OPEN LINK ↗
// 1d agoINFRASTRUCTURE
Tenstorrent Blackhole cluster runs Llama 70B locally
A solo developer bypassed expensive enterprise GPUs by assembling a local hardware setup with four Tenstorrent Blackhole cards priced at $1,299 each inside a Linux workstation. By wiring the cards directly card-to-card with QSFP-DD 800 Gbit fiber optical links, the setup achieves high-bandwidth inter-card communication to run Meta's Llama 3.3 70B model locally with high energy efficiency and minimal operational electricity costs.
// ANALYSIS
Tenstorrent's RISC-V hardware architecture and high-speed card interconnects are demonstrating viable, cost-effective alternatives to NVIDIA's dominant enterprise GPUs for local 70B model inference.
- –Replaces $30,000+ enterprise GPUs like the NVIDIA H100 with a $5,200 total hardware investment across four $1,299 Blackhole cards.
- –Utilizes direct card-to-card QSFP-DD 800 Gbit fiber interconnects to eliminate memory bandwidth bottlenecks during multi-card LLM tensor parallelism.
- –Highlights remarkable power efficiency, allowing developers to host large open-weights models locally for as little as $6 a month in power draw.
// TAGS
tenstorrentblackholehardwarellmlocal-firstllama-3.3inferencerisc-vinfrastructure
DISCOVERED
1d ago
2026-07-30
PUBLISHED
1d ago
2026-07-30
RELEVANCE
8/ 10
AUTHOR
KijAkubovs86334