GPU Compute
NVIDIA A100 Tensor Core GPU
Cost Effective Resources & Infrastructure
Reduce your cloud costs without sacrificing performance or reliability. mCloud delivers enterprise grade cloud infrastructure at a fraction of the cost of AWS, Google Cloud, or Azure.
Fault-Tolerant Tier IV Data Centre
Micron21 operates Australia’s first Tier IV-certified data centre, offering 100% uptime, redundant power, and high availability architecture.
24/7 Australian-based Expert Support
Our cloud specialists provide 24/7 Australian-based support, ensuring seamless deployments and efficient troubleshooting.
NVIDIA A100 (40 GB)
High-Performance Computing
Minimum Specifications
Provided as a High-Availability mCloud Virtual Cloud Server
- GPU: NVIDIA A100 (40 GB)
- GPU Compute: Dedicated
- vCPU: 12 Cores - XEON Gold
- RAM: 64 GB - DDR4
- Storage: 500 GB - NVMe SSD
- Bandwidth: 2 TB p/m
- IP Address: Included
- DDoS Protection: Shield
Overview
Enterprise Acceleration, Right-Sized
The A100 40GB pairs NVIDIA's Ampere architecture and third-generation Tensor Cores with 40GB of high-bandwidth HBM2 memory. It supports the full range of math precisions, from FP64 for HPC to INT8 for inference, making it a single accelerator that adapts to almost any data-centre workload, while Multi-Instance GPU lets one card serve up to seven isolated jobs at once.
Capabilities
What Makes the A100 40GB Different
Third-Gen Tensor Cores
Up to 312 TFLOPS of deep-learning performance and 20× the Tensor throughput of the previous Volta generation for training and inference.
Multi-Instance GPU (MIG)
Partition a single card into as many as seven fully isolated 5GB instances, each with its own memory, cache, and compute: ideal for multi-tenant serving.
Structural Sparsity
Tensor Cores exploit sparsity in AI models to deliver up to 2× higher performance, most notably for inference but also during training.
Next-Gen NVLink
Connect two GPUs over an NVLink bridge at 600 GB/s, double the previous generation's throughput, for workloads that outgrow a single card.
40GB HBM2 Memory
1,555 GB/s of memory bandwidth keeps the Tensor Cores fed, with enough capacity for production inference, fine-tuning, and mid-sized training runs.
Every Math Precision
One accelerator for every job: FP64 and TF32 for HPC and training, BF16/FP16 for deep learning, and INT8 for high-throughput inference.
Specifications
Technical Specifications
Compute & Tensor Cores
- FP649.7 TFLOPS
- FP64 Tensor Core19.5 TFLOPS
- FP3219.5 TFLOPS
- TF32 Tensor Core156 TFLOPS 312 TFLOPS with sparsity
- BFLOAT16 Tensor Core312 TFLOPS 624 TFLOPS with sparsity
- FP16 Tensor Core312 TFLOPS 624 TFLOPS with sparsity
- INT8 Tensor Core624 TOPS 1,248 TOPS with sparsity
Memory, Platform & Form Factor
- GPU memory40GB HBM2
- Memory bandwidth1,555 GB/s
- Max thermal design power250W
- Multi-Instance GPUUp to 7 @ 5GB
- InterconnectNVLink 600 GB/s PCIe Gen4 64 GB/s
- Form factorPCIe
- GPU architectureNVIDIA Ampere
Specifications per the NVIDIA A100 Tensor Core GPU datasheet (r4). Peak rates marked “with sparsity” require structural-sparsity-enabled models.
Performance
Built to Accelerate
Each figure compares the A100 against a different reference point, as published by NVIDIA. Note the baseline beneath each number.
Higher performance
vs the prior NVIDIA Volta generation, across AI training and inference.
AI inference throughput
BERT-Large inference vs a CPU-only server (INT8 with sparsity).
HPC throughput
Across top HPC apps vs P100, a four-year generational gain.
Sparse-model speed-up
From structural sparsity in Tensor Cores, primarily for inference.
Where It Fits
The Right Card for the Job
Ideal workloads
What the A100 40GB runs best
- Production AI inference: high-throughput serving with INT8 and structural sparsity.
- Fine-tuning & mid-sized training: accelerate models that fit comfortably within 40GB.
- HPC & scientific computing: full FP64 and TF32 precision for simulation and research.
- Data analytics: speed up large ETL, SQL, and machine-learning pipelines.
- Multi-tenant serving: partition into seven MIG instances to maximise utilisation.
Why run it on mCloud
Australian, on-demand, supported
- Tier IV data centre: Australia's first, with redundant power and high availability.
- Fast underlying platform: NVMe storage and 100Gbps networking feed the GPU.
- 24/7 Australian support: local cloud specialists for deployment and troubleshooting.
- Pay for what you use: scale instances up or down to match demand.
- OpenStack & IaC ready: automate provisioning with the API and Terraform.
Deploy the A100 40GB today
Spin up GPU compute on our Tier IV Australian cloud, or talk to our specialists about sizing the right configuration for your workload.
