Resources and Software
We offer a variety of resources and software to support your research and development needs. We also encourage inspecting our BSD 3-Clause and APL 2.0 compatible licenses.
- ALKONTEK on GitHub.
Open Source Software
Vecmat
A simple-to-use, yet robust math and linear algebra library in C for 2d/3d graphics, machine learning, physics, and science — check GitHub and the latest docs.
CPU Feature Support
- AVX supported
- AVX2 (FMA3) supported
- AVX-512F (AVX-512 FMA) supported
- AVX10 (FMA3) work in progress
- AVX10.1 (Xeon 6) coming in 2027
- AVX10.2 (Xeon 7) tbd
- SVE (ARMv8.2-A+) supported
- SVE2 (ARMv9) supported
- NEON/ASIMD (ARMv7-ARMv9) supported
MMA Support
- WMMA / MMA (NVIDIA/CUDA) work in progress
- MFMA / WMMA (AMD/ROCm) work in progress
- AMX (4th–7th generation Intel Xeon) coming in 2027
- SME / SME2 (ARMv9.2-A+) tbd
Microserve
A stand-alone HTTP server for experimentation and development — check GitHub and the latest docs.
Protocol Support
-
HTTP/3 (RFC 9114)
supported
- QUIC (RFC 9000) supported
-
HTTP/2 (RFC 9113)
supported
- Cleartext / h2c (RFC 9110) supported
- HTTP/1.1 (RFC 9112) supported
- HTTP/2 and HTTP/3 can share a port (TCP+UDP).
Sites and Listeners
- Stand-alone first. Multi-site YAML is opt-in: several
server_nameblocks, per-site TLS, locations (root/proxy_pass), named backends, index files, and access logs. - CLI listen overlay:
--listen,--address, and--protoignore YAMLserversand start a single site. - Dual-stack:
--ipv6/--no-ipv4expand unspecified binds; explicit addresses stay as written. - Process isolation: one child process per site when two or more sites are resolved. A single site still runs in-process.
Proxy, Auth, and Federation
- Reverse proxy: nginx-style
proxy_passto a host or named backend (backend://api/). Hop-by-hop headers are stripped; failures return 502 or 504. - H2C origins: an h2c-only socket keeps HTTP/1.1 closed unless configured. The proxy can speak HTTP/2 prior-knowledge (
h2c://). - HTTP Basic Authorization per location, with
{SHA512}.htpasswd(RFC 7617). - Optional federation (
-f federation.yaml): hub or member, certified or unverified membership, and certs that can be revoked. Stand-alone remains the default. - Logging and timeouts: per-site error logs, YAML
access_log, and connect, read, and write timeouts.
Library and Tests
- C++20 modules. Dependencies are Boost.Asio/Beast, nlohmann_json, nghttp2, ngtcp2/nghttp3, and OpenSSL.
- Built-in YAML parser for config — no extra YAML library. The Boost dependency is probably temporary.
- Over 100 test cases and 700 assertions, on Catch2.
- Also shipped as a static library, so it can be embedded in other applications.
