Samsung and Naver Developing an AI Chip Claiming to be 8x More Power Efficient than NVIDIA H100

[btarunr]

Naver, the firm behind the HyperCLOVA X large language model (LLM), has been working with Samsung Electronics toward the development of power-efficient AI accelerators. The collaboration brings Naver's expertise with Samsung's vast systems IP over silicon design, the ability to build complex SoCs, semiconductor fabrication, and its plethora of DRAM technologies. The two recently designed a proof of concept for an upcoming AI chip, which they iterated on an FPGA. Naver claims the AI chip it is co-developing with Samsung will be 8 times more energy efficient than an NVIDIA H100 AI accelerator, but did not elaborate on its actual throughput. Its solution, among other things, leverages energy-efficient LPDDR memory from Samsung. The two companies have been working on this project since December 2022.



View at TechPowerUp Main Site | Source

 

[ncrs]

It's kind of weird that they are calling the H100 an "AI accelerator" while it is a general purpose compute chip. Tensor cores aren't limited to AI either since they are matrix multiplication engines capable of working with data types from AI-centric INT8/FP8/FP16/BFLOAT16 to HPC-oriented FP32/FP64.
Are they saying that the new chip will be capable of the same level of general purpose compute as H100 while being 8x more power efficient? If not, and it's just for AI then this isn't something really revolutionary since we've had multiple AI-specific chips in the market already.
Oh and they still need to face the CUDA domination in the AI/ML software ecosystem.

 

[sLowEnd]

Oh and they still need to face the CUDA domination in the AI/ML software ecosystem.

Naver is like Korean Google. They'll probably just make these for their own servers and write bespoke software for it.

 

[Assimilator]

8x more energy-efficient with 0.08x the performance, I'll bet.

 

[Denver]

Competition is growing from all sides, strong and challenging. Nvidia, which previously swam alone, now has the biggest companies on the planet seeking independence to sail alone with their own hardware. Yay.

 

[chodaboy19]

By the time this comes out Nvidia will already have released the B100, competitors are always trying to compete with previous gen products from Nvidia.

 

[Minus Infinity]

By the time this comes out Nvidia will already have released the B100, competitors are always trying to compete with previous gen products from Nvidia.

b100 won't be 8x a power efficient guaranteed. Maybe 2x the performance per watt. An ASIC will always be way more efficient than a general purpose chip running the intended applications.