ionstream Expands Bare-Metal Cloud With NVIDIA B200 For Extreme GPU Inferencing and Training Performance

ionstream, provider of enterprise GPU infrastructure and managed services, announces the availability of Nvidia B200 GPUs on its GPU-as-a-Service platform as bare metal servers. 

This strategic expansion will enable organizations to build and run real-time generative AI on trillion-parameter models on a reliable and affordable platform with multiple payment and term options.

ionstream’s B200 servers are now available for deployment at the market leading cost of $3.50/hour per GPU monthly commitment giving customers access to the next generation of compute power for the most demanding AI, ML and data analytics workloads.

The Nvidia B200 sets a new benchmark for AI performance, delivering unprecedented capabilities for training and inference. With up to 15X faster real-time inference and 3X faster training speeds compared to the Nvidia Hopper architecture, the B200 is engineered to handle the most demanding AI workloads.

The offering seeks to address a broken hyperscaler public cloud GPU market that often fails under real-world pressure. Common problems include users competing for access, latency spikes that break inference workflows, inflexible stacks, and unfair pricing.

ionstream customers will secure direct access to bare-metal Nvidia B200 GPUs with no bandwidth usage billing (multiple 10g dedicated unmetered connections per server), virtualization or shared resources.

All ionstream hardware currently runs in Tier 4 designed data centers with complete fault tolerance and multi-homed BGP network. This ensures continuous and smooth traffic flow 24/7, with 99.999% uptime – the highest SLA in the neocloud landscape at 5 9’s, setting a new bar in reliability.  

Technical highlights of the Nvidia HGX B200

• 8 x B200 GPUs.
• Up to 72 PFLOPS training and 144 PFLOPS inference. 
• Total 1.44TB GPU memory with 64 TB/s bandwidth. 
• Intel Xeon 6900P Series CPU 2.40/3.9 GHz (L3 Cache 480M) / AMD 9004/9005 
• Series CPU 2.40/3.7Ghz (L3 Cache 384M). 
• 1.5TB of system memory.  
• NVLink 5 switch featuring 144 NVlink ports with 14.4 TB/s of a full line-rate bidirectional bandwidth capability.
• 3.2Tb/s of non-blocking GPU-to-GPU traffic with RoCE support over Ultra-low latency Ethernet network.
• Up to 30% lower power draw per token compared to H100-based systems. 
• Seamless integration into multi-node clusters with RDMA / RoCE and GPUDirect support.
• Fully compatible with any framework and inference servers supporting CUDA 12.8(PyTorch, TensorFlow, Triton, VLLM, and SGLANG Inference Servers). 

One of ionstream’s early adopters, Stellenium has already seen the benefits first-hand.

“Our B200 has delivered exceptional network performance and uptime that’s been critical for our business operations. It’s rare to find a vendor that combines great technology with such genuine commitment to customer success.” said Slava Yanson, CEO, Stellenium.

“GPU as a Service is projected to reach 20 billion $s by 2027. The reason is clear. AI teams need more control, more speed, and more flexibility than legacy cloud models can deliver. What we are hearing from teams coming to us is that the market isn’t delivering what they need. Cloud GPU instances are overpriced, virtualized, and unreliable, but owning servers means locking capital into hardware instead of building product.” said Jeff Hinkle, CEO, ionstream “In response, we are delighted to announce our new bare metal cloud B200 offering, giving our start-up, SME, and enterprise customers all of the benefits of GPU ownership with none of the overheads, just focused, high-performance infrastructure.”