?
- The H100 80GB HBM3: 27.05 vs 20.44 x realtime, 32% faster on our bench.
How we test
Each model runs a fixed workload on every card: a warmup, then timed runs with power, temperature and VRAM sampled every half second through NVML. Models run at the precision and settings from their model card. Datacenter cards run on Modal; consumer cards on rented machines. Non-commercially licensed models are not part of this page.