LLM Inference Latency Estimator
The spec sheet says H100 = 15× T4 TFLOPS. But a 7B model is 150× faster on H100. The bottleneck isn't compute — it's memory bandwidth. 12 models × 10 GPUs × 5 quantization levels, now including FP8, with real benchmark data.
Model Throughput by GPU →The FP8 Trap: Why 2× Becomes 1.15×
FP8 promises 2× and bills 1.15× to 1.84× depending on which switch you flipped. NVIDIA's own TensorRT-LLM numbers, INT4 beating FP8 at batch 1, the 7,010-token KV-cache break-even, and why an A100 runs FP8 at exactly BF16 speed without ever raising an error.
Read the FP8 Briefing →GPU Training Cost Estimator
arXiv says 72 GPU-hours. Your AWS bill says 3× that. Spot/preemptible pricing across 6 providers, on-demand real rates. No guesswork — just the per-hour truth.
Run 6-Provider Cost Matrix →LLM Fine-tuning Cost Estimator
Compare fine-tuning across AWS Bedrock, Google Vertex AI, Azure OpenAI, and self-hosted GPU. Full fine-tuning vs LoRA vs QLoRA. Training cost, GPU hours, inference pricing, and hidden hosting/idle fees (Azure $1,241/mo). 2026 pricing with per-platform trap warnings.
4-Platform FT Cost Model →LLM Params →FP16 VRAM
7B × 2 bytes = 14 GB. Then KV cache. Then Adam optimizer states. Then framework overhead. Your 16 GB GPU just became a bottleneck you didn't see coming.
FP16 VRAM Profiler →LLM Params →INT4 VRAM
Same 70B model. FP16 = 140 GB — you need 4× A100s. INT4 = 35 GB — one A100. You're not debating perplexity. You're debating whether the model loads at all.
INT4 VRAM Profiler →LLM API Pricing Calculator
17 models across 6 providers. Token pricing pages tell one story. Your workload's chat-vs-code-vs-summarization ratio tells another. Prompt caching cuts cached input to as little as 5% of list. Batch processing: 50%. Model the real bill before it lands.
Compare Provider Pricing →GPU Model Fit Matrix: Will It Fit on One Card?
Will it fit? A 70B at FP16 is 140 GB of weights alone. Pick a model, a quantization and a context length: get a fit / tight / fail verdict, the card count, and the cheapest card that actually runs it.
Check the Fit →How Many GPUs Do I Need? Inference Capacity Planner
The reverse question. Give it a traffic target and a latency budget instead of a card: get the GPU count, the tensor-parallel layout, the peak headroom and what the same load costs on an API.
Size From Traffic →Self-Host vs API: LLM Breakeven
One A100 80GB is $1,307/month whether or not a request arrives. The same 70B traffic over an API is $176. Size VRAM, GPUs, peak load and MLOps time — then find where self-hosting wins.
Run the Comparison →