up to Dual RTX Pro 6000 Blackwell | 192GB VRAM | Threadripper Pro
Purpose-Built for Large-Scale AI Workloads & Enterprise Workflows
This server is engineered for teams who need:
• Private, deterministic AI training and inference (no cloud variability)
• Fine-tuning multi-billion parameter models on dedicated hardware
• High concurrency inference for internal copilots and client APIs
• Secure and compliant compute environments (regulated or air-gapped)
Designed for AI startups, research labs, and enterprise ML teams that require performance, control, and predictable cost.
The Novatech RM6000 is engineered for creators, AI innovators, and technical professionals who demand the absolute best in compute performance. Built in a 6U SilverStone RM600 rackmount chassis with exceptional airflow, this powerhouse features single or dual NVIDIA RTX™ PRO 6000 Blackwell Workstation GPUs—each equipped with 96GB of ultra-fast GDDR7 VRAM, delivering a massive 192GB total VRAM for unparalleled performance in AI training, simulation, 3D rendering, and large-scale data processing.
CPU: AMD Threadripper PRO 7995WX – 96 cores / 192 threads, designed for extreme multitasking and stability.
Motherboard: ASUS Pro WS WRX90E-SAGE SE – enterprise-grade reliability with PCIe 5.0 support.
RAM Options: 128GB DDR5 ECC Registered (2×64GB, 5600 MHz) – rock-solid stability under heavy workloads.
GPU: 1x or 2x NVIDIA RTX PRO 6000 Blackwell Workstation Edition (VCNRTXPRO6000BPB) – 192GB total VRAM, 96GB per GPU.
Storage Options: Samsung 990 Pro Gen 4 NVMe SSDs (1TB, 2TB, or 4TB) for lightning-fast load times and massive project files.
Cooling: Thermalright XE360-TR5 liquid cooling for the CPU, optimized for Threadripper thermals.
Power Supply: FSP Cannon Pro 2000W Platinum-rated PSU – rock-solid stability for high TDP components.
Connectivity: Enterprise networking support, with optional 100GbE NIC compatibility.
Whether you’re training advanced AI models, rendering ultra-complex 3D environments, or running high-throughput simulations, the dual Blackwell architecture offers unmatched compute density and efficiency. With PCIe Gen 5 bandwidth and ECC memory throughout, you get maximum speed and data integrity in mission-critical environments.
Total VRAM: 192GB ECC GDDR7 – massive parallel processing for AI & GPU-heavy workloads.
ECC DDR5 Memory: Error-correcting memory for unmatched reliability.
Upgradeable Storage & Networking: Scalable for future needs.
Rackmount-Ready: Perfect for data centers, production houses, and enterprise deployments.
At Novatech, we specialize in high-performance, custom-configured systems built with precision, tested for reliability, and backed by world-class customer support. Every RM6000 is stress-tested before shipping to ensure it meets the demands of your workflow from day one.
Bring next-generation power to your AI lab, VFX pipeline, or research environment with the Novatech RM6000 – Single or Dual NVIDIA RTX PRO 6000 Blackwell.
Couldn't load pickup availability
Core Compute
• AMD Threadripper PRO 7995WX – 32 cores / 64 threads — extreme multi-tasking & parallel workloads.
GPU Architecture
• Dual NVIDIA RTX PRO 6000 Blackwell — 96 GB GDDR7 each (192 GB aggregate) — massive VRAM headroom for LLMs.
Memory & Storage
• ECC Registered DDR5 — data integrity & uptime for mission-critical workflows.
• Optional 1–4 TB PCIe Gen4 NVMe — fast local datasets and model storage.
Enterprise-Grade Infrastructure
• Liquid CPU cooling + 2000W Platinum PSU — stability under long jobs.
• Rackmount chassis — 19″ rack data center ready.
Engineered Workloads:
• On-Prem LLM Fine-Tuning (70B+ params) — local, secure, full-precision training
• Production Inference at Scale — high throughput APIs with low latency
• Private Copilots — govern models & data entirely in-house
• Computer Vision & Multi-Modal Training — large GPU memory for models and datasets
• Regulation & Compliance Computing — secure, air-gapped workflows
Enterprise-aligned estimates for model fit, inference throughput, diffusion performance, and cloud ROI framing.
Capability
| Model / Class | Quantization | Inference | Fine-tuning |
|---|---|---|---|
| Llama 3 70B | FP16 / 4-bit | ✅ | ✅ |
| Mixtral 8x22B | FP16 / 4-bit | ✅ | ✅ |
| 13B–34B class | FP16 | 🚀 | ✅ |
| 405B (sharded) | 4-bit | ✅ | ⚠️ |
Legend
Actual capability varies by framework, context length, batch size, and optimization stack.
