Skip to main content

GPU Product

Rent NVIDIA V100 GPUs

16 GB or 32 GB of HBM2 at 900 GB/s — the cost-leader for memory-bound AI fine-tuning, HPC, and batch inference. From $0.25/hour on sustainable, second-life Volta silicon.

Technical Specifications

ArchitectureNVIDIA Volta
Memory Size16 or 32 GB HBM2
Memory Bandwidth900 GB/s
CUDA Cores5 120
Tensor Cores640 (1st gen)
NVIDIA V100 SXM datacenter GPU module
Rental Options

V100 Rental Options

Rent NVIDIA V100 GPUs from $0.25/hour — 16GB SXM2 or 32GB SXM3. Hourly billing with no minimum commitment, plus 1-month and 3-month reservations for committed workloads. Capacity sourced from sustainable second-life datacenter hardware.

Start rental in consoleContact Us

Comparison

V100 32GB vs L40S

Both are datacenter GPUs in the CloudRift fleet, positioned for different workloads. The table below uses the 32 GB V100; the 16 GB part is the same Volta silicon with half the memory, at $0.25/hr. The V100 wins on price (roughly half the hourly rate), HBM2 memory bandwidth, FP64 throughput, and NVLink for multi-GPU scale-out. The L40S wins on raw FP32, modern Ada Tensor Cores, and total VRAM — choose it when you need maximum inference throughput per card.

V100 32GBL40S% Diff
CloudRift Price$0.29 / hr$0.57 / hr reserved−49%
ArchitectureVoltaAda LovelaceN/A
Memory TypeHBM2 ECCGDDR6 ECCN/A
VRAM32 GB48 GB−33.3%
Bus Width4 096-bit384-bit+966%
Memory Bandwidth900 GB/s864 GB/s+4.2%
FP64 Performance~7.8 TFLOPS~1.4 TFLOPS+457%
FP32 Performance~15.7 TFLOPS~90.5 TFLOPS−82.7%
Tensor Performance~125 TFLOPS~362 TFLOPS−65.5%
CUDA Cores5 12018 176−71.8%
Tensor Cores640 (1st gen)568 (4th gen)N/A
Multi-GPU InterconnectNVLink 2.0PCIe 4.0 onlyN/A
Form FactorSXM3Dual-slot PCIeN/A
TDP350 W350 W0%

Performance

Key performance metrics

HBM2 Memory Bandwidth

900 GB/s of HBM2 bandwidth on a 4 096-bit bus — the cost-leader for memory-bound workloads like scientific compute, mid-sized LLM fine-tuning, and large-batch inference.

Tensor Core Acceleration

640 first-generation Tensor Cores deliver up to ~125 TFLOPS of mixed-precision throughput — proven silicon for CUDA, cuDNN, and mature ML toolchains.

A Fraction of Hyperscaler Pricing

$0.25/hr on CloudRift for the 16GB part and $0.29/hr for the 32GB, vs ~$3.90/hr per GPU on AWS p3dn.24xlarge and ~$2.75/hr per GPU on Azure NDv2 — up to 13× cheaper than hyperscalers for the same V100 silicon.

NVIDIA V100 FAQ

Common Questions About the V100

V100 rentals on CloudRift start at $0.25/hour for a 16GB SXM2 GPU, or $0.29/hour for a 32GB SXM3. Pay-as-you-go, no minimum commitment.
The V100 is a Volta-architecture datacenter GPU used for AI fine-tuning, batch inference, and HPC workloads like CFD, molecular dynamics, and computational chemistry. Its HBM2 memory — 16 GB or 32 GB depending on the configuration — and strong FP64 performance make it cost-efficient for memory-bound and double-precision workloads.
CloudRift runs two V100 configurations: 16 GB HBM2 on the SXM2 part and 32 GB HBM2 on the SXM3. Both sit on a 4 096-bit bus with 900 GB/s of bandwidth — meaningfully higher than GDDR-class cards in the same price tier.
Yes — for the right workload. The V100 has 640 first-generation Tensor Cores and HBM2 memory: the 16GB part suits 7B-class fine-tuning and batch inference, and the 32GB part stretches to 13B. Both are strong on memory-bandwidth-bound HPC. For training the largest frontier models you want H100/H200 or MI350X, but the V100 remains the cost-leader for many production workloads.
A V100 32GB on CloudRift is $0.29/hr, and the 16GB part is $0.25/hr. The same 32GB silicon on AWS p3dn.24xlarge runs ~$3.90/hr per GPU, and on Azure NDv2 ~$2.75/hr per GPU — CloudRift is roughly 10–13× cheaper. Capacity comes from sustainable second-life datacenter hardware, redeployed instead of replaced.
The V100 32GB is $0.29/hr on demand. The L40S is reserved-only, from $0.57/hr on a 1-month reservation, so the V100 is roughly half the price. It has HBM2 memory (900 vs 864 GB/s), a much wider 4 096-bit bus, NVLink 2.0 for multi-GPU, and about 5× the FP64 throughput (~7.8 vs ~1.4 TFLOPS). The L40S has 50% more VRAM (48 GB), newer Ada Tensor Cores, and roughly 6× the FP32 throughput, so it is the better pick when you need maximum inference throughput per card or modern Ada-only features. Pick the V100 for memory-bandwidth-bound workloads, FP64-heavy HPC, multi-GPU training, or the lowest hourly rate. Pick the V100 for memory-bandwidth-bound workloads, FP64-heavy HPC, multi-GPU training, or the lowest hourly rate.
Yes. Open the Console, pick a V100 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