Skip to content
runcrate
Console

Benchmarks

Whisper benchmarks. Audio hours per dollar.

Whisper large-v3 and large-v3-turbo on four rented GPUs, five setups: speed against real time, word error rate and what 1,000 hours of audio cost.

Launch an RTX 5090All benchmarks

Measured

315xreal time

RTX 5090, faster-whisper large-v3-turbo, batch 16.

Run 2026-10-28, single runSee the table

Run 2026-10-28

Whisper large-v3 audio hours per dollar. Five setups on four GPUs.

Batched faster-whisper with the turbo model transcribed 315 hours of audio per hour on the RTX 5090, about 267 hours of audio per dollar at the run-day price, at a 3.1% word error rate. Plain sequential large-v3 with beam search was 6.6x slower than the same weights batched on that GPU, with a higher error rate (5.7% against 3.2%).

Single run: one run per setup, per machine, on one day. First published 2026-10-28.

The full write-up, with every command

Hardware

GPUs
RTX 5090 32 GB, L40S 48 GB, RTX A6000 48 GB and H100 PCIe 80 GB, one per machine, 6 to 20 vCPUs

Software

faster-whisper
faster-whisper BatchedInferencePipeline, batch 16: large-v3 fp16, large-v3 int8 (int8_float16), large-v3-turbo fp16
transformers
transformers 4.57.1 pipeline, large-v3 fp16, 30 s chunks, batch 16, SDPA attention
Sequential
faster-whisper large-v3 fp16, plain model.transcribe, beam 5

Method

Audio
distil-whisper/tedlium-long-form: 11 full TED talks, 8,905 s (2 h 28 min) of 16 kHz English with reference transcripts
Measured
Wall time for all 11 talks after a one-minute warm-up (model load excluded), speed as a multiple of real time, word error rate with Whisper's English normalizer, peak GPU memory from nvidia-smi

Price per hour

  • RTX 5090$1.18 on 2026-10-28, from $0.84/hr per GPU today
  • L40S$1.14 on 2026-10-28, from $0.54/hr per GPU on a 2-GPU machine today
  • RTX A6000$0.66 on 2026-10-28, from $0.57/hr per GPU today
  • H100$3.86 on 2026-10-28, from $3.87/hr per GPU on a 4-GPU machine today
Speed, turbo batched (x realtime)x realtime
Measured on 2026-10-28, from the table below.
faster-whisper large-v3-turbo, fp16, batch 16
GPU$/hour, run dayx realtimeWERPeak GPU memoryAudio hours per $, run day$ per 1,000 h, run day$ per 1,000 h, today
RTX 50901.18315x3.1%4.6 GiB267$3.75$2.67
L40S1.14192x3.1%4.5 GiB168$5.94$2.81
RTX A60000.66121x3.1%4.4 GiB183$5.45$4.71
H1003.86196x3.1%4.6 GiB51$19.65$19.70

Run day: the price on 2026-10-28, as the run paid it.

Today: the run-day figure times today's price per GPU-hour over the run-day price per GPU-hour. Today's price is the "from" price per GPU-hour shown on /pricing, which can be on a machine with more GPUs than the run used (the list of GPUs below names it). Prices read 2026-11-04 04:00 UTC.

faster-whisper large-v3, fp16, batch 16
GPU$/hour, run dayx realtimeWERPeak GPU memoryAudio hours per $, run day$ per 1,000 h, run day$ per 1,000 h, today
RTX 50901.18171x3.2%10.2 GiB145$6.89$4.90
L40S1.14109x3.2%10.3 GiB96$10.43$4.94
RTX A60000.6672x3.2%9.8 GiB109$9.15$7.90
H1003.86118x3.2%10.0 GiB30$32.80$32.88

Run day: the price on 2026-10-28, as the run paid it.

Today: the run-day figure times today's price per GPU-hour over the run-day price per GPU-hour. Today's price is the "from" price per GPU-hour shown on /pricing, which can be on a machine with more GPUs than the run used (the list of GPUs below names it). Prices read 2026-11-04 04:00 UTC.

faster-whisper large-v3, int8, batch 16
GPU$/hour, run dayx realtimeWERPeak GPU memoryAudio hours per $, run day$ per 1,000 h, run day$ per 1,000 h, today
RTX 50901.18172x3.2%8.4 GiB146$6.86$4.88
L40S1.14109x3.2%8.3 GiB95$10.49$4.97
RTX A60000.6673x3.2%8.1 GiB110$9.08$7.84
H1003.86119x3.1%8.3 GiB31$32.36$32.44

Run day: the price on 2026-10-28, as the run paid it.

