DeepX DX-M2’s 100B Low-Power Target Still Awaits Real-World Testing

Written by

in

This article analyzes DeepX DX-M2’s performance claims after tape-out, noting that its stated figures of 80 TOPS, 150 GB/s, and less than 5W have not yet been independently tested on a public chip. Based on 4-bit quantization, achieving 10 to 30 tokens/s with a 100B sparse MoE model depends on activating only 10B to 15B parameters, while the full model would still require about 50 GB of external memory. Samples are targeted for March 2027, with mass production targeted for fall 2027. Model support, the toolchain, pricing, and the actual power-consumption range also remain unclear.

Source:explainx.ai · Read the original article

Comments

Leave a Reply

Your email address will not be published. Required fields are marked *