The NVIDIA A100 80GB is the reference GPU of the modern AI era. Its GA100 Ampere die carries 6,912 CUDA cores and 432 third-generation Tensor Cores with TF32 and BF16, fed by 80 GB of HBM2e at 2,039 GB/s. It introduced Multi-Instance GPU (up to seven isolated partitions) and third-generation NVLink at 600 GB/s. Every major framework, kernel library and serving stack has been optimised for it for years.
packet.ai rents the A100 80GB as a Dedicated single-tenant card for $1.43/hr in Frankfurt, billed hourly, 99% SLA. For European teams that need 80 GB of high-bandwidth memory under GDPR-friendly residency without hyperscaler pricing, it is the obvious pick. Dynamic pricing for this SKU is launching soon.
Why rent the A100 80GB on packet.ai
80 GB of HBM2e at 2 TB/s
Enough to serve a 34B model at FP16 or a 70B model at FP8 on one card, with bandwidth that keeps time-to-first-token low under concurrency. For memory-bound inference the A100 still outperforms newer GDDR cards on latency.
The most mature stack in AI
CUDA, cuDNN, TensorRT, vLLM, DeepSpeed, Megatron, FlashAttention: all were tuned on A100 first. If a paper published a recipe, it runs here without surprises. That maturity is worth real engineering hours.
MIG for right-sized serving
Partition one A100 into up to seven isolated GPU instances with dedicated memory and compute. Run several small models or many concurrent notebooks on a single card with hard isolation.
38% below the field
$1.43/hr Dedicated is 10% under RunPod's $1.59 Secure Cloud rate, 28% below Lambda's single-card price, 47% below CoreWeave and roughly a third of AWS and Google Cloud's per-GPU rate. Only marketplace floors go lower, and they are not dedicated.
EU residency
Frankfurt capacity means data stays in the EU by default. For regulated workloads that need a single-tenant card in Europe, this is the cheapest published way to get one.
A100 80GB pricing on packet.ai
Three ways to buy the same card. Mix them on one account: Dynamic for development and bursty work, Dedicated for production, Clusters when a single node is not enough.
Dynamic
Shared card, scheduler-enforced isolation. Same peak performance and VRAM as Dedicated.
- Hourly billing, no commitment
- SSH-ready in under 5 minutes
- Never pre-empted
Dedicated
Whole card, exclusively yours. Zero scheduler interference, predictable p99.
- 99% SLA
- Single-tenant hardware
- 5–10 minute provisioning
Clusters
Multi-node, InfiniBand fabric, custom storage, named TAM.
- Bare-metal nodes and 8× configurations
- Up to 1,024+ GPUs
- 2–6 week provisioning
How it compares
Published on-demand starting rates for the same GPU at other providers, USD per GPU-hour. packet.ai's reference rate is $1.43/hr Dedicated.
| Provider | A100 80GB rate |
|---|---|
| packet.ai | $1.43/hr Dedicated |
| Vast.ai (marketplace) | $1.28/hrlower by 10% |
| RunPod Secure Cloud | $1.59/hrpacket.ai −10% |
| Hyperstack | $1.60/hrpacket.ai −11% |
| Verda | $1.79/hrpacket.ai −20% |
| Lambda (1×) | $1.99/hrpacket.ai −28% |
| Crusoe | $2.30/hrpacket.ai −38% |
| Vultr | $2.40/hrpacket.ai −40% |
| Modal | $2.50/hrpacket.ai −43% |
| CoreWeave | $2.70/hrpacket.ai −47% |
| AWS p4d (40 GB, 8× node) | $4.10/hrpacket.ai −65% |
| Google Cloud a2-ultragpu | $5.07/hrpacket.ai −72% |
Specifications
| Feature | Value |
|---|---|
| GPU architecture | NVIDIA Ampere (GA100) |
| CUDA cores | 6,912 |
| Tensor Cores | 432 (3rd generation, TF32 / BF16 / FP16 / INT8) |
| Memory | 80 GB HBM2e |
| Memory interface | 5,120-bit |
| Memory bandwidth | 2,039 GB/s |
| FP32 compute | 19.5 TFLOPS |
| TF32 Tensor compute | 156 TFLOPS dense / 312 TFLOPS sparse |
| FP16 / BF16 Tensor compute | 312 TFLOPS dense / 624 TFLOPS sparse |
| INT8 Tensor compute | 624 TOPS dense / 1,248 TOPS sparse |
| FP64 / FP64 Tensor | 9.7 / 19.5 TFLOPS |
| Interconnect | NVLink 3 · 600 GB/s (SXM) · PCIe 4.0 ×16 |
| Multi-instance GPU | Up to 7 instances (10 GB each) |
| Board power | 400 W (SXM4) · 300 W (PCIe) |
What you can run on 80 GB
Where the A100 80GB is the right card, where it is comfortable, and where you should step up.
