Launched this week

AXIM OS
Universal CUDA-Free compute & graphics runtime engine
1 follower
Universal CUDA-Free compute & graphics runtime engine
1 follower
AXIM OS is an enterprise-grade, hardware-agnostic compute and graphics runtime engineered by Cognitus Lab. Deploy modern AI inference models and high-performance workloads across Apple Silicon, AMD, and Intel without CUDA.



Free
Launch Team

AWS Startup AdvisorPersonalized AWS guidance that meets you where you’re already building
Promoted
Hey Product Hunt community! 👋
We're proud to launch AXIM OS. An enterprise-grade, universal CUDA-free compute and graphics runtime engineered by Cognitus Lab.
For too long, high-performance compute and modern AI inference models have been artificially tethered to proprietary CUDA stacks. AXIM OS frees developers to run heavy inference, post-transformer architectures, and graphics compute across Apple Silicon, AMD, and Intel hardware.
Key Highlights:
CUDA-Free Execution: Hardware-agnostic compute layer without NVIDIA lock-in.
Universal Acceleration: Full Metal acceleration on Apple Silicon, ROCm on AMD, and oneAPI on Intel.
AI Inference Optimized: Native support for next-gen post-transformer architectures and real-time execution.
PyPI Package: Install via pip install axim-os. {https://pypi.org/project/axim-os/}
License Key: Life-time ($29)
You can run and test the latency and throughput with any of the converted Models on our HF page: https://huggingface.co/GRRNMAKE
********************************************Benchmarks****************************************************
Benchmark Test
Hardware Target
AXIM OS Score (CUDA-Free / Fused)
CUDA / Eager Baseline Score
Head-to-Head Comparison & Outcome
Full 7B LLM Layer Decode
Apple M3 (Metal 3.0)
877 tokens/sec
(1.140 ms / layer)
134 tokens/sec
(7.440 ms / layer, 9 launches)
6.5
×
6.5\times
6.5× Speedup
(Eliminated 6.3 ms of multi-kernel dispatch jitter)
GPU Memory Bandwidth Saturation
Apple M3 Unified Memory
69.0 GB/s
10.8 GB/s
6.4
×
6.4\times
6.4× Higher Bandwidth
(69% vs. 10.8% of theoretical DRAM peak)
FSM Recurrence (
S
=
8
S=8
S=8 States)
Apple M3 (Metal 3.0)
1,012
μ
\mu
μs
12,099
μ
\mu
μs
12
×
12\times
12× Speedup
(Fused single dispatch vs. sequential round-trips)
FSM Recurrence (
S
=
16
S=16
S=16 States)
Apple M3 (Metal 3.0)
1,048
μ
\mu
μs
31,450
μ
\mu
μs
30
×
30\times
30× Speedup
(constant latency under scaling)
FSM Recurrence (
S
=
32
S=32
S=32 States)
Apple M3 (Metal 3.0)
1,060
μ
\mu
μs
72,100
μ
\mu
μs
68
×
68\times
68× Speedup
(Full algorithmic and dispatch unrolling)
Physical Render Engine
Apple M3 (Metal 3.0)
1,297.1 FPS
(0.771 ms / frame)
60.0 FPS
(16.6 ms standard target)
21.6
×
21.6\times
21.6× Real-Time Margin
(Shared-device compute + graphics pipeline)
NVIDIA Hopper Compute (
N
=
4.2
M
N=4.2\text{M}
N=4.2M)
NVIDIA GH200 480GB
613.0 M elem/s
4,257.0 M elem/s
(Proprietary torch/CUDA)
Zero-CUDA Portability PASS
(Unmodified bytecode on Hopper silicon)
NVIDIA Turing Compute (
N
=
4.2
M
N=4.2\text{M}
N=4.2M)
NVIDIA Tesla T4
106.4 M elem/s
20,062.0 M elem/s
(Proprietary torch/CUDA)
Zero-CUDA Portability PASS
(Unmodified SPIR-V bytecode on Turing silicon)