What is NVIDIA’s Grace CPU?
We first heard of the Grace Hopper superchip during NVIDIA’s 2022 GTC Keynote. The mention of a CPU from NVIDIA shocked everyone. The Grace Hopper superchip is essentially a Superchip featuring two chips on one motherboard. One is NVIDIA’s Hopper based GPU and the other is the Grace based CPU.
NVIDIA’s new Scalable Coherency Fabric (SCF) mesh interconnect allows for a massive bandwdith of 3.2TB/s across various Grace chip units. This mesh is scalable for up to 72+ cores where each CPU has 117MB of L3 Cache.
NVIDIA’s Grace CPU features 72 Arm v9.0 cores designed using TSMC’s 4N process node which is essentially an improved version of the 5nm process. Featuring 25.1 Billion transistors, expect extremely fast computational power.
NVIDIA is planning to use 512GB LPPDR5x memory for its Grace Hopper superchip. Why, you may ask. 512GB of LPDDR5x memory spread across 32-channels offers the best efficiency/cost saving metrics while performing just as good as the other alternatives.
To enable chip-to-chip interlink, NVIDIA introduced its NVLink technology. This interface provides a bandwidth of around 900GB/s which is 7x more than a PCIe 5.0 x 16 interface . NVLink-C2C uses just 1.3 pJ/bit which is 5x more efficient than PCIe Gen 5.0.
|Interconnect||Picojoules per Bit (pJ/b)|
|UCIe||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|
As per NVIDIA, the complete superchip is expected to use around 500W of power. That’s actually impressive considering the power it can offer. AMD EPYC 7763‘s (2x) uses around 560W (280×2) of power, so NVIDIA is actually in lead.
Sadly, even this cannot play Minecraft with RTX ON because NVIDIA states that this superchip has been designed specifically for AI workloads. The Grace CPUs are more pertained towards high performance computing, whereas the Hopper GPU is targeted for AI training, HPC.