TRANSFORMING ​the coSt of ​MACHINE LEARNING

Hardware

See us at neurips 2023: Booth 309

POSITRON

ATLAS

Transformer Inference ​Appliance

5x Better Performance per Dollar ​vs DGX-H100

• 2.3x faster
  
• ½ the cost
  
• ½ the power
  

Highlights:

• 2TB+ Device Memory
  
• Support 500B+ parameter models
  
• Individual LoRAs per user
  
• Llama 2 70B performance
  

Batch 1 → 480 tokens/sec/user

Batch 8 → 160 tokens/sec/user

Available Spring 2024

*All performance estimates are subject to change. All numbers are based on BF16 computation, and without speculative decoding or paged attention.

Building the Positronic Brains of ​the future

Positron AI, Inc.

2310 N Molter Street, Suite 308

Liberty Lake, WA 99109

510-365-2166

contact@positron.ai