48GB
VRAM
GDDR6 memory
AIVory · GPU Marketplace
Live spot pricingNVIDIA A40 — 48 GB with ECC for the price of a coffee.
The A40 packs 48 GB of ECC-protected GDDR6 and Ampere tensor cores into a passively-cooled data center form factor. At $0.35/hr on spot, it's the cheapest 48 GB GPU available — undercutting even the A6000 — and the ECC memory means bit-flip errors won't silently corrupt your model outputs during long-running inference sessions.
At a glance
A40 specifications.
Key hardware specs that determine what workloads this GPU handles.
696 GB/s
Memory Bandwidth
peak throughput
300W
TDP
thermal design power
Ampere
Architecture
NVIDIA GPU architecture
Spot pricing
A40: 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 A40.
Tested model-GPU pairings with notes on why each is a good fit.
Mistral Small 3.2 24B
24B in FP16 uses ~48 GB — a tight fit, but quantizing to INT8 drops it to 24 GB, leaving half the VRAM for generous KV caching and concurrent users.
View model pricing →
Mixtral 8x7B
47B total parameters with 13B active per forward pass. The full MoE model in INT8 needs ~24 GB, fitting on the A40 with room for all 8 experts to stay resident.
View model pricing →
Llama 3.3 70B
Fits in INT4 quantization at ~35 GB. Not ideal for high-throughput production, but at $0.35/hr the A40 makes 70B inference accessible for development and testing.
View model pricing →
Use cases
What the A40 is built for.
-
Budget-friendly 48 GB inference for MoE and mid-range models
The A40 at $0.35/hr is the cheapest 48 GB GPU on the spot market — cheaper than the A6000 ($0.33/hr with less ECC protection) and far cheaper than the L40 ($0.51/hr) or L40S ($0.57/hr). For workloads where you need 48 GB of VRAM but don't need the absolute fastest throughput, the A40 delivers the most memory per dollar.
-
ECC-protected inference for regulated and financial workloads
The A40's ECC memory corrects single-bit errors in real time, preventing silent data corruption during inference. This matters for financial modeling, healthcare AI, and any workload where a wrong output has legal or monetary consequences. Non-ECC GPUs like the RTX 4090 can produce bit flips under sustained load — the A40 eliminates that risk.
-
Batch processing and offline analysis on large models
Document summarization, classification, and embedding generation at scale don't need sub-100ms latency — they need cheap GPU-hours with enough VRAM. The A40 handles 30B models in FP16 or 70B in INT4 at $0.35/hr. Process a million documents overnight for under $3.
FAQ
Common questions.
A40 vs A6000 — which 48 GB Ampere GPU should I pick?
The A40 ($0.35/hr) and A6000 ($0.33/hr) are priced similarly with the same 48 GB VRAM and Ampere architecture. Key differences: the A40 has ECC memory (error correction) and is passively cooled for data centers; the A6000 has higher bandwidth (768 vs 696 GB/s) and display outputs for workstation use. For inference reliability, the A40's ECC is the tiebreaker. For raw throughput, the A6000's extra bandwidth gives it a slight edge.
How does the A40 compare to the newer L40 and L40S?
The L40/L40S use Ada Lovelace architecture (one generation newer) with FP8 support and higher tensor throughput. They cost more ($0.51-$0.57/hr) but deliver roughly 40% better inference performance per GPU-hour. If your workload is latency-sensitive, the L40S wins on speed. If it's throughput-at-minimum-cost, the A40 wins on price.
Does the A40 support multi-GPU configurations?
Yes. A40 servers support up to 4 or 8 GPUs connected via PCIe 4.0. NVLink is not available on the A40 (unlike the A100), so inter-GPU bandwidth is limited to PCIe speeds (~32 GB/s per direction). This is fine for independent model instances but too slow for tensor-parallel inference on models sharded across cards. Use multi-A40 for multi-model hosting, not multi-GPU sharding.
Is the A40 suitable for model training?
For fine-tuning models up to ~15B parameters in BF16, the A40 is usable. LoRA and QLoRA fine-tuning of 30B-70B models works within the 48 GB envelope. Full pre-training at scale should use A100 or H100 for the HBM bandwidth advantage (2.0+ TB/s vs 696 GB/s). The A40 is a training-capable inference card, not a training-first card.
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 A40. Right now.
Spot pricing, per-second billing, no commitment.
Browse the live marketplace, pick your GPU, deploy in one click. Credits from $10.