Today: the run-day figure times today's price per GPU-hour over the run-day price per GPU-hour. Today's price is the "from" price per GPU-hour shown on /pricing, which can be on a machine with more GPUs than the run used (the list of GPUs below names it). Prices read 2026-11-04 04:00 UTC.

transformers large-v3, fp16, batch 16
GPU$/hour, run dayx realtimeWERPeak GPU memoryAudio hours per $, run day$ per 1,000 h, run day$ per 1,000 h, today
RTX 50901.1847x2.9%27.9 GiB40$25.05$17.83
L40S1.1423x2.9%27.6 GiB20$48.93$23.18
RTX A60000.6619x2.9%27.5 GiB29$34.92$30.16
H1003.8636x2.9%27.9 GiB9$106.93$107.21

Run day: the price on 2026-10-28, as the run paid it.

Today: the run-day figure times today's price per GPU-hour over the run-day price per GPU-hour. Today's price is the "from" price per GPU-hour shown on /pricing, which can be on a machine with more GPUs than the run used (the list of GPUs below names it). Prices read 2026-11-04 04:00 UTC.

faster-whisper large-v3, fp16, sequential, beam 5
GPU$/hour, run dayx realtimeWERPeak GPU memoryAudio hours per $, run day$ per 1,000 h, run day$ per 1,000 h, today
RTX 50901.1826x5.7%5.6 GiB22$45.74$32.56
L40S1.1419x5.5%5.5 GiB17$60.00$28.42
RTX A60000.6610x6.0%5.4 GiB15$65.35$56.44
H1003.8619x5.5%5.6 GiB5$207.53$208.07

Run day: the price on 2026-10-28, as the run paid it.

Today: the run-day figure times today's price per GPU-hour over the run-day price per GPU-hour. Today's price is the "from" price per GPU-hour shown on /pricing, which can be on a machine with more GPUs than the run used (the list of GPUs below names it). Prices read 2026-11-04 04:00 UTC.

What stood out

  • int8 gave no speedup on any of the four GPUs. It saved about 1.8 GiB of memory and changed the error rate by 0.1 points at most.
  • Turbo was 1.7 to 1.8x faster than large-v3 at the same error rate (3.1% against 3.2%) on this English set.
  • The transformers pipeline had a 2.9% error rate, under every faster-whisper setup, at the highest cost per audio hour of the batched setups.

Keep in mind

  • One English test set of TED talks: clean, single-speaker audio. Noisy calls, accents or other languages give different error rates.
  • Peak memory comes from nvidia-smi, which includes what the frameworks reserve.
  • The machines differ in CPU, disk and network (6 to 20 vCPUs), and the batched pipeline uses the CPU, so part of the GPU-to-GPU difference may come from the rest of the machine.

Costs

How the costs are worked out. Run day and today, side by side.

Run day

The cost the write-up printed: the hourly price the machine was launched at that day, against what it measured. Prices move daily, so each run-day figure carries its date. Every run was a single run.

Today

The run-day figure moved by today's price: the "from" price per GPU-hour shown on /pricing, over the run-day rate per GPU-hour. A GPU that is not in stock right now reads "Not in stock" rather than a stale price. Prices read 2026-11-04 04:00 UTC.

GPUs

The GPUs on this page. From price per GPU-hour, today.

  • RTX 5090from $0.84/hr per GPU
  • L40Sfrom $0.54/hr per GPU on a 2-GPU machine
  • RTX A6000from $0.57/hr per GPU
  • H100from $3.87/hr per GPU on a 4-GPU machine

Benchmarks

More benchmarks.

LLM inferenceCost per GPU.$0.12 per 1M output tokens. RTX A6000, vLLM, 64 concurrent requests, at the run-day price, run 2026-10-26, single run.A100 vs H100 vs H200One script, six GPUs.718.46 bf16 TFLOPS. H200, dense bf16 matmul at 8192, run 2026-10-25, single run.Text embeddingsCost per passage.$0.032 per 1M passages. L40S, bge-small, 128-token passages, at the run-day price, run 2026-10-29, single run.Training and fine-tuningMemory, speed, cost.14.5 GiB peak memory. LoRA of Qwen2.5-1.5B on an L40S, micro-batch 4, run 2026-10-28, single run.Image generationImages per dollar.$0.91 per 1,000 images. RTX A6000, compiled, batch 1, at the run-day price, run 2026-10-27, single run.GPU startup timeTimed from the API call.186 s to the first CUDA op. L40S, from the create call to a 4096 x 4096 matmul, run 2026-10-26, single run.