Fish-Audio-S2-Pro (5B) VRAM & GPU Requirement Calculator
Interactive hardware sizing for Fish-Audio-S2-Pro (5B) (4.8B by Fish Audio). Estimated 8K context footprint: 5.0 GB (INT4) • 7.5 GB (FP8) • 12.6 GB (FP16).
Fixed 30s audio window / codec frame cap
GPUs
Memory Requirements 7.54 GB
Requires 1 GPUs (based on memory capacity)
5.65 GB
All model weights
0.28 GB
Conversation history cache
1.4 GB
Concurrent processing cache
0.21 GB
Temporary computation cache
Performance Metrics
Visualize 179 tok/s →Max Streams (This GPU)
19 streams
Duplex Architecture
Stream / Turn-Based
Estimated Latency
840 ms
Max QPS
1.2 queries/s
Estimated using memory bandwidth and tensor compute throughput across 1x NVIDIA RTX 4090 (24GB).
Quick Deploy CLI Command
python -m tools.api_server --llama-checkpoint-path checkpoints/s2-pro --compile
Flags reflect your selected precision, context length, and tensor parallel GPU count.Open Fish-Audio-S2-Pro spec page →
Fish-Audio-S2-Pro (5B) Architecture & Memory Profile
Total Parameters
4.8B
Architecture
Dense Transformer
Developer / Provider
Fish Audio
Recommended Precision
FP8
Fish-Audio-S2-Pro (5B) VRAM Matrix by Precision & Concurrent Streams
Precomputed audio-stream VRAM requirements including unquantized acoustic/codec submodules and runtime workspace.
| Precision | 1 Stream | 4 Streams | 16 Streams |
|---|---|---|---|
| FP16 | 12.6 GB | 16.2 GB | 30.4 GB |
| FP8 | 7.5 GB | 10.3 GB | 21.1 GB |
| INT8 | 7.5 GB | 10.3 GB | 21.1 GB |
| INT4 | 5.0 GB | 7.3 GB | 16.4 GB |
Recommended GPU Hardware for Fish-Audio-S2-Pro (5B) (FP8)
Minimum GPU count and maximum concurrent real-time voice streams per node.
| GPU Hardware | Min GPUs | Total VRAM Pool | Max Streams |
|---|---|---|---|
| NVIDIA RTX 4090 (24GB) | 1× | 24 GB | 19 streams |
| NVIDIA RTX 5090 (32GB) | 1× | 32 GB | 28 streams |
| Apple M4 Max (128GB) (128GB) | 1× | 128 GB | 134 streams |
| NVIDIA A100 (80GB) (80GB) | 1× | 80 GB | 81 streams |
| NVIDIA H100 (80GB) | 1× | 80 GB | 81 streams |
| NVIDIA H200 (141GB) | 1× | 141 GB | 148 streams |
| Huawei Ascend 910B (64GB) | 1× | 64 GB | 63 streams |
Fish-Audio-S2-Pro (5B) Deployment FAQ
How much VRAM is needed to run Fish-Audio-S2-Pro (5B)?
For single-stream inference, Fish-Audio-S2-Pro (5B) uses about 12.6 GB in FP16, 7.5 GB in FP8, and 5.0 GB in INT4 including resident acoustic codecs and stream buffers.
Can I run Fish-Audio-S2-Pro (5B) on RTX 4090 or H100 GPUs?
Running Fish-Audio-S2-Pro (5B) in FP8 precision requires at least 1x RTX 4090 (24GB) or 1x NVIDIA H100 (80GB) GPUs.
What precision should I choose for Fish-Audio-S2-Pro (5B) in production?
For Fish-Audio-S2-Pro (5B) (4.8B by Fish Audio), FP8 provides the best balance between inference speed, VRAM footprint, and output quality.