|Introduction||November 7, 2017 (announced)|
May 3, 2018 (launched)
|Frequency||1,000 MHZ, 1,300 MHz|
|TDP||80 W, 110 W|
MLU100 is a high-performance neural processor designed by Cambricon and introduced for the server market in mid-2018. The MLU100 is manufactured on TSMC's 16 nm process and operates at 1 GHz with a high-performance mode of 1.3 GHz.
In balanced mode which has a power envelope of 80 W, this processor has a theoretical peak performance of 128 trillion fixed-point (8-bit int) operations per second and 64 half-precision floating point (16-bit) TFLOPS. The efficiency of the chip slightly worsens at high-performance mode in favors of higher performances. In this mode the processor has a peak performance of 166.4 TOPS and 83.2 half-precision TFLOPS.
|Balanced Mode||High-Perf Mode|
|Frequency||1 GHz||1.3 GHz|
|8-bit Int||128 TOPS||166.4 TOPS|
|16-bit FP||64 TFLOPS||83.2 TFLOPS|
Cambricon also sells the MLU100 in a PCIe accelerator card form factor. The x16 PCIe card comes in two versions with a 16 GiB and a 32 GiB of DDR4.
|base frequency||1,300 MHz (1.3 GHz, 1,300,000 kHz) +|
|first announced||November 7, 2017 +|
|first launched||May 3, 2018 +|
|full page name||cambricon/mlu/mlu100 +|
|has ecc memory support||true +|
|ldate||May 3, 2018 +|
|market segment||Server +|
|max memory bandwidth||95.37 GiB/s (165.944 GB/s, 97,658.88 MiB/s, 0.0931 TiB/s, 0.102 TB/s) +|
|max memory channels||4 +|
|model number||MLU100 +|
|supported memory type||DDR4-3200 +|
|tdp||80 W (80,000 mW, 0.107 hp, 0.08 kW) + and 110 W (110,000 mW, 0.148 hp, 0.11 kW) +|