- [[Nvidia]]の[[SoC]] # スペック - アーキテクチャ: [[Grace Blackwell]] - ただしCPUはGraceではなくMediaTek - TSMC 3nm - [[Unified Memory]]: 128GB - バス幅: 256bit - 帯域幅: 273.2GB/s - SKhynix H58G78CK8B x 8 chip - 動作速度: 8533 MT/s - 1 chipあたり2ch. 合計16ch. 1chあたり16bitで、合計$16 \times 2 \times 8=256$ bitのバス幅 - CPU: 20Core ARM (Cortex-X925 x 10, Cortex-A725 x 10) - GPU: - [[CUDA]] Core: 6144 - [[Streaming Multiprocessors]]: 48 - Tensor Core: 192 - 周波数: 1665 MHz (Boost: 2418 MHz) ### cudaGetDevicePropertiesの結果 ``` Name : NVIDIA GB10 Compute Capability : 12.1 SM count : 48 Total Global Mem : 121.6 GB Shared Mem / Block : 48 KB Shared Mem / SM : 100 KB L2 Cache Size : 24 MB Warp Size : 32 Max Threads / Block : 1024 Max Threads / SM : 1536 Memory Bus Width : 256 bit Unified Addressing : 1 Integrated : 1 Concurrent Kernels : 1 Regs per Block : 65536 Regs per SM : 65536 ``` - https://docs.nvidia.com/dgx/dgx-spark/hardware.html - https://hc2025.hotchips.org/assets/program/conference/day2/21_nvidia_skende_final.pdf - https://www.chiplog.io/p/analysis-of-nvidia-dgx-sparks-gb10 - https://www.chargerlab.com/teardown-of-nvidia-dgx-spark-4tb/ - https://chipsandcheese.com/p/inside-nvidia-gb10s-memory-subsystem - https://chipsandcheese.com/p/analyzing-nvidia-gb10s-gpu