Entry-level dual-GPU AI workstation — single or dual NVIDIA RTX 5090 (32 GB GDDR7) on AMD Ryzen 9 9950X. EMARQUE-assembled in Malaysia.
Manufacturer-defined features from the published datasheet.
Single or dual NVIDIA RTX 5090 GPUs — 32 GB of GDDR7 each, 64 GB pooled — enough headroom to load and serve a 70B-class quantized model entirely on local hardware. No cloud round-trip, no per-token billing, no data leaving the office.
Sixteen Zen 5 cores at 5.7 GHz boost handle data preprocessing, vector indexing, and parallel evaluations without bottlenecking the GPUs. Same CPU class as production servers, in a chassis that draws under 1.2 kW under load.
Validated to sub-45 dBA idle and sub-55 dBA under multi-hour training runs — comparable to a high-end gaming rig, half the volume of a server chassis. Runs in shared office space without acoustic complaints from neighbours.
Memory and storage headroom large enough to hold the model, the working dataset, and the experiment cache simultaneously. RAID 0/1 NVMe configuration means model load times measure in seconds, not minutes.
Built and QA-tested locally by the EMARQUE Lab — multi-point QA on CPU / GPU / memory / disk, benchmark report shipped with the unit. Warranty and parts handled in-country; no international RMA queue.
No special electrical work — runs from a standard Malaysian 13 A wall socket. UPS recommended for graceful shutdown but not required. Sits under a desk or on a side table; no rack, no DC, no facilities request to procurement.
The four sub-systems that determine real-workload behaviour. We tune each before delivery.
Tell us your workload. EMARQUE sizes the AI Work 100 and sends a quote.
How EMARQUE scopes this system: who it suits, when to pick it, when to pick something else, and what we add beyond the hardware.
Individual researchers, developers, and small teams who want a real on-prem AI workstation under the desk — EMARQUE-built, locally supported, without server-class price or noise.
Two RTX 5090s give 64 GB of GPU memory across both cards. Comfortably runs 7B–13B at FP16, and 30B–70B with INT4 / INT8 quantization on a single card. Tensor parallelism across the two cards is over PCIe (no NVLink).
Tower workstation on a standard 240 V wall outlet. Air-cooled, sub-55 dBA under load. 2.5 GbE on-board with optional 10 GbE add-in. No special HVAC or rack needed.
Sits between DGX Spark (single-superchip, ≤ 70B quantized) and AI PRO 500 (multi-GPU pedestal). Cheaper than DGX Spark on a per-GPU-memory basis when dual RTX 5090 is selected; step up to AI PRO 500 when departmental concurrency grows.
Step up to AI PRO 500 when you need 4 GPUs, ECC memory, or quieter office operation under sustained departmental load. Step to the EMARQUE AI Server (RTX PRO 6000 SE config) for production rackmount deployments.
Workload categories documented in the manufacturer's reference materials. Sizing is confirmed with your technical team during scoping.
Load a quantized 70B model, run LoRA / QLoRA adapter training on a domain dataset, evaluate the result — all on the dual-GPU pool. Iterate as fast as your patience, not as fast as your cloud credit limit refreshes.
Run a private RAG stack — embedder, vector store, 13B–34B chat model — for a team of 5–10 people querying internal docs. Throughput holds steady; the company's IP never reaches a third-party API.
Build the RAG pipeline, prove the latency targets, demonstrate the cost model — then scale into the AI PRO 500 or AI Server when team and traffic justify it. The AI Work 100 is the cheapest path to operational confidence.
Run a local code-completion model (StarCoder 2, Qwen Coder, DeepSeek Coder) alongside the IDE — sub-100ms suggestions with no cloud dependency. Spare GPU cycles power evening experiments without renting a cloud GPU.
Configurable. Final BOM, GPU mix, RAM and storage, and networking topology are confirmed in writing at quotation.
Yes, with quantization — Llama 3 70B at 4-bit (Q4_K_M GGUF) fits in ~40 GB of VRAM and runs well on dual RTX 5090s. Full FP16 70B does not fit; for that you want the AI PRO 500 with RTX PRO 6000 Blackwell. The AI Work 100 is sized for 7B–34B at full precision and 70B quantized.
Single 5090 for individual research on ≤34B models or 70B quantized at lower throughput. Dual 5090 doubles inference throughput, enables larger context windows via tensor parallelism, and is required for serving 5+ concurrent users on a 13B+ model. EMARQUE recommends starting dual if the budget supports it — the upgrade cost later includes labour and re-validation.
A dual-5090 AI Work 100 amortises against ~18 months of continuous cloud GPU rental at typical Malaysian developer usage. The break-even shifts shorter when you count data egress, idle-time billing, and the time cost of provisioning. Most teams see payback in 9–12 months; air-gapped workloads see it immediately because cloud isn't an option.
For inference and short-run fine-tuning, no — modern consumer DDR5 with on-die ECC and the Ryzen 9 9950X's IO die have proven reliable for AI workloads under EMARQUE's multi-point QA. If you require server-grade ECC RDIMMs and PCIe lane count for sustained 24/7 multi-user serving, the AI PRO 500 on Threadripper PRO is the right step up.
Ubuntu 24.04 LTS or Windows 11 Pro at the customer's choice, with NVIDIA driver stack, CUDA 12.6+, PyTorch, vLLM, llama.cpp, and a benchmark suite pre-installed and validated against your stated workload at delivery. Each unit ships with a written benchmark report against your reference model.
Standard EMARQUE warranty (3-year parts and labour on EMARQUE-assembled components; NVIDIA's warranty on the GPUs). Priority phone + email support from EMARQUE in Malaysia, with optional Priority Care add-on for on-site response in Klang Valley.
Two paths: (1) add a second AI Work 100 and load-balance across them at the application layer, or (2) consolidate into an AI PRO 500 (single workstation, up to 4 × RTX PRO 6000 Blackwell, ECC memory, dedicated AI-optimised cooling). EMARQUE handles trade-in and migration; the AI Work 100 retains residual value for a secondary research or developer role.
Manufacturer specifications and warranty terms apply. EMARQUE issues a formal quotation through your Key Account Manager.
Model size, concurrency, latency budget, deployment site. EMARQUE returns a quote in MYR within one Malaysian business day, sized to the workload — not the salesperson’s quota.