Throughput
| Model | Precision | Estimated tok/sec |
|---|---|---|
| 7B | FP16 | 600–900 tok/sec |
| 13B | FP16 | 350–600 tok/sec |
| 34B | FP16 | 180–300 tok/sec |
| 70B | FP16 | 120–200 tok/sec |
| 70B | 4-bit | 220–350 tok/sec |
Estimates assume an optimized inference stack (e.g., vLLM / TensorRT-LLM). Batch inference can materially increase throughput.
Generative
| Workflow / Model | Estimated performance |
|---|---|
| Stable Diffusion XL | ~1–2 sec/image (optimized) |
| SD Turbo | Sub-second generations (pipeline dependent) |
| Video diffusion (optimized) | Near real-time depending on resolution |
Render time depends on resolution, steps, scheduler, and pipeline optimizations.
ROI
| Scenario | Estimate |
|---|---|
| H100-class cloud instance | $90–$110/hr (varies by region/commit) |
| 8 hrs/day utilization | ~$20k–$26k/month |
| Directional break-even | Often ~4–6 months (utilization dependent) |
Cloud rates vary by region, commitment, and availability. Use as directional procurement framing.
Deploy an on-prem AI workstation designed for LLM training hardware, inference, and secure internal workflows—without giving up performance or control.
Directional, production-aligned expectations for 3D rendering, simulation, and professional video workflows.
Works great with
3D Rendering
| Software | Workload | Expected performance |
|---|---|---|
| Blender (Cycles) | GPU path tracing / final-frame rendering | 🚀 Strong dual-GPU scaling (scene/VRAM dependent); often ~2× throughput vs single GPU |
| Cinema4D + Redshift | GPU rendering + interactive lookdev | 🚀 High RTX throughput; dual-GPU scales well for final renders |
| OctaneRender | Photoreal path tracing | 🚀 Excellent multi-GPU scaling; ideal for heavy RTX workloads |
| Unreal Engine 5 | Nanite + Lumen viewport / virtual production | 🚀 Real-time viewport with complex assets; GPU + RAM help large projects |
| Houdini / Sim workloads | CPU simulation + caching | 🚀 96 cores + high RAM enables large sims and heavy caching |
Multi-GPU scaling depends on renderer support and whether the scene fits in GPU VRAM. CPU simulations benefit from core count and memory bandwidth.
Legend
Video Production
| Software | Workflow | Expected capability |
|---|---|---|
| DaVinci Resolve | 8K editing + color grading (GPU accelerated) | 🚀 Smooth timeline playback in many workflows; strong GPU effects performance |
| Adobe Premiere Pro | Multi-stream 4K editing + GPU effects | ✅ Excellent real-time editing; performance depends on codecs/effects stack |
| After Effects | Large comps + RAM-heavy projects | 🚀 High RAM supports massive compositions and heavy caching |
| Media Encoder / Delivery | H.264 / HEVC export (GPU accelerated) | ✅ Fast exports with GPU acceleration (workflow dependent) |
Playback/export depends on codec, effects stack, GPU acceleration settings, and storage throughput. Use these as directional expectations.
Large Project Handling
| Resource | Production benefit |
|---|---|
| 512GB ECC RAM | Massive scenes and simulations |
| 96-core Threadripper Pro | High performance physics, CPU rendering, simulation |
| Dual RTX Pro 6000 GPUs | High ray-tracing throughput |
| Large VRAM capacity | Handles complex textures and geometry |
| Up to 10TB NVMe storage | Fast loading of large asset libraries |
Best results come from aligning GPU VRAM capacity to your largest scenes (textures/geometry) and keeping active projects on fast NVMe storage.
Render Speed Comparison
| System | Relative render speed |
|---|---|
| Typical RTX 5080 workstation | 1× |
| Single RTX 5090 or Pro 6000 workstation | ~1.4–1.6× |
| Dual RTX 5090 or Pro 6000 workstation | ~2.4–2.8× |
Note: Estimates are directional and based on expected workstation-class behavior for comparable hardware. Actual results vary by scene complexity, codecs, effects, renderer settings, and software versions.
Is it powerful enough for machine learning?
Can it handle big datasets and multitasking?
Is it reliable for long workloads and 24/7 operation?
Will it stay relevant as technology advances?
Is this workstation only for AI?
Cut model training times and speed up simulations with the AMD Threadripper PRO 9995WX (96 cores, 192 threads) and NVIDIA RTX PRO 6000 with 96GB VRAM.
Run AI, HPC, data analytics, 3D rendering, and engineering simulations without bottlenecks — keeping teams productive and projects on schedule.
With 512GB DDR5 ECC memory, 10TB of Gen 5 NVMe storage, and a future-ready WRX90E platform, your workstation grows with your company’s needs.
