Lab Resources
Computational Resources
IIT Gandhinagar has a number of computational resources available for research and teaching purposes. In addition, the department also has a number of workstations available for students and faculty members.
The following is the list of computational resources in our lab apart from the institute and departmental resources.
Servers
Our servers are named after great mathematicians/scientists, plus Sustain after the lab’s focus on sustainability:
- Ramanujan (रामानुजन): Srinivasa Ramanujan made groundbreaking contributions to number theory.
- Bhaskar (भास्कर): Bhaskar (aka Bhaskar II) is known for his work Siddhānta Shiromani (सिद्धांत शिरोमणि) which includes advances in algebra, calculus, and astronomy.
- Sustain: named after the lab’s focus on sustainability and energy research.
| Ramanujan | Bhaskar | Sustain | |
|---|---|---|---|
| RAM | 512 GB (16 x 32GB) | 256 GB (4 x 64GB) | 192 GB (6 x 32GB) |
| CPU | AMD EPYC 7452 32-Core Processor @ 2.30 GHz | Intel Gold ICX 6326 @ 2.90 GHz | Intel(R) Xeon(R) Silver 4208 CPU @ 2.10GHz |
| Storage | 8 TB | 5 TB | 2 TB |
| Number of CPUs | 64 | 32 | 16 |
| GPU | 4 x NVIDIA A100-SXM4 (80GB) | 2 x NVIDIA RTX A5000 (24GB) | 2 x NVIDIA RTX A4000 (16GB) |
| Total VRAM | 320 GB | 48 GB | 32 GB |
Compact AI System
Our NVIDIA DGX Spark is named Mani (मणि), after Anna Mani, the pioneering Indian physicist and meteorologist whose work on solar-radiation and wind-energy measurement closely reflects the lab’s sustainability focus. The DGX Spark is a compact Arm64 AI system-on-chip rather than a conventional x86 server or discrete-GPU workstation, so we list it separately.
Watch the four-minute India Science profile, Anna Mani: The Pioneer Indian Meteorologist.
| Mani | |
|---|---|
| System | NVIDIA DGX Spark |
| Processor | NVIDIA GB10 Grace Blackwell Superchip |
| CPU | 20-core Arm (10 x Cortex-X925 + 10 x Cortex-A725) |
| GPU | Integrated NVIDIA Blackwell GPU (6,144 CUDA cores) |
| Unified memory | 128 GB LPDDR5x shared coherently by CPU and GPU |
| Memory bandwidth | 273 GB/s |
| Storage | 4 TB NVMe M.2 |
| AI tensor performance | Up to 1 PFLOP at FP4 with sparsity (theoretical peak) |
| Power | 140 W GB10 TDP; 240 W power supply |
Memory and performance note. Mani’s 128 GB is coherent unified system memory shared by the CPU, integrated GPU, operating system, and applications; it is not 128 GB of dedicated VRAM. This gives a single GPU workload more addressable memory than any one GPU currently listed, which is especially useful for fitting large AI models, but less than 128 GB is available in practice. Ramanujan also retains more aggregate dedicated GPU memory (320 GB across four A100s). Memory capacity is not a speed rating: Mani’s integrated GPU, 273 GB/s LPDDR5x bandwidth, Arm64 software requirements, and compact power envelope make it materially different from the discrete x86/GPU systems. Hardware figures follow NVIDIA’s DGX Spark specifications.
Workstations
Our workstations are named after Indian scientists: Aryabhata (आर्यभट), well-known for the concept of zero (शून्य) and the place value system, and Vikram.
| Aryabhata | Vikram | |
|---|---|---|
| RAM | 32 GB | 64 GB |
| CPU | Intel Core i9-13900 | Intel Core i9-14700 |
| Storage | 2 TB | 4 TB |
| Number of CPUs | 24 | 32 |
| GPU | 1 x NVIDIA RTX Titan XP (12 GB) | 1 x NVIDIA RTX A2000 (12 GB) |
| Total VRAM | 12 GB | 12 GB |
Relative Performance
Measured with our open gpu-benchmark-suite using the same workload definitions on every machine, grouped into raw compute, training, inference, and data-loading. Mani’s results were measured on 31 August 2026 with the current GB10-compatible Arm64 stack (NVIDIA PyTorch 25.11 and CUDA 13.0); the earlier machines used NVIDIA PyTorch 24.03 and CUDA 12.4. Its results are therefore a useful workload-level comparison, but not a strict common-software-stack comparison. These workloads fit within 12 GB, so they measure speed rather than Mani’s distinctive ability to fit much larger models in its 128 GB unified-memory pool.
| Benchmark | Ramanujan | Bhaskar | Sustain | Aryabhata | Vikram | Mani |
|---|---|---|---|---|---|---|
| GPU | A100-SXM4-80GB | RTX A5000 | RTX A4000 | TITAN Xp | RTX A2000 | GB10 (integrated Blackwell) |
| Raw compute (TFLOPS, fp16) | 269.7 | 89.2 | 61.3 | 8.7 | 26.4 | 96.2 |
| Training — ResNet50 (img/s) | 1059.9 | 717.7 | 508.5 | 257.3 | 304.5 | 345.0 |
| Inference — ResNet50 (img/s) | 3139.4 | 1419.5 | 888.6 | 726.9 | 522.7 | 614.1 |
| Inference — detection (FPS) | 57.9 | 46.2 | 28.9 | 22.5 | 19.0 | 19.3 |
| Inference — LLM gen (tok/s) | 70.8 | 87.2 | 46.8 | — | 71.7 | 25.5 |
| Data loading (img/s) | 2483.8 | 2216.5 | 957.6 | 4805.0 | 4413.1 | 5297.3 |
Raw compute (TFLOPS) and training throughput track the GPUs’ real power most reliably — the A100 leads the GPU workloads overall, while Mani places second in measured FP16 matrix throughput. Small-batch inference (LLM tokens/sec, detection FPS) is latency-bound, so it can look flat on large GPUs and varies with concurrent load. Data loading primarily reflects the CPU, NVMe storage, and software pipeline rather than GPU performance. Numbers are point-in-time, single-GPU measurements; the TITAN Xp’s LLM generation is omitted as its older (Pascal) architecture is impractically slow at FP16 generation.
Server Policy
How to get access: Send an email to Dr. Supin Gopi, keeping Prof. Nipun Batra in cc. Mention the following details in your request: i) server name; ii) purpose for access.
Don’ts:
- Do not use the server for personal use or anything not related to the project. This includes any classwork or homework.
- Do not share your server credentials with anyone else.
Fair usage:
- Sometimes we have multiple projects running on the same server. Please be considerate of other projects and do not use up all the resources.
- Sometimes we may allocate specific cores or GPUs to specific projects. Please do not use cores or GPUs that are not allocated to your project.