MangoBoost Outperforms All Other Ethernet-Based Results With Its Industry-First Data Processing Unit (DPU) Results in MLPerf Storage v1.0

MangoBoost Outperforms All Other Ethernet-Based Results With Its Industry-First Data Processing Unit (DPU) Results in MLPerf Storage v1.0

MangoBoost is proud to announce that it has set a new industry standard by becoming the first company to submit Data Processing Unit (DPU) accelerated storage systems in the MLPerf Storage benchmark. In both the v1.0 and v0.5 MLPerf submissions, MangoBoost is the only company to fully accelerate communications between AI servers (“Initiators”) and remote disaggregated storage servers (“Targets”) using DPU hardware solutions over standard Ethernet with the NVMe/TCP protocol.

Unprecedented Performance in Ethernet-Based Systems

MangoBoost has achieved the best performance in the Ethernet-based system category for MLPerf Storage v1.0. In the single-host system category, MangoBoost's solution demonstrated a remarkable 1.2x to 4.7x speedup, with an average of 2.7x, outperforming 6 other single-host systems submitted by Nutanix and Hammerspace. For the multi-host system category, when scaled to the equivalent number of hosts, MangoBoost's solution provided estimated speedups ranging from 1.4x to 7.8x, with an average of 4.3x, out of 16 total multi-host systems submitted by Lightbits Labs, Nutanix, Hammerspace.

These results underscore MangoBoost's leadership in delivering top-tier performance in MLPerf Storage, reaffirming the company's commitment to pushing the boundaries of AI and data processing technology.

Revolutionizing NVMe/TCP with DPU Acceleration

MangoBoost’s DPU solutions set a new benchmark by accelerating the entire NVMe/TCP protocol stack, including TCP packet processing, using specialized hardware. This approach bypasses the limitations of general-purpose CPUs, whether on the host or the DPU card, resulting in unmatched performance and efficiency.

Showcased Solutions:

Overcoming Traditional NVMe/TCP Challenges

Traditionally, NVMe/TCP has been favored for its scalability and ease of deployment over alternatives like RDMA/RoCEv2, which require specialized hardware. However, NVMe/TCP has often suffered from performance overheads compared to RDMA/RoCEv2 solutions. MangoBoost's fully hardware-accelerated NVMe/TCP solution is the first to overcome these limitations, delivering the high performance and efficiency required for cutting-edge AI and data processing applications.

Ease of Deployment and Compatibility

MangoBoost's DPU solutions are designed for straightforward deployment on standard systems. The hardware can be installed in a standard PCIe slot and connected via Ethernet QSFP ports, similar to a NIC. The software is compatible with standard NVMe/PCIe and NVMe/TCP protocols, ensuring seamless integration with existing systems without the need for modifications.

To demonstrate the robustness and versatility of its solutions, MangoBoost submitted MLPerf Storage results using DPUs on two different Supermicro servers, featuring CPUs from both major vendors and SSDs from two major vendors, across all GPU types and models in MLPerf Storage v1.0.

MangoBoost is gearing up to showcase its latest MLPerf-based products at two major events this fall: the Open Compute Project Global Summit in San Jose in October and the Supercomputing Conference in Atlanta in November. These events, the largest of their kind in the global open IT hardware and computer architecture sectors, provide key opportunities for MangoBoost to highlight its cutting-edge technologies.