Benchmarks
Real performance data, not marketing claims. Our benchmarks test every GPU we offer across LLM inference, image generation, OCR, and TTS workloads on dedicated GPU servers. See our tokens/sec benchmark for the latest results.
RTX 4090 24GB Prefill vs Decode Benchmark
Isolated prefill (~12000 t/s) and decode (~195 t/s) numbers on the RTX 4090 24GB across Llama 3 8B FP8, Mistral 7B FP8, Qwen 14B/32B AWQ and Llama 70B INT4 with…
RTX 4090 24GB Concurrent Chat Users: Capacity Benchmark
Deep concurrent-user benchmark on the RTX 4090 24GB across Llama 3 8B FP8, Mistral 7B, Qwen 14B/32B AWQ, Phi-3 mini…
RTX 4090 24GB FLUX.1-schnell Benchmark
Full FLUX.1-schnell benchmark on the RTX 4090 24GB - 1.8s per 1024px image at FP8, batch throughput, FP16 vs FP8…
RTX 4090 24GB Qwen 2.5 14B Benchmark: Full Quant Sweep with AWQ, FP8 and GGUF
Comprehensive RTX 4090 24GB benchmark for Qwen 2.5 14B - 135 t/s AWQ INT4, 110 t/s FP8 at batch 1,…
RTX 4090 24GB Qwen 2.5 32B Benchmark: AWQ INT4 at the VRAM Ceiling
Comprehensive RTX 4090 24GB benchmark for Qwen 2.5 32B - AWQ INT4 fits at 18GB, decodes 65 t/s, sustains 4…
RTX 4090 24GB Whisper Benchmarks: large-v3, Turbo, WhisperX
Deep real-time-factor measurements for Whisper large-v3, large-v3-turbo and medium on the RTX 4090 24GB, including batched WhisperX throughput, WhisperX alignment,…
RTX 4090 24GB Fine-Tuning Throughput: LoRA, QLoRA, Unsloth
LoRA, QLoRA and Unsloth fine-tuning throughput on the RTX 4090 24GB across Llama 3 8B, Mistral 7B, Qwen 14B, Qwen…
RTX 4090 24GB FLUX.1-dev Benchmark: 30-step in 4 Seconds FP8
FLUX.1-dev FP16 just fits a single RTX 4090 24GB at 22GB peak with 30-step renders in 6.2s; FP8 drops to…
RTX 4090 24GB SDXL Benchmark: 1024×1024 in 2.0s
SDXL 1.0 at 1024x1024 on the RTX 4090 24GB renders a 30-step image in 2.0 seconds, batch of four in…
RTX 4090 24GB Stable Video Diffusion Benchmark: 25-frame in 25s
Deep Stable Video Diffusion benchmark on the RTX 4090 24GB: 25-frame SVD-XT in 25s FP16 / 18s FP8, full VRAM…
Explore GPU Hosting Solutions
From the blog to your next deployment — pick the right platform for your workload.
Tokens/sec Benchmarks
Real-world tokens per second data across every GPU we offer, tested on popular LLMs.
View BenchmarksTTS Latency Benchmarks
Time-to-first-audio for Coqui, Bark, Kokoro, and XTTS-v2 across GPU tiers.
View TTS BenchmarksOCR Speed Benchmarks
Pages per second for PaddleOCR and Tesseract across our GPU server lineup.
View OCR BenchmarksCost per 1M Tokens
What does it cost to process a million tokens on each GPU? Interactive calculator.
Calculate CostDedicated GPU Hosting
Bare-metal servers with a dedicated GPU, NVMe, full root access, and 1Gbps networking from our UK datacenter.
Browse GPU ServersOpen Source LLM Hosting
Deploy LLaMA, Mistral, DeepSeek, and more on dedicated hardware with no per-token API fees.
Explore LLM HostingReady to deploy your AI workload?
Dedicated GPU servers from our UK datacenter. NVMe storage, 1Gbps networking, full root access.
65 Victory Avenue
PE7 1XU, England
GPUAIR © All Rights Reserved.