Unverified50% confidenceFactExact time
OjasX packs ReLU/ReLU2 states into 1-bit hardware ballot masks (32 times smaller than FP32), reconstructing smooth activations such as SiLU/GELU directly in registers, enabling GEMM without backward recomputation.
1
Sources
50%
Confidence
Long-term
Relevance
10/14/2026
First Seen
Sources
Related Entities
Related Claims
Unverified自Compute Capability 2.0（Fermi架构）起，GPU硬件引入了更灵活的内存合并机制，即使线程访问顺序打乱，只要地址落在同一128字节缓存行内仍可通过一次事务完成62% similarUnverifiedPR #642 fixes the ARM64 2B batched step size memory overwrite, FFN activation alignment (SiLU vs ReLU²), and tokenizer inference defects.61% similarUnverifiedx86-64 处理器提供 MUL 指令可将两个64位整数相乘产生128位结果，在GCC和Clang中通过 __uint128_t 类型暴露；ARM AArch64 提供 UMULH 指令获取乘法的高64位结果60% similarUnverifiedThe tester used a standard ComfyUI workflow configured with the ref2va int8 pruned base model, 1.2MP resolution, 20 steps, on a single RTX 5090.60% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/1025339API
curl https://kongchang.com/api/v1/knowledge/claims/1025339MCP
get_claim(id=1025339)