Unverified50% confidenceReleaseExact time
Hanxu Technology has released a new architecture that keeps model weights resident in memory chips, attempting to bypass the HBM bandwidth bottleneck
1
Sources
50%
Confidence
Medium-term (~90 days)
Relevance
9/11/2026
First Seen
Valid until: 12/10/2026
Sources
Related Entities
Related Claims
UnverifiedHuawei's Ascend chips are based on its self-developed DaVinci architecture, with the 910D targeting performance on par with NVIDIA's H100.69% similarUnverifiedMoE架构的稀疏激活特性适合定制化芯片针对性优化片上内存容量与带宽分配、专家路由调度及稀疏矩阵乘法单元69% similarUnverifiedHBM vertically integrates multiple layers of memory chips through 3D stacking technology, achieving bandwidth 5 to 10 times that of conventional DDR5, and is a core component of AI accelerators such as NVIDIA's H100/H200.69% similarVerifiedGroq's LPU and Cerebras's WSE are chips optimized specifically for inference, breaking through bandwidth bottlenecks through customized memory architectures66% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/905201API
curl https://kongchang.com/api/v1/knowledge/claims/905201MCP
get_claim(id=905201)