Engineered for Next-Generation AI and HPC
The NVIDIA HGX™ B300 platform is built for the next wave of accelerated computing. Powered by Blackwell Ultra GPUs and high-speed NVLink interconnects, HGX B300 enables massive generative AI models, advanced data analytics, and high-performance simulations at scale. With a fully integrated software stack and support for up to 800 Gb/s networking, it delivers the infrastructure backbone needed to meet growing compute demands across enterprise and cloud environments.
AMAX Platforms with HGX B300
The AceleMax® AXG-828U is AMAX’s 8U platform engineered to support the full capabilities of the NVIDIA HGX B300. Designed for demanding AI training, inference, and simulation workloads, it pairs the performance of 8x Blackwell Ultra GPUs with robust system infrastructure built around dual AMD EPYC™ 9005 Series Processors or dual Intel® Xeon® 6 processors. This platform is optimized for high-throughput, large-memory applications in enterprise and industrial environments.
The AceleMax® AXG-928U is a 9U server platform designed for high-power GPU deployments and large-scale AI infrastructure. Equipped with 8x NVIDIA Blackwell Ultra GPUs and dual Intel® Xeon® 6 processors, the AXG-928U offers a modular design optimized for thermal performance, reduced cable complexity, and streamlined assembly.
With built-in 800G networking and support for up to 4TB of DDR5 memory, this system is built for enterprises, cloud providers, and research institutions running high-throughput LLM training, inference, and scientific workloads. The 5+5 redundant 3200W Titanium power supplies and high-efficiency air cooling architecture ensure stable operation even under the most demanding AI workloads.
Inference Performance for Large Language Models
Up to 11x Higher Inference Throughput on Llama 3.1 405B
The NVIDIA HGX B300 delivers significant improvements in real-time inference, reaching up to 11 times the throughput of HGX H100 systems for models such as Llama 3.1 405B. This gain is achieved using NVIDIA’s second-generation Transformer Engine with custom Blackwell Tensor Cores and optimizations from TensorRT-LLM.
This performance is based on per-GPU comparisons in an 8-GPU HGX configuration, using token-to-token latency of 20ms and first token latency of 5 seconds. Inference was served through disaggregated architecture with sequence lengths of 32,768 tokens (input) and 1,028 tokens (output).
Accelerated Training at Scale
4x Faster Training for Llama 3.1 405B
HGX B300 systems enable up to 4x faster training for large-scale models compared to the previous generation. This is powered by Blackwell’s second-generation Transformer Engine with FP8 precision, improved numerical formats, and high interconnect bandwidth.
With 1.8 TB/s of GPU-to-GPU bandwidth via fifth-generation NVLink, combined with InfiniBand networking and NVIDIA Magnum IO software, the HGX B300 platform supports efficient model scaling across GPU clusters for enterprise training workloads.
NVIDIA HGX B300 vs. NVIDIA HGX B200
HGX B300 is powered by Blackwell Ultra GPUs, offering higher FP4 throughput, significantly greater INT8 performance, and larger memory capacity compared to HGX B200. It also doubles the networking bandwidth (1.6 TB/s vs. 0.8 TB/s) and delivers 2x attention performance, making it better suited for large-scale inference and transformer-based models. HGX B200, while based on standard Blackwell GPUs, features higher FP64 compute, which may benefit mixed workloads requiring double-precision performance.
| Specification | HGX B300 | HGX B200 |
|---|---|---|
| Form Factor | 8x NVIDIA Blackwell Ultra SXM | 8x NVIDIA Blackwell SXM |
| FP4 Tensor Core** | 144 PFLOPS | 105 PFLOPS | 144 PFLOPS | 72 PFLOPS |
| FP8/FP6 Tensor Core* | 72 PFLOPS | 72 PFLOPS |
| INT8 Tensor Core* | 2 POPS | 72 POPS |
| FP16/BF16 Tensor Core* | 36 PFLOPS | 36 PFLOPS |
| TF32 Tensor Core* | 18 PFLOPS | 18 PFLOPS |
| FP32 | 600 TFLOPS | 600 TFLOPS |
| FP64/FP64 Tensor Core | 10 TFLOPS | 296 TFLOPS |
| Total Memory | Up to 2.3 TB | 1.4 TB |
| NVLink | Fifth generation | Fifth generation |
| NVIDIA NVSwitch™ | NVLink 5 Switch | NVLink 5 Switch |
| NVSwitch GPU-to-GPU Bandwidth | 1.8 TB/s | 1.8 TB/s |
| Total NVLink Bandwidth | 14.4 TB/s | 14.4 TB/s |
| Networking Bandwidth | 1.6 TB/s | 0.8 TB/s |
| Attention Performance | 2X | 1X |
* With sparsity
** With sparsity | without sparsity
Scalable Deployment with AMAX
AMAX offers complete rack scale deployment for NVIDIA HGX B300 platforms, starting with the AceleMax® AXG-828U. This 8U system is purpose-built for AI training and inference at scale, combining the power of 8x Blackwell Ultra GPUs with enterprise-grade CPU, memory, and storage configurations. Every system is integrated, validated, and optimized in-house by AMAX engineers to ensure high performance and efficiency from day one. Whether you're building infrastructure for generative AI or complex simulation workloads, AMAX delivers ready-to-deploy solutions backed by expert support.