24GB
VRAM
GDDR6 memory
AIVory · GPU Marketplace
Live spot pricingNVIDIA RTX A5000 — professional Ampere at the price of a legacy card.
The RTX A5000 brings 24 GB of GDDR6 with Ampere tensor cores and professional-grade ECC memory support into the spot market at $0.17/hr — matching the RTX A4000 as the cheapest Ampere professional card available. For teams that need enterprise driver certification with 24 GB of VRAM, the A5000 delivers workstation reliability at commodity pricing.
At a glance
RTX A5000 specifications.
Key hardware specs that determine what workloads this GPU handles.
768 GB/s
Memory Bandwidth
peak throughput
230W
TDP
thermal design power
Ampere
Architecture
NVIDIA GPU architecture
Spot pricing
RTX A5000: 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 A5000.
Tested model-GPU pairings with notes on why each is a good fit.
Mistral Nemo 12B
12B in FP16 uses ~24 GB — tight on VRAM but the A5000's 768 GB/s bandwidth delivers fast tokens. INT8 quantization at ~12 GB gives more headroom for concurrent users.
View model pricing →
Phi-4 14B
14B in INT8 at ~14 GB leaves 10 GB for KV cache. The professional drivers ensure stable multi-day uptime, making this a reliable always-on reasoning endpoint.
View model pricing →
Gemma 2 9B
9B parameters in FP16 need ~18 GB. Fits on the A5000 with 6 GB free, and the 768 GB/s bandwidth — the highest of any 24 GB Ampere card — keeps generation fast.
View model pricing →
Use cases
What the RTX A5000 is built for.
-
Highest-bandwidth 24 GB Ampere GPU for inference
At 768 GB/s, the RTX A5000 has the fastest memory bandwidth of any 24 GB Ampere GPU — the same bandwidth as the A6000's 48 GB. This makes it the fastest Ampere card for token generation on models that fit in 24 GB, outpacing the A10G (600 GB/s) and RTX 3090 (936 GB/s GDDR6X, but consumer-grade). At $0.17/hr, it's also one of the cheapest.
-
Professional workstation AI with NVLink support
The RTX A5000 supports NVLink, enabling two cards to share a unified 48 GB memory pool with 112 GB/s interconnect. This makes a dual-A5000 workstation ($0.34/hr total) a cost-effective alternative to a single A6000 ($0.33/hr) — same total VRAM, with the flexibility to run independent models on each card when NVLink isn't needed.
-
ISV-certified inference for healthcare and scientific computing
Medical imaging AI, computational chemistry, and scientific simulation tools often require ISV-certified GPU drivers. The RTX A5000's professional driver stack is validated against applications like MONAI, Clara, and ANSYS. At $0.17/hr, it's the cheapest way to run AI inference in an ISV-certified environment.
FAQ
Common questions.
RTX A5000 vs RTX 3090 — both 24 GB Ampere, which to pick?
The RTX 3090 has higher raw bandwidth (936 GB/s GDDR6X vs 768 GB/s GDDR6) and costs less ($0.10/hr vs $0.17/hr). But the A5000 offers ECC memory support, professional drivers, NVLink, and lower power draw (230W vs 350W). Choose the RTX 3090 for maximum speed-per-dollar on inference. Choose the A5000 when you need enterprise driver certification, ECC, or NVLink multi-GPU.
Is the RTX A5000 being phased out?
Yes, NVIDIA's professional lineup has moved to Ada Lovelace (RTX 5000 Ada replaces the A5000). But the A5000 remains widely available on spot markets at deeply discounted prices. The deprecation is actually good news for spot users — providers are offloading A5000 inventory at rock-bottom rates with plenty of supply.
RTX A5000 vs A10G — both 24 GB Ampere, what's the difference?
The A5000 is a professional workstation card; the A10G is a cloud-native data center card. The A5000 has higher bandwidth (768 vs 600 GB/s), display output, NVLink, and professional drivers. The A10G integrates with AWS/GCP infrastructure and is available in their spot markets. If you're on AWS, the A10G is the default. If you want the cheapest 24 GB Ampere from any provider, the A5000 at $0.17/hr wins.
Can I run LLM training on the RTX A5000?
Fine-tuning models up to 9B parameters in BF16 fits within the 24 GB envelope. LoRA fine-tuning of 13B-14B models is possible with gradient checkpointing. Full pre-training should target larger GPUs — the A5000's GDDR6 bandwidth is adequate for inference but becomes a bottleneck for training throughput at scale. The A5000 is a strong inference and fine-tuning card.
Rent a RTX A5000. Right now.
Spot pricing, per-second billing, no commitment.
Browse the live marketplace, pick your GPU, deploy in one click. Credits from $10.