Session17_Invited Industry.pdf

1、ISSCC 2026SESSION 17Highlighted Chip Releases for AI 2026 IEEEInternational Solid-State Circuits ConferenceNVIDIA GB10:An SoC Built for AI Acceleration A.Skende1,G.Rosseel2,N.Pinckney31NVIDIA,Westford,MA,2NVIDIA,Santa Clara,CA,3NVIDIA,Austin,TX 1 of 13 2026 IEEEInternational Solid-State Circuits Con

2、ferenceOUTLINEOverview of Blackwell ArchitectureScaling Blackwell to create GB10 SoC and DGX SparkDGX Spark capabilities GB10 SoC specifications GB10 scalabilityEngineering Collaboration between NVIDIA and MediaTek Key workloads and use cases accelerated on DGX Spark systemNVIDIA GB10:An SoC Built f

3、or AI Acceleration 2 of 13 2026 IEEEInternational Solid-State Circuits ConferenceGPCGPCGPCGPCGPCGPCGPCGPCL2CACHEL2CACHEL2CACHEL2CACHEHBM CTRLHBM CTRLHBM CTRLHBM CTRLHBM CTRLHBM CTRLHBM CTRLHBM CTRLNVLINKC2CNVLINKMIGMIGNVDECGIGATHREAD ENGINE&FABRICGIGATHREAD ENGINE&FABRICCCCCNVDECNV-HBINV-HBIPCIEGEN6

4、NVSwitchBlackwell Architecture,Scaling BlackwellSame architecture,optimized for developersDGX SparkHigh Bandwidth Interface 10 TB/s Die-to-Die288GB HBM3E Memory(8 Stacks,Up to 8 TB/s)GB300 NVL72 Rack72 Blackwell Ultra GPUsin a coherent domainDatacenter IntegrationSHARPIn NetworkComputeNVLink 51,800

5、GB/sGrace CPUNVLink-C2CUp to 900GB/s Coherent CPU-GPU InterfaceNVFP4 Tensor cores15 PetaFLOPS Dense NVFP4Compute Density4x more TFLOPs/mm2vs HopperNVIDIA Spectrum-X&ConnectX-8 800 Gbps NetworkingNVIDIA GB10:An SoC Built for AI Acceleration 3 of 13 2026 IEEEInternational Solid-State Circuits Conferen

6、ceBuilding a“Mini”AI Supercomputer with BlackwellSmall Form Factor Multi-die Packaging(Advanced 2.5D)Low PowerC2CUnified Memory Architecture(UMA)GB10 AdaptationsDGX SparkInheriting from the DatacenterNVFP4Blackwell ArchGrace-Blackwell Programming ModelCUDANVLink C2CTMEMCX-7NICvLLMSGLangTensorRTNVIDI