Grace Blackwell
A high-performance AI accelerator chip introduced by NVIDIA, integrating the Grace CPU with the Blackwell GPU architecture. It is designed for large-scale AI training and inference workloads and is widely deployed in data centers.
Core Facts
The GB10 Grace Blackwell Superchip integrates an ARM-architecture Grace CPU and a Blackwell GPU within the same package via NVLink-C2C chip-to-chip interconnect technology.
Mistral Large 4 was trained using 3,800 NVIDIA Grace Blackwell GPUs in its own data centers, and the public preview also runs on the same infrastructure.
DGX Spark is a compact AI workstation launched by NVIDIA for individual developers and small teams, featuring the Grace Blackwell chip.
Timeline (last 90 days)
By redesigning the scheduling policy, 27% more available GPU resources can be obtained from the same set of racks without additional hardware procurement.
When tightly coupled architectures like Grace Blackwell become mainstream, the scheduler's design philosophy needs to shift from "allocating devices" to "orchestrating collaboration".
There is a technical discussion on Hacker News about Grace Blackwell scheduling requirements, which received 42 points.
Within a Grace Blackwell node, the internal resource topology exhibits stronger heterogeneous affinity: some tasks have lower communication overhead on tightly coupled CPU-GPU paths, while cross-GPU communication relies on the NVLink switching network.
The new generation of Grace Blackwell achieves 100% liquid cooling, whereas only 80% of the previous generation used liquid cooling.
The assembly time of the new generation of Grace Blackwell has been shortened from two hours in the previous generation to five minutes.
At GTC, Nvidia announced that Grace Blackwell products have officially entered full production.
ML4's training was completed in Europe, using about 3,800 NVIDIA Grace Blackwell GPUs.
Grace Blackwell is NVIDIA's new-generation GPU platform for large-scale AI training, with Grace responsible for the CPU side and Blackwell for the GPU side, forming a superchip through a high-bandwidth interconnect.
Le Chonk's training used 3,800 NVIDIA Grace Blackwell GPUs located in European data centers.
11 more timeline events
All Facts (20)
The GB10 Grace Blackwell Superchip integrates an ARM-architecture Grace CPU and a Blackwell GPU within the same package via NVLink-C2C chip-to-chip interconnect technology.
75%VerifiedMistral Large 4 was trained using 3,800 NVIDIA Grace Blackwell GPUs in its own data centers, and the public preview also runs on the same infrastructure.
70%VerifiedDGX Spark is a compact AI workstation launched by NVIDIA for individual developers and small teams, featuring the Grace Blackwell chip.
65%VerifiedNVIDIA's GB300 Blackwell Ultra GPU integrates two GPU chips and one Grace CPU, interconnected via NVLink-C2C, with memory bandwidth exceeding 8 TB/s and inference performance 4 to 5 times higher than the H100.
65%UnverifiedDGX Spark基于NVIDIA的Grace Blackwell架构，采用Grace CPU与Blackwell GPU的紧密耦合设计
90%UnverifiedBy redesigning the scheduling policy, 27% more available GPU resources can be obtained from the same set of racks without additional hardware procurement.
50%UnverifiedWithin a Grace Blackwell node, the internal resource topology exhibits stronger heterogeneous affinity: some tasks have lower communication overhead on tightly coupled CPU-GPU paths, while cross-GPU communication relies on the NVLink switching network.
50%UnverifiedThe new generation of Grace Blackwell achieves 100% liquid cooling, whereas only 80% of the previous generation used liquid cooling.
50%UnverifiedThe assembly time of the new generation of Grace Blackwell has been shortened from two hours in the previous generation to five minutes.
50%UnverifiedML4's training was completed in Europe, using about 3,800 NVIDIA Grace Blackwell GPUs.
50%UnverifiedGrace Blackwell is NVIDIA's new-generation GPU platform for large-scale AI training, with Grace responsible for the CPU side and Blackwell for the GPU side, forming a superchip through a high-bandwidth interconnect.
50%UnverifiedLe Chonk's training used 3,800 NVIDIA Grace Blackwell GPUs located in European data centers.
50%UnverifiedMistral Large was trained in European data centers using 3,800 NVIDIA Grace Blackwell chips, with model weights and license to be published by the end of the month.
50%UnverifiedGrace Blackwell is Nvidia's latest-generation chip platform for AI training and inference, with significant improvements in energy efficiency and interconnect bandwidth compared to the previous-generation H100
50%UnverifiedThe Grace Blackwell platform is currently supported primarily through CUDA 12.x and the corresponding cuDNN and NCCL versions, with PyTorch and JAX having optimized kernels for this architecture.
50%UnverifiedThe GB300 is the latest iteration of NVIDIA's Grace Blackwell architecture, with the previous generation being the GB200.
50%UnverifiedThe GB300 superchip is based on NVIDIA's Grace Blackwell architecture, integrating the CPU and GPU via high-bandwidth interconnects to form a unified memory addressing space.
50%UnverifiedThe DGX Spark's Grace Blackwell architecture uses a unified memory pool shared between CPU and GPU, so model weights do not need to be copied between system memory and VRAM, but two large models cannot reside in memory simultaneously.
50%UnverifiedThere is a technical discussion on Hacker News about Grace Blackwell scheduling requirements, which received 42 points.
50%UnverifiedWhen tightly coupled architectures like Grace Blackwell become mainstream, the scheduler's design philosophy needs to shift from "allocating devices" to "orchestrating collaboration".
50%