White Papers

The following technical white papers are published by Openzeka Teknoloji A.Ş.

Title Topic Platform
Local LLM Usage Guide Hardware → model → software decision guide Jetson, RTX PRO, DGX Spark, DGX/HGX
LLM Quantization Selection Guide Quantization selection: workload × format × GPU architecture × checkpoint reliability, with the selection flow and case examples NVIDIA Ampere → Blackwell (Jetson, RTX, DGX)
Engram Offloading in LLM Inference Large embedding tables read a few rows per token: offloading them to host RAM or NVMe, with measured results DGX Spark, RTX PRO 6000; DGX B300 memory architecture
NVIDIA DGX B300 vs GB300 NVL72 Cluster Architecture Comparison Technical comparison of two Blackwell Ultra architectures, workload-based platform selection guide NVIDIA DGX B300, GB300 NVL72 (Blackwell Ultra)
DGX Spark 2-Node AI Cluster Setup Guide Point-to-point topology cluster setup, RoCEv2/RDMA, sparkrun 2x NVIDIA DGX Spark (GB10)
DGX Spark 3-Node AI Cluster Setup Guide Ring (mesh) topology cluster setup, RoCEv2/RDMA, sparkrun 3x NVIDIA DGX Spark (GB10)
DGX Spark 4-Node AI Cluster Setup Guide Switch-based cluster setup, RoCEv2/RDMA, sparkrun, NAS 4x NVIDIA DGX Spark (GB10)
DGX Spark 8-Node AI Cluster Setup Guide Switch-based cluster setup, RoCEv2/RDMA, sparkrun, NAS 8x NVIDIA DGX Spark (GB10)
Qwen3.6-27B DGX Spark Benchmark LLM quantization comparison (FP8/AWQ/NVFP4 + MTP) NVIDIA DGX Spark (GB10)
Qwen3.6-27B DGX Spark Cluster Scaling Multi-node scaling (TP1/TP2/TP4), SLO-driven capacity planning 1x/2x/4x NVIDIA DGX Spark (GB10)
KV Cache Offloading with LMCache in LLM Inference Repeated contexts and a KV cache bigger than GPU memory: three cache tiers (GPU, CPU, NVMe), measured TTFT and the settings to verify before production NVIDIA DGX Spark (GB10)
Network Link Speed for Tensor-Parallel Inference on 8× DGX Spark (200G vs 100G) What tensor parallelism sends over the network, an estimate of the 100G penalty from the model’s architecture, and a measured 200G vs emulated 100G comparison with GLM-5.3 at TP8 8× NVIDIA DGX Spark (GB10)
The First Open-Weight Alternatives to Jev Six open-weight typed-decision models compared on one DGX Spark: accuracy, calibration, option-order sensitivity and latency, and which model fits which job NVIDIA DGX Spark (GB10)
DeepSeek-V4.1-Flash 4× DGX Spark Deployment 763B MoE model deployment on 4× DGX Spark (GB10) with TP4: vLLM build chain, 7 SM121 patches, Engram-on-disk, DSpark k=5 4× NVIDIA DGX Spark (GB10)
DeepSeek-V4.1-Flash 8× DGX Spark TP8 Deployment 763B MoE model on 8× DGX Spark (GB10) with TP8: two configurations (300K Engram-in-memory, 1M Engram-on-disk), NCCL optimization, benchmark results and TP4 comparison 8× NVIDIA DGX Spark (GB10)
Kimi K3 Inference Benchmark on DGX-B300 Inference engine + speculative decoding comparison (vLLM vs SGLang, direct vs DSpark) NVIDIA DGX-B300 (8x Blackwell Ultra, TP=8)
LLM Inference Benchmark Explorer Interactive benchmark table: filter by device, model and quantization, set your own service-level thresholds DGX Spark (GB10), DGX B300, RTX PRO 6000, Jetson Thor
CV Inference Benchmark Explorer Interactive computer-vision benchmark: pick a device and a detection model to see sustained FPS, how it falls as cameras are added, and how many cameras it carries at your target FPS DGX Spark (GB10), Jetson AGX Thor, Jetson Orin Nano, RTX 3060, RTX 3090
VLM Inference Benchmark Explorer Interactive vision-language model benchmark: pick the image size and the longest a camera may wait for its answer, and see how many cameras each configuration keeps up with RTX PRO 6000 Blackwell, Jetson AGX Orin 32GB, Jetson Orin NX 16GB

On this page