2024-07-09 16:13:18
Nvidia AI Chip Using Low-Power DRAM Becomes the ‘Problem Solver’
Grace CPU power consumption is 1/8th that of general DRAM
Memory capacity is likely to increase with the advent of on-device AI
As Nvidia prepares to launch its new AI (artificial intelligence) semiconductor, ‘Blackwell’, its low-power DRAM (LPDDR) is once again attracting attention.
Low-power DRAM has been used primarily in mobile products, but it is increasingly being used in the AI semiconductor field, which is likened to a ‘power-hungry hippopotamus’.
There are even voices saying that the hit product that will heat up the market the most in the second half of this year, along with HBM (high bandwidth memory), is low-power DRAM.
According to industry sources on the 9th, NVIDIA will begin mass production of the ‘GB200 Grace Blackwell Superchip’ in the second half of this year. This superchip consists of one NVIDIA Grace CPU (central processing unit) and two new Blackwell GPUs (graphics processing units).
In particular, the Blackwell GPU is considered the next-generation GPIU that will replace the existing ‘Hopper (H100)’ chip, which is the representative of AI semiconductors. It contains a total of 16 HBM, the superstar of memory semiconductors. This is to support data processing functions for AI operations.
The Grace CPU, which acts as the ‘brain’ of the super chip, is also closely coupled with the Blackwell GPU to double computing power.
Above all, Grace CPUs are specialized in reducing power consumption.
The memory mounted on this CPU is also a 7th generation low-power DRAM, ‘LPDDR5X’, with 16 units. This means that the power required per GB (gigabyte) is only 1/8 of that of DDR5 used in general CPUs. This is why Grace CPUs are far more power efficient than general CPUs.
Nvidia CEO Jensen Huang recently said, “As generative AI advances, memory becomes more important,” and “The memory used in the Grace CPU is LPDDR, which can save a lot of power.”
◆Low-power DRAM advances beyond mobile to data center market
The Grace CPU performance, backed by the low power performance of LPDDR5, is already well known.
According to ‘Supercomputing’ data, which compiles rankings of supercomputers around the world, systems using Nvidia Grace CPUs swept the top ranks in the Green500, which selects the most energy-efficient systems.
As power consumption decreases, so does the heat generated. The heat generated by data sensors, which are increasing rapidly around the world, is being pointed out as a cause of global warming.
Accordingly, low power consumption in data centers could become a hot topic in the semiconductor industry, and the GB200 superchip equipped with Grace CPU could be a new solution.
According to Taiwanese media, GB200 is manufactured mainly through Taiwanese supply chains, including TSMC, the world’s number one foundry. It is currently so popular that supply cannot keep up with demand.
On-device AI, such as AI smartphones and AI PCs, also absolutely require low-power DRAM.
Market research firm TrendForce predicted that low-power DRAM will continue to grow by 3-8% in the third quarter of this year despite concerns over oversupply.
Apple’s recently released generative AI service ‘Apple Intelligence’ is only available on the latest Apple devices, such as iPhone 15 Pro and Pro Max. This is due to insufficient memory capacity on older devices.
Apple is going one step further and will be installing 8GB (gigabyte) LPDDR5X in the iPhone 16 series to be released in the second half of this year. As AI functions become increasingly important in mobile phones, LPDDR, which can simultaneously secure power and performance, is essential.
PC manufacturers are also actively considering utilizing low-power memory to perform generative AI services.
DRAM manufacturers have now developed a new memory form, LPCAMM2, which is a high value-added form that combines LPDDR packages into one, and has now begun commercializing it. LPCAMM2 reduces the mounting area by 60% compared to existing products while using less power.
[서울=뉴시스]
2024-07-09 16:13:18