NVIDIA A100 80GB PCIe Price, Specs & Performance
- Feb 9, 2024
- 4 min read
Updated: 15 hours ago
The NVIDIA A100 80GB PCIe is a data center GPU built on the Ampere architecture, designed for AI training, AI inference, and high-performance computing (HPC).
NVIDIA A100 80GB PCIe GPUs
✔️ 5-Year Warranty – No Risk: Pay Only After Testing
This guide covers full specifications, FP64 clarification, memory bandwidth, MIG support, and realistic market pricing.
NVIDIA A100 80GB PCIe – Technical Specifications
Specification | NVIDIA A100 80GB PCIe |
Architecture | NVIDIA Ampere |
GPU Memory | 80GB HBM2e |
Memory Bandwidth | 1,935 GB/s |
CUDA Cores | 6,912 |
Tensor Cores | 432 (3rd Generation) |
FP64 (CUDA Cores) | 9.7 TFLOPS |
FP64 (Tensor Cores) | 19.5 TFLOPS |
TF32 (Tensor Cores) | 156 TFLOPS |
FP16 (Tensor Cores) | 312 TFLOPS |
Interface | PCIe Gen4 x16 |
TDP | 300W |
Cooling | Passive |
MIG | Up to 7 GPU instances |
These values apply specifically to the 80GB PCIe variant, not the SXM model.
FP64 Performance Explained (9.7 vs 19.5 TFLOPS) - NVIDIA A100 80GB PCIe
The NVIDIA A100 80GB PCIe lists two FP64 figures:
9.7 TFLOPS → Standard double precision via CUDA cores
19.5 TFLOPS → FP64 using Tensor Cores
The higher number applies when workloads are optimized for Tensor Core acceleration. For traditional CUDA-only double-precision workloads, 9.7 TFLOPS is the relevant baseline. This distinction is important when sizing HPC clusters.
Memory and Bandwidth - NVIDIA A100 80GB PCIe
The NVIDIA A100 80GB PCIe delivers:
80GB HBM2e memory
1,935 GB/s memory bandwidth
High bandwidth is critical for:
Large transformer model training
Memory-bound simulations
Multi-user GPU partitioning
For comparison, the NVIDIA Tesla V100 32GB provides 900 GB/s bandwidth and 32GB memory. The A100 significantly increases both capacity and throughput.
The SXM version of the A100 reaches ~2,039 GB/s due to higher clocks and power limits.
Compute Performance by Precision Mode - NVIDIA A100 80GB PCIe
FP64 (CUDA): 9.7 TFLOPS
FP64 (Tensor Core): 19.5 TFLOPS
TF32: 156 TFLOPS
FP16: 312 TFLOPS
Real-world performance depends on:
Model architecture
Software stack (CUDA, cuDNN, TensorRT)
Precision mode
PCIe bandwidth
CPU platform
Performance claims such as “20X faster” or “249X vs CPU” originate from NVIDIA benchmark scenarios (e.g., BERT-Large inference with optimized INT8 or sparsity). Actual results vary.
MIG (Multi-Instance GPU) Support - NVIDIA A100 80GB PCIe
The NVIDIA A100 80GB PCIe supports Multi-Instance GPU (MIG). It can be partitioned into up to seven isolated GPU instances.
MIG is useful for:
Multi-tenant AI environments
Kubernetes deployments
Inference workloads
Resource isolation in shared clusters
Each MIG profile allocates defined compute cores and memory slices.
Power and Cooling Requirements - NVIDIA A100 80GB PCIe
NVIDIA A100 80GB PCIe Power
300W TDP per GPU
Adequate PSU headroom required in multi-GPU systems
NVIDIA A100 80GB PCIe Cooling
Passive cooling design
Requires high-airflow rack server chassis
NVIDIA A100 80GB PCIe PCIe Compatibility
PCIe Gen4 recommended
PCIe Gen3 supported (reduced bandwidth)
The A100 PCIe is a compute-only GPU and does not include display outputs.
NVIDIA A100 80GB PCIe Price (New & Refurbished)
The price of the NVIDIA A100 80GB PCIe depends on:
Market demand cycles
Supply from hyperscalers
Warranty coverage
Region
Bulk quantity
There is no fixed global street price.
Typical Market Price Range (2024–2026) - NVIDIA A100 80GB PCIe
New Units (when available): NVIDIA A100 80GB PCIe
Historically: €18,000 – €25,000+ per GPU
Availability from distribution is limited due to focus on newer generations.
Refurbished / Secondary Market Units: NVIDIA A100 80GB PCIe
Typically: €9,000 – €16,000 per GPU
Price varies based on:
Warranty length (1–3 years)
Testing documentation
Physical condition
Market demand
During AI demand spikes, refurbished pricing has increased significantly.
Cost-Per-Performance Perspective: NVIDIA A100 80GB PCIe
When evaluating price, consider:
80GB HBM2e capacity
1.9 TB/s bandwidth
FP64 capability (9.7 / 19.5 TFLOPS)
MIG partitioning
For many AI and HPC deployments, the A100 still delivers strong price-to-performance value, especially in refurbished cluster builds.
Refurbished NVIDIA A100 80GB PCIe – What to Check
If buying refurbished:
Request stress test documentation
Verify firmware and driver compatibility
Confirm warranty terms
Inspect for data center extraction damage
Silicon performance is identical between new and properly tested refurbished units.
FAQ – NVIDIA A100 80GB PCIe
What is the FP64 performance of NVIDIA A100 80GB PCIe?
It delivers 9.7 TFLOPS FP64 via CUDA cores and up to 19.5 TFLOPS using Tensor Cores.
What is the memory bandwidth of the NVIDIA A100 80GB PCIe?
The GPU provides 1,935 GB/s memory bandwidth with 80GB HBM2e memory.
How much does NVIDIA A100 80GB PCIe cost?
Typically €9,000–€25,000 per unit depending on condition (new vs refurbished), warranty, and market demand.
Does NVIDIA A100 80GB PCIe support MIG?
Yes. It supports up to seven isolated GPU instances via Multi-Instance GPU (MIG).
What is the power consumption?
The NVIDIA A100 80GB PCIe has a 300W TDP and requires high-airflow server cooling.
What is the difference between NVIDIA A100 PCIe and SXM?
The PCIe version runs at 300W over PCIe Gen4. The SXM version supports NVLink, higher power limits, and slightly higher memory bandwidth (~2,039 GB/s).
Technical Summary - NVIDIA A100 80GB PCIe
The NVIDIA A100 80GB PCIe provides:
80GB HBM2e memory
1,935 GB/s bandwidth
9.7 TFLOPS FP64 (CUDA)
19.5 TFLOPS FP64 (Tensor Core)
156 TFLOPS TF32
MIG partitioning
300W TDP
For PCIe-based enterprise AI and HPC infrastructure, it remains one of the most widely deployed Ampere-generation GPUs.
NVIDIA A100 80GB PCIe GPUs
✔️ 5-Year Warranty – No Risk: Pay Only After Testing
Sources - NVIDIA A100 80GB PCIe
Official specifications: https://www.nvidia.com/en-us/data-center/a100/
Ampere architecture whitepaper: https://resources.nvidia.com/en-us-ampere-architecture/nvidia-ampere-ga100-architecture-whitepaper
NVIDIA A100 PCIe product brief (PDF): https://resources.nvidia.com/en-us-data-center-overview/a100-datasheet
Tensor Core and MIG documentation: https://docs.nvidia.com/datacenter/tesla/mig-user-guide/
CUDA performance and precision formats: https://docs.nvidia.com/cuda/