Smallest AI reports an effective real-time factor of about 1,000 for Pulse Pro on one NVIDIA L40S GPU when audio is sent in batched 50-second chunks without word timestamps.

Quote found word for word on the live source, last checked October 19, 2026

In the source's own words

“Sustained throughput on L40S (batched, 50-second chunks) Mode RPS (requests / second) Effective RTFx No word timestamps 19–21 ~1,000×”

Parallelism and Latency (Smallest AI self-hosted Pulse STT), Smallest AI, n.d., Parallelism and Latency, Pulse Pro, Sustained throughput on L40S.

What it measures: The page states that the figures come from internal benchmarks on a single GPU host, with the L40S (48 GB) used for reference. The source writes the figure as 1,000 with a tilde, and it assumes optimal batching of typical-length audio. The page says a single very long file measured as low as 68 times real time.
What it does not show: A batched, steady-state figure on chunked audio, not the speed of one file, and so not comparable with single-file rows. It is a vendor's internal benchmark. The source writes the figure as approximately 1,000; it is recorded here as 1000.

A row of Speech Recognition Speed: Real-Time Factor Data, as Smallest AI, batched, no timestamps: 1,000 times faster than real time.

System: Smallest AI Pulse Pro (self-hosted)
Kind: Cloud service
Hardware: NVIDIA L40S GPU (48 GB)
Test set: Batched 50-second chunks, no word timestamps
Note: Source writes ~1,000× (approximate, batched 50-second chunks); recorded as 1000.

Fact · Speech recognition

1000times faster than real time

Quote found word for word on the live source, last checked October 19, 2026

“Sustained throughput on L40S (batched, 50-second chunks) Mode RPS (requests / second) Effective RTFx No word timestamps 19–21 ~1,000×”

Parallelism and Latency (Smallest AI self-hosted Pulse STT), Smallest AI, n.d., Parallelism and Latency, Pulse Pro, Sustained throughput on L40S.

What it measures: The page states that the figures come from internal benchmarks on a single GPU host, with the L40S (48 GB) used for reference. The source writes the figure as 1,000 with a tilde, and it assumes optimal batching of typical-length audio. The page says a single very long file measured as low as 68 times real time.
What it does not show: A batched, steady-state figure on chunked audio, not the speed of one file, and so not comparable with single-file rows. It is a vendor's internal benchmark. The source writes the figure as approximately 1,000; it is recorded here as 1000.

A row of Speech Recognition Speed: Real-Time Factor Data, as Smallest AI, batched, no timestamps: 1,000 times faster than real time.

System: Smallest AI Pulse Pro (self-hosted)
Kind: Cloud service
Hardware: NVIDIA L40S GPU (48 GB)
Test set: Batched 50-second chunks, no word timestamps
Note: Source writes ~1,000× (approximate, batched 50-second chunks); recorded as 1000.