Skip to main content

GPU Product

Rent NVIDIA A100 80GB GPUs

80 GB of HBM2e at 2,039 GB/s, up to seven hardware-isolated MIG instances, and NVLink 3.0 for multi-GPU scale-out. The proven workhorse for large-model training, high-throughput inference, and HPC — from $1.05/hour.

Technical Specifications

ArchitectureNVIDIA Ampere
Memory Size80 GB HBM2e
Memory Bandwidth2 039 GB/s
CUDA Cores6 912
Tensor Cores432 (3rd gen)
NVIDIA A100 80GB datacenter GPUs
Rental Options

A100 Rental Options

Rent NVIDIA A100 80GB SXM4 GPUs from $1.05/hour. Each GPU ships with 15 CPU cores, 210 GB of RAM, and 800 GB of NVMe, and nodes scale to 8 GPUs over NVLink 3.0. Hourly billing with no minimum commitment, plus 1-month and 3-month reservations for committed workloads.

Start rental in consoleContact Us

Comparison

A100 80GB vs V100 32GB

The A100 is the direct generational successor to the V100, and both live in the same CloudRift datacenter — so this is the upgrade decision most tenants actually face. For roughly 3.6× the hourly rate you get 2.5× the memory, 2.3× the bandwidth, and 2.5× the Tensor throughput, plus TF32, BF16, MIG partitioning, and NVLink 3.0. The V100 remains the better value when your model fits in 32 GB; the A100 earns its premium the moment it does not.

Comparison dimensionA100 80GBV100 32GB% Diff
ArchitectureAmpere (GA100)Volta (GV100)N/A
Memory TypeHBM2e ECCHBM2 ECCN/A
VRAM80 GB32 GB+150%
Bus Width5 120-bit4 096-bit+25%
Memory Bandwidth2 039 GB/s900 GB/s+126.6%
FP64 Performance9.7 TFLOPS~7.8 TFLOPS+24.4%
FP64 Tensor Core19.5 TFLOPSNot supportedN/A
FP32 Performance19.5 TFLOPS~15.7 TFLOPS+24.2%
TF32 Tensor Core156 TFLOPSNot supportedN/A
BF16/FP16 Tensor312 TFLOPS~125 TFLOPS+149.6%
CUDA Cores6 9125 120+35%
Tensor Cores432 (3rd gen)640 (1st gen)−32.5%
Multi-Instance GPUUp to 7 MIGNot supportedN/A
Multi-GPU InterconnectNVLink 3.0NVLink 2.0N/A
Form FactorSXM4SXM3N/A
TDP400 W350 W+14.3%

Performance

Key performance metrics

HBM2e Memory Bandwidth

2 039 GB/s of HBM2e bandwidth across an 5 120-bit bus — more than double the V100 and enough to keep 70B-class models fed during high-batch inference and training.

Third-Generation Tensor Cores

432 third-generation Tensor Cores deliver 312 TFLOPS of BF16/FP16 and 156 TFLOPS of TF32 — TF32 accelerates existing FP32 training code with no changes to your model.

Multi-Instance GPU (MIG)

Partition a single A100 into up to seven hardware-isolated instances, each with dedicated memory, cache, and compute — run separate tenants or jobs without noisy-neighbour interference.

Use Cases

What the A100 Is Built For

Large-Model Training & Fine-Tuning

80 GB of HBM2e holds 30B–70B parameter models for LoRA and QLoRA runs, or full fine-tunes of 13B-class models — with NVLink 3.0 for multi-GPU scale-out.

High-Throughput LLM Inference

Token generation is memory-bandwidth bound. At 2 039 GB/s the A100 sustains high concurrency on vLLM and TensorRT-LLM without spilling to host memory.

Multi-Tenant Serving

MIG carves one A100 into up to seven isolated instances — serve several models or customers from a single card with guaranteed quality of service.

HPC & Scientific Computing

9.7 TFLOPS of native FP64 — and 19.5 TFLOPS through the FP64 Tensor Cores — for CFD, molecular dynamics, genomics, and computational chemistry.

NVIDIA A100 FAQ

Common Questions About the A100

A100 80GB rentals on CloudRift start at $1.05/hour for a single GPU. Pay-as-you-go, no minimum commitment, and nodes scale up to 8 GPUs.
The A100 is an Ampere-architecture datacenter GPU used for large-model training and fine-tuning, high-throughput LLM inference, multi-tenant model serving via MIG, and HPC workloads such as CFD, molecular dynamics, and genomics. Its 80 GB of HBM2e and 2,039 GB/s of bandwidth make it a strong fit for memory-bound and bandwidth-bound work.
The A100 80GB has 80 GB of HBM2e memory on a 5,120-bit bus delivering 2,039 GB/s of bandwidth — more than twice the bandwidth of the V100 32GB and roughly 2.4× that of a GDDR6 card like the L40.
MIG lets you partition a single A100 into as many as seven hardware-isolated GPU instances, each with its own memory, cache, and streaming multiprocessors. Because the isolation is enforced in hardware rather than by the scheduler, one tenant cannot affect another’s throughput — which makes it well suited to multi-tenant inference and shared research clusters.
The A100 80GB has 2.5× the VRAM (80 vs 32 GB), 2.3× the memory bandwidth (2,039 vs 900 GB/s), roughly 2.5× the Tensor throughput (312 vs ~125 TFLOPS), and adds TF32, BF16, MIG, and NVLink 3.0. The V100 costs about a third as much per hour at $0.29 vs $1.05. Choose the V100 for smaller memory-bound jobs on a tight budget, and the A100 when the model does not fit in 32 GB or when memory bandwidth is your bottleneck.
An A100 80GB on CloudRift is $1.05/hr. The same silicon on AWS p4de.24xlarge runs roughly $5.12/hr per GPU and on Azure NDm A100 v4 roughly $4.10/hr per GPU, so CloudRift is about 4–5× cheaper for equivalent hardware.
A single A100 80GB comfortably serves models up to about 70B parameters in 4-bit quantization, or roughly 34B in FP16. For full-precision fine-tuning it handles 13B-class models on one card, and multi-GPU nodes with NVLink 3.0 scale to larger runs.
Yes. Open the Console, pick an A100 instance, and launch — most instances are running in under a minute. Hourly billing, no long-term commitment.
Get in touch

Ready to get started?

Get in touch with our team to discuss your requirements and find the right solution for your infrastructure.

Contact Us