GPU × use-case guide · Speech-to-text

Is the H100 PCIe a good GPU for speech-to-text?

The H100 PCIe is a hopper NVIDIA GPU with 80 GB HBM3 (2000 GB/s), 756 TFLOPS BF16 / 1513 TFLOPS FP8, and a 350 W TDP. Speech-to-text workloads care most about encoder throughput and real-time factor; models are small so mid-range cards are plenty. Here's how the H100 PCIe measures up.

What models fit on a single H100 PCIe?

Weights only, reserving ~25% of the 80 GB for KV cache, activations and fragmentation. ✓ = fits on one card.

Largest single-card fit: Gemma 2 27B at BF16, Mixtral 8x7B (MoE) at FP8, Qwen 2.5 72B at INT4. Bigger models need tensor-parallel across 8 cards.

H100 PCIe for speech-to-text, specifically

Speech-to-text is throughput-bound rather than VRAM-bound, so the H100 PCIe's 2000 GB/s of bandwidth and 756 TFLOPS matter more than raw capacity. Because it's latency-sensitive, run it at low batch sizes on a fast framework (vLLM ≈ 85% MFU) rather than maximising batch. Size it precisely on the calculator.

H100 PCIe pricing across providers

Verdict

With 80 GB, the H100 PCIe is a data-center-class card that comfortably handles speech-to-text for models up to Gemma 2 27B at full precision on a single card — a strong pick if your budget supports ~$2.29/hr.