6GB
VRAM
GDDR6 memory
AIVory · GPU Marketplace
Live spot pricingNVIDIA RTX A2000 — the cheapest GPU on the spot market, period.
At $0.12/hr, the RTX A2000 is the absolute lowest-cost GPU you can rent for tensor-core-accelerated compute. Six gigabytes of GDDR6 and Ampere architecture in a compact, 70-watt single-slot form factor. VRAM limits you to 2B-3B models, but for embedding generation, classification, and tiny model inference, nothing else comes close on price.
At a glance
RTX A2000 specifications.
Key hardware specs that determine what workloads this GPU handles.
288 GB/s
Memory Bandwidth
peak throughput
70W
TDP
thermal design power
Ampere
Architecture
NVIDIA GPU architecture
Spot pricing
RTX A2000: live hourly rates.
Every provider offering this GPU on the spot market, sorted cheapest first.
Prices in USD per GPU-hour · spot instances · sorted cheapest first
Recommended models
AI models that run well on RTX A2000.
Tested model-GPU pairings with notes on why each is a good fit.
Gemma 2 2B
2B in FP16 at 4 GB leaves 2 GB for KV cache. The largest model that runs comfortably on the 6 GB RTX A2000 — surprisingly capable for summarization and simple Q&A.
View model pricing →
TinyLlama 1.1B
1.1B in FP16 at just 2.2 GB. Run with 3.8 GB free for generous KV caching and high concurrency — multiple simultaneous conversations at $0.12/hr.
View model pricing →
Qwen 2.5 0.5B
500M parameters in FP16 at 1 GB. Ultra-lightweight classification and extraction tasks at the lowest possible GPU cost. Run 5-6 instances on a single card.
View model pricing →
Use cases
What the RTX A2000 is built for.
-
Text embedding and vector indexing at the lowest GPU cost
Embedding models like E5-small, MiniLM, and all-MiniLM-L6 use 0.2-0.5 GB of VRAM. The RTX A2000's 6 GB handles large batch embedding with 5+ GB free for input buffers. At $0.12/hr, indexing millions of documents for vector search costs $2.88/day — less than a cup of coffee for enterprise-scale RAG infrastructure.
-
Classification and NER with compact transformer models
Named entity recognition, sentiment analysis, and text classification with DistilBERT, DeBERTa-v3-small, or custom fine-tuned models require 0.3-1 GB of VRAM. The RTX A2000 processes these at Ampere tensor-core speeds for $0.12/hr, making it viable to run GPU-accelerated NLP pipelines that would otherwise fall back to CPU for cost reasons.
-
GPU compute testing in CI/CD pipelines
Automated tests that validate CUDA compatibility, model loading, and inference pipeline correctness need the cheapest available GPU. The RTX A2000 at $0.12/hr is perfect for CI/CD GPU runners — run your test suite on real tensor core hardware at $2.88/day, catching GPU-specific bugs that CPU-only CI misses.
FAQ
Common questions.
Can the RTX A2000 run any useful LLM?
Yes, but only very small ones. Gemma 2 2B in FP16 (4 GB) and TinyLlama 1.1B (2.2 GB) both fit. Llama 3.2 3B in FP16 needs 6 GB — it fits but with zero headroom for KV cache, making it impractical for multi-turn conversations. For any model above 3B, you need more VRAM. The RTX A2000 excels at embedding, classification, and sub-2B inference.
RTX A2000 6 GB vs 12 GB — which am I getting?
NVIDIA released both a 6 GB and 12 GB variant. Spot listings should specify the VRAM size. The 12 GB variant opens up 3B-7B models in INT8, which dramatically expands the card's usefulness. At $0.12/hr, either variant is a bargain — but check the listing carefully before renting.
Is the RTX A2000 faster than a CPU for inference?
For transformer models, significantly. Even a high-end 64-core CPU running llama.cpp serves Gemma 2 2B at 10-15 tokens/second. The RTX A2000's Ampere tensor cores deliver 30-50+ tokens/second on the same model. The GPU advantage is smaller than with larger cards (fewer tensor cores, lower bandwidth), but the price-performance ratio still favors the GPU for any workload that fits in 6 GB.
Why is the RTX A2000 cheaper than every other GPU?
Three factors: tiny chip (GA106), low VRAM (6 GB GDDR6), and market positioning. NVIDIA designed the A2000 as the entry-level professional card for CAD and light compute. On spot markets, these cards come from workstations being retired or upgraded — providers price them at $0.12/hr because demand for 6 GB GPUs is low. This makes them a hidden gem for workloads that don't need more VRAM.
Need 8, 16 or 32 GPUs?
Pick the GPU, the count and how long. See every live offer, priced per GPU hour and for the whole job.
Do you run GPUs or an inference API?
Send a test key and we benchmark your endpoint. You keep the numbers either way.
Rent a RTX A2000. Right now.
Spot pricing, per-second billing, no commitment.
Browse the live marketplace, pick your GPU, deploy in one click. Credits from $10.