· GPU Compute Catalogue

World-Class
GPU Compute.
Locally Deployed.

On-demand access to NVIDIA's latest data centre GPUs — from proven H100 workhorses to flagship Blackwell B200 clusters. Deployed in Malaysia for data sovereignty, low-latency regional access, and full compliance with local regulations.

Malaysia-Hosted On-Demand & Reserved NVLink Clusters PDPA Compliant 24/7 SRE Support
— Available Hardware
NVIDIA B200 GPU
B200
Blackwell Architecture · Flagship
B200
Ultra-Scale AI Compute · Dual-GPU Module

NVIDIA's most powerful data centre GPU. The B200 integrates two Blackwell dies in a single module, delivering unprecedented memory bandwidth and compute density for frontier model training and ultra-large inference workloads.

Architecture Blackwell
GPU Memory 192 GB HBM3e
Memory BW 8 TB/s aggregate
Interconnect NVLink 5.0 · 1.8 TB/s
Precision FP4 / FP8 / BF16 Tensor Cores
vs. H100 ~5× AI training throughput
Frontier LLM Training Mixture-of-Experts Ultra-Scale Inference Multimodal Foundation Models High-Speed HPC Simulation
NVIDIA B100 GPU
B100
Blackwell Architecture
B100
Next-Gen Training & Inference

The Blackwell B100 delivers a decisive leap in compute efficiency over Hopper, with native FP4 tensor cores and double the memory bandwidth of the H100 — purpose-built for large-scale AI training and high-throughput inferencing.

Architecture Blackwell
GPU Memory 192 GB HBM3e
Memory BW 8 TB/s
Interconnect NVLink 5.0
Precision FP4 / FP8 / BF16
vs. H100 ~3× AI training throughput
LLM Pre-Training Fine-Tuning at Scale High-Throughput Inference RAG & Retrieval Systems
NVIDIA H200 GPU
H200
Hopper Architecture · HBM3e
H200
Memory-Optimised AI Powerhouse

The H200 upgrades NVIDIA's battle-proven Hopper architecture with HBM3e memory — delivering 4.8 TB/s of bandwidth and 141 GB of capacity per GPU. The optimal choice for large-model inference where memory capacity is the binding constraint.

Architecture Hopper (GH200)
GPU Memory 141 GB HBM3e
Memory BW 4.8 TB/s
Interconnect NVLink 4.0 · 900 GB/s
Precision FP8 / BF16 / TF32
vs. H100 1.76× memory bandwidth
Large Model Inference 70B+ Parameter Models Multi-GPU Training Memory-Bound Workloads
NVIDIA H100 GPU
H100
Hopper Architecture · Industry Standard
H100
The AI Infrastructure Workhorse

The NVIDIA H100 SXM5 remains the industry standard for production AI — trusted by global cloud providers and research labs alike. Offers a proven, well-supported platform for training, fine-tuning, and deploying transformer-based models at scale.

Architecture Hopper (SXM5)
GPU Memory 80 GB HBM2e
Memory BW 3.35 TB/s
Interconnect NVLink 4.0 · 900 GB/s
Precision FP8 / BF16 / TF32
Cluster Size Up to 8-way NVLink
Model Fine-Tuning Production Inference 7B–70B Training Runs Computer Vision Recommendation Systems
— Infrastructure & Performance
NVLink cluster topology
Interconnect
NVLink Cluster Topology

All GPU nodes are interconnected via NVIDIA NVLink, enabling direct GPU-to-GPU communication at up to 1.8 TB/s — eliminating PCIe bottlenecks for multi-GPU training runs.

GPU performance comparison
Benchmark
Relative AI Training Throughput

Measured across LLM pre-training workloads (FP8/BF16). B200 leads at ~5× H100 baseline. Select the right tier for your budget and timeline.

· Request Compute

Reserve Your
GPU Cluster.

On-demand or reserved capacity, single-node to full-rack NVLink clusters. Our team will size a solution matched to your training or inference requirements.

Get in Touch