News Deep Dive: Nvidia Inference Research Chip Scales to 32 Chiplets

[Admin]

Nvidia has presented its RC 18 research chip for inference acceleration. The small 16nm chip with RISC-V core scales to 32 units in a package, delivering over 100 TOPS at over 100W.

Deep Dive: Nvidia Inference Research Chip Scales to 32 Chiplets : Read more

 

[bit_user]

In a 6x6 36-die configuration, the frequency is lowered to 1.8GHz, resulting in 128TOPS at 106W power consumption.

Hmmm... a RTX 2080 Ti is quoted at ~228 int8 TOPS in 260 W.

Source: https://devblogs.nvidia.com/nvidia-turing-architecture-in-depth/

So, not an absolute improvement. Efficiency-wise, that's 1.2 TOPS/W vs. 0.88 TOPS/W or a 38% improvement! However, the RTX card is 12 nm and this is 16 nm. And, if you compare the silicon area of 216 mm^2 (@ 16 nm process) the TU102's 754 mm^2 (@ 12 nm process), it's clear they're onto something.

BTW, nice article, Arne. IMO, it has just the right amount of detail and explanation.