Huawei announces self-made HBM memory to boost AI chips potential
Today, Huawei unveiled several chip technologies that will launch, starting later this year and one of them is the self-made HBM memory for AI chips. The particular product suggests the company’s win over Nvidia’s lead in the AI chipset market.
HBM or High-Bandwidth Memory significantly improves AI chip functions and performance. It is responsible for stacking DRAM dies vertically to lessen signal paths and increase the overall bandwidth of the chipset.
On the greener side, HBM memory can reduce power consumption for data-intensive AI workloads like LLM training and inference. It minimizes data movement as it places the memory directly beside the processor.
Due to the US-China tensions and sanctions, Huawei can’t access HBM chips from foreign firms. Hence, it decided to turn the tables in its favor with self-made tech.
According to the details, Huawei has two versions of its self-made HBM memory for AI chips: HiBL 1.0 and HiZQ 2.0. Both memory versions claim to provide a new boost and potential for the upcoming Ascend processors.
HiBL 1.0: It has a bandwidth of 1.6TB/s and a capacity of 128GB. The OEM will use this memory for the Ascend 950PR, debuting in the first quarter of next year.
This HBM memory will help 950PR to support low-precision data formats like FP8, MXFP8, HIF8: 1 PFLOPS, MXFP4: 2 PFLOPS, and more. It will further focus on improving vector computing power and increasing the overall interconnection by 2.5 times.
HiZQ 2.0: On the other hand, the Ascend 950DT will use HiZQ 2.0 memory to improve the inference decoding performance. It supports a bandwidth of 4TB/s and 144GB capacity.
Perhaps the new Huawei self-made HBM memory will eventually add more potential to Ascend processors, making them a strong competitor against Nvidia AI chips.
READ the latest news shaping the AI Chips market at AI Chips News
Huawei announces self-made HBM memory to boost AI chips potential, source
Add comment