| Workload | Fit | Notes |
|---|---|---|
| 30–34B LLM inference at FP16 | Single card | Qwen 2.5 32B, Yi-34B, Command-R |
| 70B LLM inference at FP8 / 4-bit | Single card; 2 cards for FP16 | Llama 3.1 70B with NVLink pair |
| Latency-sensitive serving (TTFT under 400 ms) | Strong: 2 TB/s HBM2e | Better than GDDR cards at high concurrency |
| Full fine-tuning to 13B, LoRA to 70B | Single card for LoRA; 4–8 cards for full FT | BF16 training, mature recipes |
| Many small models via MIG | Up to 7 isolated instances | Multi-tenant notebooks or micro-services |
| Scientific computing (FP64) | 19.5 TFLOPS FP64 Tensor | The only card on this list with real FP64 |
| FP4 / FP8 Transformer Engine work | Not supported (Ampere) | Use B200 or RTX 6000 Pro for FP8/FP4 |
Deploy in minutes
- Create an accountSign up at dash.packet.ai. No credit card required to start; billing is usage-based.
- Pick the A100 80GB and a tierChoose Dedicated (Dynamic launching soon), hourly or monthly billing, and the region. Multi-GPU and bare metal are on request.
- DeployFrom the dashboard, the API, or the packet CLI. Dedicated instances are typically SSH-ready in five to ten minutes.
- SSH in and runUbuntu, current NVIDIA drivers and CUDA are in place, with local NVMe attached. Bring your containers or install directly.
# after the instance shows Running in the dashboard ssh ubuntu@<instance-ip> nvidia-smi # confirm the A100 80GB and driver # example: serve a model with vLLM pip install vllm python -m vllm.entrypoints.openai.api_server --model <hf-model-id> --port 8000
CLI and API reference, including how to launch and stop instances programmatically, is in the packet.ai docs.
FAQ
How much is an A100 80GB on packet.ai?
$1.43 per GPU-hour on Dedicated capacity, billed hourly, no minimum term. About $1,044 per month at 730 hours; monthly commits take up to 20% off. Dynamic pricing for this card is launching soon.
SXM or PCIe?
packet.ai lists the A100 80GB HBM2e without a form-factor suffix; check the instance details in the dashboard for the exact variant and NVLink availability on multi-GPU configurations.
Why choose the A100 over the newer RTX 6000 Pro?
Bandwidth and maturity. The A100's 2 TB/s HBM2e beats the RTX 6000 Pro's 1.79 TB/s GDDR7 for latency-sensitive serving at high concurrency, it has NVLink for multi-card tensor parallelism, MIG for partitioning, and every recipe in the literature was written for it. Choose the RTX 6000 Pro for 96 GB, FP8/FP4 and a lower price; choose the A100 for HBM bandwidth, NVLink and EU residency.
How does it compare to the H100?
The H100 SXM has roughly 3.2× the dense FP16 tensor throughput, 1.6× the bandwidth and adds FP8. For training and very high-throughput inference the H100 is the faster card. For memory-bound inference at moderate concurrency the A100 remains excellent value, and on packet.ai it is available today while the H100 is on the notify list.
Can I use MIG?
Yes. The A100 supports up to seven Multi-Instance GPU partitions with isolated memory and compute; configure them with nvidia-smi on your instance.
Which region?
EU (Frankfurt) today. Use the pricing page region filter for the live list; US capacity for this SKU is on the roadmap.
What is pre-installed?
Ubuntu with current NVIDIA drivers and CUDA, SSH access and local NVMe. Bring containers or install frameworks directly.
Hidden fees?
No platform fees, no ingress. Egress $0.04/GB. Up to 2 TB local NVMe included; object storage from $0.018/GB-month.
Same silicon. Smarter economics.
A100 80GB from $1.43/hr, billed hourly, live in EU · Frankfurt. No quota approval, no credit card to start.