ALKONTEK
building for digital abundance

Accelerated Mathematics

ALKONTEK builds both open-source and commercial modules for the Vecmat math and linear library that provide SIMD capability abstraction for AMD, Intel, and ARM platforms. Our modules and tools speed up prototyping math and physics focusing on edge embedded platforms, SBCs, and servers.

There are many platforms and CPU and GPU architectures available, but platform integration and compatibility can be challenging. Our modules and tools simplify the process by providing a unified interface for SIMD operations across different platforms, allowing developers to focus on their core algorithms rather than dealing with platform-specific intricacies.

Vecmat is self-contained and modular, allowing vendors to integrate with their hardware in a public or private way. It includes a custom testing and benchmarking framework that enables vendors to test their hardware and optimize their implementations for maximum performance.

Supported Hardware and Platforms

The core SIMD Vecmat contract is the following dispatch set:

  • SVE2
  • SVE
  • NEON
  • AVX-512F
  • AVX2
  • AVX
  • Scalar

However, different processors have different vector and matrix instructions and acceleration peripheries, and most libraries do not support or integrate all of them. There is also the scenario you're building for one platform, but for whatever reason decide to use a different platform. Our goal is to reduce the change of a platform to the bear minimum — that often translates into huge savings in time and money.

We provide special vendor implementations for these platforms:

Upcoming CPU Modules and Features

  • arm/cortex-x92x (SVE and SVE2)
  • intel/xeon-6 (AMX, AVX10.1)
  • intel/xeon-7 (AMX, AVX10.2)
  • amd/epyc-zen-4 (AVX-512, 256-bit pipeline)
  • amd/epyc-zen-5 (AVX-512, VNNI/BF16, 512-bit pipeline)
  • riscv/rva22 (RVV 1.0; optional)
  • riscv/rva23 (RVV 1.0; mandatory)

Upcoming GPU Modules and Features

AMD GPUs

  • amd/mi1xx (CDNA1)
  • amd/mi2xx (CDNA2)
  • amd/mi30x (CDNA3)
  • amd/mi32x (CDNA3)
  • amd/mi35x (CDNA4)

NVIDIA GPUs

  • nvidia/ampere-sm80 (3rd Gen WMMA + mma.sync)
  • nvidia/ada-l4x-sm89 (4th Gen WMMA + mma.sync)
  • nvidia/hopper-sm90 (4th Gen WMMA + mma + wgmma)
  • nvidia/blackwell-sm10 (5th Gen tcgen05.mma)

Intel GPUs

  • intel/alchemist-a3xx-a5xx (Xe-HPG, XMX/DPAS)
  • intel/alchemist-a7xx (Xe-HPG, XMX/DPAS)
  • intel/battlemage-b5xx (Xe2-HPG, XMX/Enhanced)