Hot fries 34 has been the event of the month for us (so far). A lot of information has been published such as that of NVIDIA Superchip Grace Hopper, NVIDIA GPU Hopper and the Birentech BR100.
What is the NVIDIA Grace processor?
We first heard about superchip Grace Hopper during NVIDIA Keynote GTC 2022. The mention of a CPU of Nvidia shocked everyone. The Grace Hopper superchip is essentially a superchip with two chips on a motherboard. One is NVIDIA’s Hopper-based GPU and the other is Grace-based CPU.
New from NVIDIA Evolutionary consistency Tissue (SCF) mesh interconnect enables massive bandwidth of 3.2 TB/s through various Grace chip units. This mesh is scalable up to 72+ cores where each CPU has 117 MB of L3 cache.
NVIDIA Grace Processor Features 72 Arms v9.0 cores designed using TSMC 4N process node which is basically an improved version of the 5nm treat. With 25.1 Billion transistors, expect lightning-fast computing power.
Nvidia plans to use 512 GB LPPDR5x memory for his Grace Hopper superchip. Why, you may ask. 512 GB of LPDDR5x memory distributed over 32 channels offers the best measures of efficiency/savings while performing as well as other alternatives.
To enable chip-to-chip interconnection, NVIDIA introduced its NVLink Technology. This interface offers a bandwidth of approximately 900 GB/s who is 7x just one more PCIe 5.0 x16 interface. NVLink-C2C uses only 1.3 pJ/bit Which one is 5x more efficient than PCIe generation 5.0.
|interconnection||Picojoules per bit (pJ/b)|
|OPCe||0.5 – 0.25 pJ/b|
|infinity fabric||~1.5 pJ/b|
|TSMC CoWoS||0.56 pJ/b|
|Bunch of Wires (BoW)||0.7 to 0.5 pJ/b|
According to NVIDIA, the full superchip should use approximately 500W power. It’s actually impressive considering how much power it can deliver. AMD EPYC 7763‘s (2x) uses approx. 560W (280×2) of power, so NVIDIA is actually ahead.
Unfortunately, even that can’t play Minecraft with RTX enabled because NVIDIA states that this superchip was designed specifically for AI workloads. Grace CPUs are more focused on high performance computing, while GPU Hopper is for AI, HPC training.