Need guidance on GPU costs and compatibility for building an LLM inference and image generation server?
Need guidance on GPU costs and compatibility for building an LLM inference and image generation server?
Sorry I'm asking multiple questions, but I'm really enthusiastic about this build and want to get it right! If this isn't the correct way to ask questions here, please let me know.
My Planned Build:
CPU
: AMD Ryzen 9 9950X
GPU
: RTX 5090 at ₹2,70,000 ($3,200) OR RTX 5000 ADA at ₹3,07,000 ($3,650) -
Need advice on pricing!
Motherboard
: ASRock B850M Steel Legend WiFi
RAM
: G.Skill DDR5-6000 64GB (2x32GB) Kit
Storage
: Samsung 990 Pro 2TB with heatsink
PSU
: MSI MAG A1000GL 1000W 80+ Gold (upgrading to 1200W based on research)
Cooling
: Arctic Liquid Freezer III 280 AIO (upgraded from 240mm)
Case
: Cooler Master NR200P V3 (ITX case)
Monitor
: GIGABYTE 27" 1440p 165Hz Gaming Monitor
Budget
: ₹4,00,000 (~$4,750)
Primary Use Cases (Priority Order):
LLM inference
(Llama 3.0/3.1, 7B-30B models)
AI image generation
(Flux, Stable Diffusion)
Personal experimentation
with AI models
Gaming
(Call of Duty - not a priority, but would like decent performance)
Main Questions:
1. GPU Pricing Concerns from Local Retailer (India)
I've found a local retailer in Bangalore (SP Road - major computer market) with 25+ years experience and 4.2 Google rating offering:
RTX 5090
: ₹2,70,000 ($3,200)
RTX 5000 ADA
: ₹3,07,000 ($3,650)
But online these cards cost ₹4,50,000+ ($5,350+)
Questions:
Is this pricing realistic or red flag for counterfeit/gray market?
Do US/European users get similar discounts from local retailers vs online? I am in India.
How can I verify sealed packaging authenticity?
(They claim sealed boxes, 3-year warranty)
2. ITX Build Thermal Concerns
Research suggests my RTX 5090 (575W) + Ryzen 9950X (170W) = 745W in 18.6L case might face thermal issues.
Questions:
Is 280mm AIO sufficient for 9950X under AI workloads?
Will RTX 5090 throttle in NR200P V3 with sustained AI inference?
Should I consider RTX 5000 ADA (250W) for better thermals in ITX?
3. PSU & Power Delivery
Will 1000W PSU handle RTX 5090 power spikes, or should I upgrade to 1200W?
4. GPU Choice for AI Workloads
RTX 5090 vs RTX 5000 ADA for my use case:
RTX 5090: 32GB VRAM, higher performance, consumer drivers
RTX 5000 ADA: 32GB ECC VRAM, professional stability, lower power
Which would you recommend for 12 hour per day AI server operation?
Location & Market Context:
Located
: Bangalore, India
Buying from
: Established SP Road retailers (India's major computer hardware market)
Timeline
: Planning to build within 2-3 weeks
Any advice on build optimization, GPU authenticity verification, or thermal management would be greatly appreciated!
Thank you for taking the time to help!
Purchasing from the USA will likely be significantly cheaper, though it's unclear which online stores can reliably ship to India. Perhaps eBay sellers could work, but it's difficult to confirm their reliability.
With the Ada generation GPU you won't require a 1.2KW device. Choosing the RTX 5090 would mean needing a 1.2KW unit instead.
I recommend purchasing locally to ensure warranty support. As someone from India, it would be wise to research importers and compare prices at different stores, possibly checking unboxed options.
Regarding the build, you should consider a bigger case with improved airflow if you plan to run LLMs alongside the system. While an ITX model works on paper, thermal management will become problematic with the RTX 5090. Thermal performance should be acceptable for Ada but not ideal for RTX 5090.