MIMD via 1088 cores per chip, each core has 512 bit short vector SIMD, and a 1024 bit tensor unit. ~100 MB of SRAM on chip, 8 GDDR busses to DRAM off chip. It

MIMD via 1088 cores per chip, each core has 512 bit short vector SIMD, and a 1024 bit tensor unit. ~100 MB of SRAM on chip, 8 GDDR busses to DRAM off chip. It's purpose designed for parallel sparse matrix ML problems. It's more efficient than both a CPU and GPU at these, as well as faster in absolute terms, taking their numbers at face value.