WORKLOAD TRADE-OFF MATRIX

GPU Comparison Engine

Select any two GPUs for a side-by-side evaluation of VRAM capacity, memory bandwidth, TFLOPS throughput, and live spot pricing across workload profiles.

Methodology: How GPU Comparison Works

OpenGPU Radar compares GPU instances across four dimensions: Price, Performance, Memory, and Availability. Pricing data is aggregated from Spheron, RunPod, Vast.ai, and Lambda Labs APIs, refreshed every 6 hours.

Cost per 1M Tokens = (Hourly GPU Cost) / (Tokens/Second × 1,000,000) Lower values indicate better economics. Throughput Index = Effective Tokens/Second (vLLM FP8, Llama 70B) Measured end-to-end output throughput including KV-cache retrieval. Example: H100 SXM5 at $1.89/hr delivers ~120 tok/s → $1.89 / (120 × 1M) = $0.000158 per 1M tokens

Price Data Source

Cloud Provider APIs

Refresh Rate

Every 6 hours

Baseline GPU

RTX 4090 ($0.34/hr)

Benchmarks

vLLM FP8 inference

Hardware Requirements by Workload

Minimum VRAM and GPU specs required for each workload category. Determined via entity-graph VRAM calculations.

Workload	Min GPU	Min VRAM	Precision	Recommended GPU
70B LLM Inference	H100 SXM5	80 GB	FP8	H200 or B200
30B LLM Inference	L40S	48 GB	FP8	A100 80GB
7B LLM Inference	RTX 4090	24 GB	FP16	RTX 4090
Fine-Tuning (LoRA)	A100 80GB	80 GB	FP16	H100 × 2
Pre-Training	B200	192 GB	FP8	B200 × 8
Vision Model (72B)	H200	141 GB	FP8	B200