GPU Servers with NVIDIA Tesla V100
Introduced in 2017, the NVIDIA Volta architecture (Tesla V100) faces significant limitations for training or fine-tuning modern Large Language Models (LLMs). This is primarily due to its limited VRAM capacity (16 GB or 32 GB) and the lack of native hardware support for modern compute precision formats like TF32, BF16, and FP8.
Nevertheless, V100-based configurations remain highly viable for a specific set of workloads where budget optimization is critical, and demands for raw compute power and memory bandwidth are moderate:
- Small-Scale Model Inference: Deploying lightweight models that do not require massive amounts of VRAM.
- Legacy Environments: Running legacy codebases and specialized enterprise software certified specifically for the Volta architecture.
- Foundational HPC Workloads: Numerical methods and simulations (e.g., molecular dynamics) that parallelize efficiently but do not bottleneck on HBM2 memory bandwidth.
- Development and Debugging: Testing pipeline logic and writing code before scaling up workloads to higher-performance configurations (such as H100 or H200).
Hardware Specifications for Tesla V100 Servers
All NVIDIA Tesla V100 compute nodes at immers.cloud are built on robust, enterprise-grade architecture:
- Processors: Dual Intel® Xeon® Gold 6240R processors, featuring a 2.4 GHz base clock with dynamic boosting up to 4.0 GHz via Intel® Turbo Boost Technology.
- ML Optimization: Each processor features dual Intel® AVX-512 units and supports Intel® AVX-512 Deep Learning Boost. This instruction set significantly accelerates low-precision multiply-and-add operations, which are critical for the inner loops of deep learning algorithms.
- Memory: Up to 3,072 GB of 2933 MHz DDR4 ECC Registered RAM. Error-Correcting Code (ECC) support guarantees data integrity during extended computational sessions.
- Storage Subsystem: Up to 1,920 GB of local storage powered by enterprise-grade Intel® Data Center SSDs. This ensures the high I/O throughput and reliability required for handling massive datasets.
GPU Tesla V100
When NVIDIA introduced the Volta architecture and the V100 GPU in 2017, the market was highly skeptical. The generative AI boom had not yet begun, and many industry experts questioned the viability of dedicating specialized silicon (Tensor Cores) exclusively to matrix computations. At the time, many researchers believed that general-purpose CUDA cores were entirely sufficient, and that the inclusion of expensive HBM2 memory would make the final product too niche.
However, this strategic bet paid off handsomely. The V100 was the world’s first GPU to feature Tensor Cores, laying the foundational hardware blueprint for every subsequent generation of AI accelerators, from Turing to Blackwell. The architectural refinements pioneered in the V100 directly informed the design of more powerful neural network accelerators. It definitively proved that hardware specialization for linear algebra delivers multiplicative performance gains over general-purpose computing.
The Tesla V100 is a legendary architecture that ushered in the modern era of hardware-accelerated deep learning. While it is not intended to tackle the bleeding-edge workloads designed for the latest generation of chips, it remains a highly reliable, predictable, and cost-effective solution for its target class of workloads.
| Video memory capacity |
32 GB |
| Type of video memory |
HBM2 (ECC) |
| Memory bandwidth |
900 Gb/s |
| Tensor cores |
640 |
| CUDA cores |
5120 |
GPU performance benchmarks
Performance benchmarks results in a virtual environment for 1 Tesla V100 graphics card.
-
OctaneBench 2020
pts
-
Matrix multiply example
GFlop/s
-
Hashcat bcrypt
H/s
100% Performance
Each physical core or GPU adapter is dedicated to a single client.
This means:
- 100% vCPU time is available;
- Physical pass-through of GPUs inside virtual servers;
- Reduced storage and network load on hypervisors, delivering more storage and network performance to clients.
Up to 75,000 IOPS1 for RANDOM READ and up to 20,000 IOPS for RANDOM WRITE on Virtual Servers with local SSDs
Up to 70 000 IOPS1 for RANDOM READ and up to 60 000 IOPS for RANDOM WRITE on Virtual Servers with block storage volumes
Advanced immersion cooling technology maintains optimal processor temperatures, ensuring stability and maximum turbo boost frequencies across all CPU and GPU cores.
Ready-made OS images with the required software
Create virtual servers by utilizing our pre-configured OS images with either Windows or Linux, along with specialized pre-installed software.
View all the pre-installed images
in the Marketplace.
Pure OpenStack API
Developers and system administrators can manage the cloud using the full OpenStack API.
Authenticate ninja_user example:
$ curl -g -i -X POST https://api.immers.cloud:5000/v3/auth/tokens \
-H "Accept: application/json" \
-H "Content-Type: application/json" \
-H "User-Agent: YOUR-USER-AGENT" \
-d '{"auth": {"identity": {"methods": ["password"], "password": {"user": { "name": "ninja_user", "password": "ninja_password", "domain": {"id": "default"}}}}, "scope": {"project": {"name": "ninja_user", "domain": {"id": "default"}}}}}'
Create ninja_vm example:
$ curl -g -i -X POST https://api.immers.cloud:8774/v2.1/servers \
-H "Accept: application/json" \
-H "Content-Type: application/json" \
-H "User-Agent: YOUR-USER-AGENT" \
-H "X-Auth-Token: YOUR-API-TOKEN" \
-d '{"server": {"name": "ninja_vm", "imageRef": "8b85e210-d2c8-490a-a0ba-dc17183c0223", "key_name": "mykey01", "flavorRef": "8f9a148d-b258-42f7-bcc2-32581d86e1f1", "max_count": 1, "min_count": 1, "networks": [{"uuid": "cc5f6f4a-2c44-44a4-af9a-f8534e34d2b7"}]}}'
STOP ninja_vm example:
$ curl -g -i -X POST https://api.immers.cloud:8774/v2.1/servers/{server_id}/action \
-H "Accept: application/json" \
-H "Content-Type: application/json" \
-H "User-Agent: YOUR-USER-AGENT" \
-H "X-Auth-Token: YOUR-API-TOKEN" \
-d '{"os-stop" : null}'
START ninja_vm example:
$ curl -g -i -X POST https://api.immers.cloud:8774/v2.1/servers/{server_id}/action \
-H "Accept: application/json" \
-H "Content-Type: application/json" \
-H "User-Agent: YOUR-USER-AGENT" \
-H "X-Auth-Token: YOUR-API-TOKEN" \
-d '{"os-start" : null}'
SHELVE ninja_vm example:
$ curl -g -i -X POST https://api.immers.cloud:8774/v2.1/servers/{server_id}/action \
-H "Accept: application/json" \
-H "Content-Type: application/json" \
-H "User-Agent: YOUR-USER-AGENT" \
-H "X-Auth-Token: YOUR-API-TOKEN" \
-d '{"shelve" : null}'
DELETE ninja_vm example:
$ curl -g -i -X DELETE https://api.immers.cloud:8774/v2.1/servers/{server_id} \
-H "User-Agent: YOUR-USER-AGENT" \
-H "X-Auth-Token: YOUR-API-TOKEN"
Documentation
Subscribe to our newsletter
Get notifications about new promotions and special offers by email.