GPU catalog
Available GPUs
Meshia's visible GPU catalog is generated from the same product catalog used by the launcher. Live rates and capacity can change by provider, region, and time of day; the table below shows the fallback rate Meshia uses when a provider has not returned a live quote.
| GPU | VRAM | Fallback rate | Best for |
|---|---|---|---|
| RTX A5000 | 24 GB | $0.18/hr fallback | low-cost eval replay and smoke reruns for small receipt workloads |
| A10G | 24 GB | $1.15/hr fallback | Steady 24 GB receipt worker for baseline vs candidate comparison and moderate eval replay |
| RTX 3090 | 24 GB | $0.25/hr fallback | Budget 24 GB CUDA option for smoke reruns, eval replay, and quick verifier passes |
| T4 16GB | 16 GB | $0.40/hr fallback | Cost-effective legacy CUDA option for lightweight verifier passes and log-only proof checks |
| P4 8GB | 8 GB | $0.23/hr fallback | Ultra-low cost legacy GPU for tiny eval replay jobs and receipt sanity checks |
| P100 16GB | 16 GB | $0.92/hr fallback | Legacy Pascal-class GPU for low-cost eval replay and older verifier harnesses |
| RTX 4090 | 24 GB | $0.39/hr fallback | Fast 24 GB card for smoke reruns, baseline vs candidate checks, and report generation |
| L4 24GB | 24 GB | $0.55/hr fallback | Efficient Ada card for batched eval replay, video-adjacent evidence, and receipt generation |
| L40S | 48 GB | $2.88/hr fallback | 48 GB card for large eval batches, vision-heavy verifier passes, and report generation |
| V100 16GB | 16 GB | $1.32/hr fallback | Legacy Volta option for compatibility eval replay and scientific verifier workloads |
| RTX 5090 | 32 GB | $0.79/hr fallback | 32 GB Blackwell card for fast candidate reruns and report iteration |
| RTX A6000 | 48 GB | $0.38/hr fallback | 48 GB card for large eval batches, multi-case verifier passes, and stable long runs |
| RTX 6000 Ada | 48 GB | $0.85/hr fallback | Stable 48 GB Ada card for vision eval replay and report generation |
| A100 PCIe 80GB | 80 GB | $1.37/hr fallback | 80 GB workhorse for long-context replay, large eval batches, and receipt packaging |
| A100 SXM 40GB | 40 GB | $2.19/hr fallback | 40 GB A100 for baseline vs candidate comparison and steady verifier passes |
| A100 SXM 80GB | 80 GB | $1.60/hr fallback | 80 GB A100 for large eval batches, long-context replay, and receipt generation |
| RTX PRO 6000 96GB | 96 GB | $1.94/hr fallback | 96 GB Blackwell card for memory-heavy verifier passes and report generation |
| H100 NVL 94GB | 94 GB | $2.98/hr fallback | High-throughput Hopper card for large eval batches and long-context replay |
| H100 SXM 80GB | 80 GB | $3.09/hr fallback | Fast Hopper card for verifier passes, report generation, and candidate reruns |
| H200 141GB | 141 GB | $4.13/hr fallback | 141 GB memory tier for long-context replay and the largest receipt workloads |
| B200 180GB | 180 GB | $6.88/hr fallback | 180 GB Blackwell card for verifier passes, report generation, and extreme eval replay |
| B300 288GB | 288 GB | $7.98/hr fallback | 288 GB memory tier for extreme long-context replay and report generation |
| TPU v2 | 8 GB | $1.29/hr fallback | Legacy TPU option for compatibility eval replay and receipt checks |
| TPU v3 | 32 GB | $2.30/hr fallback | High-memory TPU option for legacy eval replay and verifier passes |
| TPU v4 | 32 GB | $1.61/hr fallback | High-throughput TPU option for large eval batches and receipt generation |
| TPU v5p | 95 GB | $3.68/hr fallback | Largest Cloud TPU option for long-context replay and report generation |
| TPU v5e | 16 GB | $1.38/hr fallback | Cost-effective TPU option for batched eval replay and verifier passes |
| TPU v6e | 32 GB | $2.53/hr fallback | Cloud TPU option for high-performance eval replay and receipt workloads |
Availability
GPUs are provisioned on demand. Availability varies by live provider capacity. The current catalog includes: RTX A5000, A10G, RTX 3090, T4 16GB, P4 8GB, P100 16GB, RTX 4090, L4 24GB, L40S, V100 16GB, RTX 5090, RTX A6000, RTX 6000 Ada, A100 PCIe 80GB, A100 SXM 40GB, A100 SXM 80GB, RTX PRO 6000 96GB, H100 NVL 94GB, H100 SXM 80GB, H200 141GB, B200 180GB, B300 288GB, TPU v2, TPU v3, TPU v4, TPU v5p, TPU v5e, TPU v6e.
If a GPU isn't available, Meshia queues your request and provisions it as soon as one opens up. You're not charged during the queue wait.
Recommendations by use case
Fine-tuning 7B to 13B models (LoRA/QLoRA): start with a 24 GB or 32 GB card such as RTX 4090, RTX 5090, A5000, A10G, L4, or RTX 3090.
Fine-tuning 30B+ models: choose 48 GB+ options such as A6000, L40S, A100, H100, or H200 depending on capacity and budget.
Pre-training or full fine-tuning: A100 or H100. These workloads are compute-bound. The H100's FP8 support and higher memory bandwidth make a real difference for large training runs.
Inference and prototyping: start with the lowest-cost card that fits your model in VRAM, then scale up if latency or batch size demands it.
Running 70B+ models: H200. The 141 GB VRAM lets you fit models that won't fit on anything else in a single GPU.
CUDA and driver versions
All GPUs run CUDA 12.4+ with the latest stable NVIDIA drivers. PyTorch, JAX, and TensorFlow are pre-installed in the pod environment. The agent handles versions. If your code needs a specific one, just ask.