1. Fit the model in VRAM
Model weights are only the baseline. Leave room for the framework, activations, KV cache and the batch size you actually need.
Crypto-only workspace funding
Pay for cloud GPU rental with crypto.
Choose the GPU and estimate its runtime first, then fund the workspace through the hosted crypto checkout. Use only the asset and blockchain network displayed on the live invoice.
24–288 GB GPU memoryFrom $0.16 per GPU-hourParis and Frankfurt regionsCrypto-only workspace funding
GPU specifications and rates
Compare the hardware that changes the result.
VRAM answers “will it fit?” Memory bandwidth and supported precision help explain throughput. The only reliable performance test is still your model, batch, framework and output target.
2. Check memory bandwidth
LLM decoding and other memory-bound jobs can benefit when weights and cache move faster. Capacity alone does not predict throughput.
3. Match the precision
FP16, BF16, FP8 and FP4 peaks measure different operating modes. Use the precision your model and software stack can run safely.
4. Price the completed job
The cheapest hourly GPU is not always the cheapest result. Measure runtime, accepted outputs, storage and retry cost together.
| GPU | Memory | Memory bandwidth | Official peak metric | On-demand | Decision fit | Source |
|---|---|---|---|---|---|---|
| RTX A5000Ampere | 24 GB GDDR6 ECC | 768 GB/s | 222.2 TFLOPS Tensor* Effective Tensor peak with sparsity | $0.16/hr$3.84/24h | Low-cost LoRA, diffusion and general CUDA development. | NVIDIA RTX A5000 datasheet |
| RTX 4090Ada | 24 GB GDDR6X | 1,008 GB/s | 1,321 AI TOPS* Ada vendor AI peak with sparsity | $0.34/hr$8.16/24h | Strong price/performance for image, video and 24 GB inference. | NVIDIA Ada architecture |
| RTX 5090Blackwell | 32 GB GDDR7 | 1,792 GB/s | 3,352 AI TOPS* Blackwell vendor AI peak with sparsity | $0.69/hr$16.56/24h | Fast single-GPU generation with 32 GB and native FP4 support. | NVIDIA RTX 5090 specifications |
| L40SAda | 48 GB GDDR6 ECC | 864 GB/s | 1,466 TFLOPS FP8* Ada Tensor peak with sparsity | $0.79/hr$18.96/24h | 48 GB ECC for inference, fine-tuning, rendering and video. | NVIDIA L40S specifications |
| A100 80GBAmpere | 80 GB HBM2e | 1.94–2.04 TB/s | 312 TFLOPS FP16/BF16 624 TFLOPS with sparsity | $1.19/hr$28.56/24h | Mature 80 GB platform for training, fine-tuning and HPC. | NVIDIA A100 specifications |
| H100 NVLHopper | 94 GB HBM3 | 3.9 TB/s | 1,671 TFLOPS FP16/BF16* Hopper Tensor peak with sparsity | $2.59/hr$62.16/24h | 94 GB per GPU and NVLink for high-throughput LLM inference. | NVIDIA H100 specifications |
| H100 SXMHopper | 80 GB HBM3 | 3.35 TB/s | 1,979 TFLOPS FP16/BF16* Hopper Tensor peak with sparsity | $2.69/hr$64.56/24h | High-throughput training and FP8 inference with 900 GB/s NVLink. | NVIDIA H100 specifications |
| H200 SXMHopper | 141 GB HBM3e | 4.8 TB/s | 1,979 TFLOPS FP16/BF16* Hopper Tensor peak with sparsity | $3.59/hr$86.16/24h | 141 GB and higher bandwidth for long-context and large-model inference. | NVIDIA H200 specifications |
| B200Blackwell | 180 GB HBM3e | Up to 8 TB/s | 9 PFLOPS FP4 dense† Per-GPU equivalent from HGX total | $5.98/hr$143.52/24h | 180 GB for frontier training and low-precision inference. | NVIDIA HGX B200 specifications |
| B300Blackwell Ultra | 288 GB HBM3e | Up to 8 TB/s | 13.5 PFLOPS FP4 dense† Per-GPU equivalent from HGX total | $6.94/hr$166.56/24h | 288 GB for frontier reasoning, very long context and large batches. | NVIDIA HGX B300 specifications |
Peak figures come from NVIDIA and are not cross-GPU benchmark scores. * Vendor figure includes sparsity. † Per-GPU equivalent calculated from the official eight-GPU HGX total. Different precisions, sparsity modes and form factors are not directly interchangeable.
Catalog rates reviewed September 1, 2026. Regions: EU West · Paris · EU Central · Frankfurt
Model memory planning
Estimate weight memory before choosing VRAM.
A useful first approximation is parameter count multiplied by bytes per weight: 2 bytes for FP16/BF16, 1 for INT8 and roughly 0.5 for 4-bit weights.
These are raw weight estimates, not complete runtime requirements. KV cache grows with context and batch size; training also needs activations, gradients and optimizer state.
