[WTS] [IN-KA] Complete AI Compute Rig — 3× RTX 5090 + 3× RTX 4090 | 168GB VRAM | Threadripper 7960X | WRX90E | 128GB DDR5 | 4TB NVMe — Bengaluru
---
## Specs
| Component | Details |
|---|---|
| **GPU** | 3× RTX 5090 (32GB GDDR7) + 3× RTX 4090 (24GB GDDR6X) |
| **Total VRAM** | 168GB |
| **CUDA Cores** | 114,432 (combined) |
| **CPU** | AMD Threadripper 7965WX — 24C/48T, 5.3GHz boost, 128MB L3 |
| **Motherboard** | ASUS WRX90E (88× PCIe 5.0 lanes) |
| **RAM** | 128GB DDR5 |
| **Storage** | 4TB NVMe SSD (internal) |
| **GPU Connection** | All 6 via PCIe riser extension cables, same board |
---
## What this rig runs locally (no cloud)
- ✅ Qwen2.5-72B at **full FP16** — fits in 144GB VRAM with headroom
- ✅ LLaMA 3.1 70B / LLaMA 3.3 70B (full precision)
- ✅ DeepSeek R1 70B, Mistral Large 2, Mixtral 8×22B
- ✅ DeepSeek V2.5 236B at Q3 quantization
- ✅ Local vLLM / llama.cpp inference server for team use
> 168GB VRAM is the threshold for running frontier open-weight 70B models without quality-degrading quantization. This replaces ₹80K–1.5L/month in cloud GPU costs.
---
## Hashcat benchmarks — all 6 GPUs combined
| Hash type | Speed |
|---|---|
| MD5 | ~1.15 TH/s |
| NTLM | ~2.05 TH/s |
| SHA-256 | ~151 GH/s |
| WPA2-PMKID/EAPOL | ~19.2 MH/s |
| bcrypt | ~1.47 MH/s |
| SHA-512 | ~50 GH/s |
*Per-card figures: RTX 5090 — 220 GH/s MD5, 340 GH/s NTLM, 3,800 kH/s WPA2. RTX 4090 — 164 GH/s MD5, 344 GH/s NTLM, 2,600 kH/s WPA2.*
---
## Condition & logistics
- All GPUs tested, stable, no artifacts
- Full system benchmark proof available (nvidia-smi, hashcat output, llama.cpp inference)
- **Location:** Bengaluru, Karnataka (IN-KA)
- **Shipping:** Pan-India, buyer pays insured courier. In-person inspection welcome in Bengaluru.
- **Preferred buyers:** AI startups, ML labs, security research teams, mining ops
- **Will not split** — selling as a complete unit only
**Price:** Comment or DM. Negotiable for serious buyers only.
---
