OPEN_SOURCE ↗
REDDIT · REDDIT// 20d agoINFRASTRUCTURE
Modded 2080 Tis: 176GB VRAM for $2,000
A budget-conscious LLM builder explores filling a server with modded 22GB RTX 2080 Ti cards to achieve 176GB of VRAM for under $2,000. This configuration offers an unprecedented VRAM-to-price ratio for massive models, though the aging Turing architecture lacks modern features like BF16 and FlashAttention 2 support.
// ANALYSIS
This high-VRAM "Frankenstein" build is the ultimate budget king for raw capacity, but it trades modern architectural features for sheer memory volume.
- –Modded 22GB cards provide 2x the VRAM of standard units, allowing 70B+ models to run entirely on-GPU without the massive cost of enterprise-grade silicon.
- –Lack of native BF16 support can lead to numerical precision issues and slower processing in modern model weights originally trained in Bfloat16.
- –Absence of FlashAttention 2 significantly handicaps prefill speeds, making the user experience noticeably slower for long-context prompts.
- –Reliability remains a primary concern as these third-party hardware modifications lack warranties and are prone to failure under sustained high-heat server workloads.
- –Power draw for eight 250W+ cards exceeds 2kW at peak, making this setup expensive to operate despite the low initial acquisition cost.
// TAGS
gpullmself-hostedinferencertx-2080-ti
DISCOVERED
20d ago
2026-03-23
PUBLISHED
20d ago
2026-03-23
RELEVANCE
8/ 10
AUTHOR
PossiblePossible2571