Parameter class
7B class
- FP16 / BF16
- ≈14 GB
- INT8
- ≈7 GB
- 4-bit weights
- ≈3.5 GB
24 GB GPUs usually provide practical inference headroom.
Parameter class
13B class
- FP16 / BF16
- ≈26 GB
- INT8
- ≈13 GB
- 4-bit weights
- ≈6.5 GB
32–48 GB gives room for runtime overhead and larger context.
Parameter class
70B class
- FP16 / BF16
- ≈140 GB
- INT8
- ≈70 GB
- 4-bit weights
- ≈35 GB
48 GB is a tight low-bit floor; 80–288 GB adds useful headroom.
Workload recommendations
Match the GPU to the bottleneck.
Creative workloads often reward RTX price/performance. Training and large-model inference increasingly depend on ECC memory, Tensor precision, memory bandwidth and multi-GPU interconnect.
RTX 4090 · RTX 5090 · L40S
Stable Diffusion & ComfyUI
Start with the 4090 for value, move to the 5090 for 32 GB and stronger low-precision acceleration, or L40S for 48 GB ECC and data-center operation.
Read workload guideL40S · H100 NVL · H200
LLM inference
Choose by model footprint and context. L40S covers many quantized models; H100 NVL and H200 add memory, bandwidth and stronger Tensor throughput.
Read workload guideA100 · H100 SXM · B200/B300
Training & fine-tuning
A100 remains a mature baseline. H100 adds Transformer Engine and FP8; Blackwell targets the largest low-precision training and reasoning jobs.
Read workload guideRTX 4090 · RTX 5090 · L40S
Rendering & video
RTX cards offer strong creative value. L40S combines 48 GB ECC with data-center graphics, AI and media acceleration for persistent pipelines.
Read workload guidePaying with crypto funds ordinary cloud compute. Mining is not advertised as a supported workload and still requires explicit policy confirmation.
Crypto wallet
Pay after you know which GPU you need.
The payment flow is intentionally separate from GPU selection: choose the resource, estimate the job, then add workspace credit with an exact asset and network from the hosted invoice.
Choose the GPU first
Check memory fit, expected runtime, region and hourly rate before opening the wallet.
Fund through the hosted checkout
The minimum initial top-up is $50. It is wallet credit, not an extra service fee. Use only the asset, amount and network shown on the live invoice.
Create after confirmation
Configure resources directly in the dashboard. Prepaid usage uses wallet credit; monthly rentals have a separate checkout.
Which cryptocurrencies and networks are accepted?+
Bitcoin, USDC, USDT or another asset can be used only when that exact asset and blockchain network appear on the current CoinGate invoice. The live checkout is authoritative because merchant settings and network availability can change.
Provider currency and network listWhat happens while a payment confirms or expires?+
Pending and confirming orders do not add workspace credit. Wait for final confirmation and do not pay twice. If the invoice expires, generate a new one so its amount, address, network and timer are current.
Official order status definitionsWhat if the amount or network is wrong?+
Blockchain transfers can be irreversible. Keep the order ID and transaction hash, then contact payment support. GPURento does not auto-credit an order until the server verifies the matching final provider state.
Official payment instructionsWhat security and identity rules apply?+
Use only the hosted HTTPS checkout and never share a seed phrase or private key. Crypto is the payment rail, not a promise of anonymity; provider or platform risk and compliance checks may still apply.
Hosted crypto checkout overviewGPU rental with crypto FAQ
The remaining payment questions.
Performance and prices are covered above; these answers focus on funding, account access and workload policy.
Can I rent a GPU with Bitcoin or USDC?+
Yes, when Bitcoin or USDC and the exact blockchain network appear on the live hosted invoice. The available payment routes can change, so the checkout—not a static list—is authoritative.
Is the minimum crypto top-up a fee?+
No. The deposit becomes prepaid wallet credit. The $50 minimum is a deposit limit, not a fee or an amount charged to unlock the site.
When does the workspace balance update?+
Only after the payment provider reports a final confirmed order and GPURento verifies that order from the server. Pending or confirming transfers do not add wallet credit.
What if the invoice expires or I use the wrong network?+
Generate a new invoice after expiry. A wrong-network transfer can be irreversible and is not auto-credited; keep the order ID and transaction hash and contact payment support.
Can unused wallet credit be withdrawn or refunded?+
There is no self-service crypto withdrawal. Unused credit remains in the workspace balance; any refund depends on the applicable payment and service terms and may require manual reconciliation.
Does paying with crypto mean mining is allowed?+
No. Crypto describes the payment method, not the workload policy. Mining is not advertised as supported and requires explicit confirmation before any workload is submitted.
Hardware specifications use linked NVIDIA product documentation. Payment behavior uses official CoinGate documentation. Last reviewed September 1, 2026.
Crypto-funded GPU cloud
Choose by memory and throughput—not by the biggest model name.
Compare the catalog, estimate runtime and storage, then create the workspace that matches the job.