According to a Gartner study, 47% of all companies want to re-patriate or re-home at least the crucial parts of their IT back onto on-premises data centers. This move away from the Cloud is due to many reasons, Data Sovereignty and security as well as TCO (operating costs based on excessive power consumption and enormous requirements for cooling) among the most important ones. While Cloud Service Providers can build their data centers where the power and cooling are cheap and abundant, on-premises operators do not have that luxury. Their focus is instruction- and energy-efficiency of the systems executing their crucial workloads.

Today’s CPUs and GPGPUs are constrained by the lack of seamless Inter Processor Communication (IPC) and by relying on traditional memory (DRAM and even HBM) with well-known bottlenecks such as the often-cited Memory Wall. Those shortcomings limit performance scaling for Artificial Intelligence (AI, i.e. training and inference as well as agentic AI), high-performance computing (HPC), and cloud workloads. At Abacus Semi, we’ve developed the Kloth Architecture™, a patented compute platform based on adding a scale-out port onto each CPU and accelerator core, and unified interfaces between processors, accelerators and smart multi-homed memories. That architecture delivers:

  • Seamless Integration of Processors and Accelerators with Smart Multi-Homed Memory Systems
  • Processors' and Accelerators' ability to communicate without hardware or software bridges
  • Efficient and coherent sharing of data across Processors and Accelerators
  • Lower TCO: reduced operating costs (power, cooling) for AI and HPC Server clusters
  • On-prem data centers and smart carrier edge solutions within power-constrained environments
  • Compatibility with most AI and HPC API frameworks

Abacus Semi is on a mission to accelerate the future of compute – with first silicon prototypes in development, dev kits for research institutions, and OEM partnerships underway.

The resulting advantage of this novel computing architecture is at least an order of magnitude improvement in performance, power consumption (and heat generation), or both.

Our patent-protected Heterogeneous Accelerated Compute™ technology removes the fundamental bottlenecks affecting inter-processor communication and memory access in existing systems, enabling server manufacturers to build supercomputers for AI and beyond with unprecedented scalability. In addition, Abacus Semi’s unique architecture provides intrinsic cybersecurity features that make these widely distributed servers impervious to even the most advanced cyberthreats.

There is also consistent agreement across the industry that a Hybrid Cloud approach has become essential, where new investment into on-premises AI data centers and high-performance computing including AI and proxying at the smart (carrier) edge is needed to power proprietary AI foundational models trained on private datasets and agentic AI for secure internal operations and product development.

Industry consensus is that we have reached the end of the usefulness of PCIe and of legacy memory interfaces such as DDR5, even if High Bandwidth Memory (HBM) stacks are now integrated into processors and GPGPUs. Several attempts have been made to delay the inevitable. NVidia has introduced NVLink to connect its GPGPUs directly, circumventing the existing GPGPU to GPGPU communication via CPU and NIC bottlenecks. AMD, Intel and others countered with UALink that tries to achieve the same. The memory size and shared access bottlenecks are unsuccessfully addressed by CXL. However, none of these technologies address or solve the fundamental underlying problem in connectivity. Abacus Semi is solving these challenges with the patent-pending Heterogeneous Accelerated Compute™ or Kloth Architecture for processors and accelerators.