Open Source Licenses
Meanwhile, you can check out our open source licenses that will tackle in the future this probably exactly:
- TEK Public License 1.0
(first draft) - TEK Apache License 1.0
(first draft)
Vectorized Type Reference by Vendor
Intel (CPU)
| Generation | AVX Supported Types | AMX Supported Types |
|---|---|---|
| 4th Gen Xeon | AVX, AVX2, AVX-512 | AMX-TILE, AMX-BF16, AMX-INT8 |
| 5th Gen Xeon | AVX, AVX2, AVX-512 | AMX-TILE, AMX-BF16, AMX-INT8 |
| Xeon 6 | AVX, AVX2, AVX-512, AVX10.1 | AMX-TILE, AMX-BF16, AMX-INT8, AMX-FP16 |
| Xeon 7 | AVX, AVX2, AVX-512, AVX10.1, AVX10.2 | AMX-TILE, AMX-BF16, AMX-INT8, AMX-FP16, AMX-FP8, AMX-TF32, AMX-AVX512 |
AMD (CPU)
| Generation | Architecture | Supported Vector Extensions |
|---|---|---|
| 2nd/3rd Gen EPYC (Rome/Milan) |
Zen 2 / Zen 3 | AVX, AVX2 |
| 4th Gen EPYC (Genoa/Bergamo) |
Zen 4 / Zen 4c | AVX, AVX2, AVX-512 (256-bit data path) |
| 5th Gen EPYC (Turin) |
Zen 5 / Zen 5c | AVX, AVX2, AVX-512 (native 512-bit data path) |
ARM (CPU)
| Generation / Extension | SVE Support | Supported Types |
|---|---|---|
| ARMv8.2-A+ | SVE (optional) | Z0–Z31 (128–2048-bit); P0–P15; INT8, INT16, INT32, INT64, FP16, FP32, FP64 |
| ARMv9-A | SVE2 (baseline) | Z0–Z31 (128–2048-bit); P0–P15; INT8, INT16, INT32, INT64, FP16, FP32, FP64 |
| ARMv9.2+ | SVE2 + SME/SME2 (optional) | Z0–Z31, P0–P15, ZA tiles;
INT8, INT16, INT32, INT64, BF16, FP16, FP32, FP64 |
NVIDIA (GPU)
| Generation | Tensor Core Gen / API | Supported Types (A/B → Acc) |
|---|---|---|
| Ampere (SM 8.0) (A100) |
3rd Gen WMMA + mma.sync |
FP16 → FP16 / FP32 BF16, TF32 → FP32 FP64 → FP64 INT8, INT4 → INT32 |
| Ada Lovelace (SM 8.9) (L40/L40S/L4) |
4th Gen WMMA + mma.sync |
FP16 → FP16 / FP32 BF16, TF32 → FP32 FP64 → FP64 INT8, INT4 → INT32 FP8 → FP16 / FP32 |
| Hopper (SM 9.0) (H100/H200) |
4th Gen WMMA + mma + wgmma |
FP16 → FP16 / FP32 BF16, TF32 → FP32 FP64 → FP64 INT8 → INT32 FP8 → FP16 / FP32 |
| Blackwell (SM 10.x) (B100/B200) |
5th Gen tcgen05.mma |
FP16 → FP16 / FP32 BF16, TF32 → FP32 FP64 → FP64 INT8 → INT32 FP8 → FP16 / FP32 FP6, FP4, MXFP, NVFP4 → FP32 |
AMD (GPU)
| Generation | Architecture | Supported Types (A/B → Acc) |
|---|---|---|
| MI100 series (MI100) |
CDNA1 |
FP32 → FP32 FP16, BF16 → FP32 INT8 → INT32 |
| MI200 series (MI210/MI250/MI250X) |
CDNA2 |
FP32 → FP32 FP16, BF16 → FP32 INT8 → INT32 FP64 → FP64 |
| MI300 series (MI300A/X, MI325X) |
CDNA3 |
FP32 → FP32 FP16, BF16, FP8, BF8, TF32 → FP32 INT8 → INT32 |
| MI350 series (MI350X/MI355X) |
CDNA4 |
FP32 → FP32 FP16, BF16, FP8, BF8, FP6, FP4, TF32, MXFP → FP32 INT8 → INT32 |
Intel (GPU)
| Generation | Architecture / XMX | Supported Types (A/B → Acc) |
|---|---|---|
| Alchemist (Arc A-series A310/A380/A580, A750/A770) |
Xe-HPG XMX (DPAS) |
FP16, BF16 → FP32 INT8, INT4, INT2 → INT32 |
| Battlemage (Arc B-series B570/B580) |
Xe2-HPG XMX (enhanced) |
FP16, BF16, TF32 → FP32 INT8, INT4, INT2 → INT32 |