CN
重新定义 AI 推理基础设施
GPNPU架构融合通用编程与专用效率，打造面向大模型推理时代的新一代AI推理芯片
构建面向下一代AI推理的
全新智算底座
'%3e%3cpath%20stroke-width='1.092'%20d='M35.3%2038h-26c-1.7%200-3.1-1.4-3.1-3.1V9.5c0-1.7%201.4-3.1%203.1-3.1h26c1.7%200%203.1%201.4%203.1%203.1v25.4c0%201.7-1.4%203.1-3.1%203.1Z'/%3e%3cpath%20d='M29.4%2015.2H15.2v13.9h14.2V15.2Z'/%3e%3cpath%20stroke-width='1.092'%20d='M13.7%200v6.2M22.3%200v6.2M30.9%200v6.2M13.7%2037.5v6.2m8.6-6.2v6.2m8.6-6.2v6.2m7.5-12.9h6.2m-6.2-8.7h6.2m-6.2-8.6h6.2M0%2030.8h6.2M0%2022.1h6.2M0%2013.5h6.2'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h44.6v43.7H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
高性能计算'%3e%3cpath%20d='M20.6.6%201.2%2010.8l19.9%2010.5%2019.8-10.8L20.6.6Z'/%3e%3cpath%20d='M40.2%2014.6v16.6L20.7%2042.1V21.3M1.2%2013.7V32l17.2%209.5'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h42.1v43.1H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
灵活编程'%3e%3cpath%20d='M18.9.9%201.1%2010.1c-.8.4-.8%201.6%200%202l17.8%208.4c.9.4%201.9.4%202.8%200l18.2-8.8c.8-.4.8-1.6%200-2L21.8.9c-.9-.5-2-.5-2.9%200ZM7.4%2023.7%201.1%2027c-.8.4-.8%201.6%200%202l17.8%208.4c.9.4%201.9.4%202.8%200l18.2-8.8c.8-.4.8-1.6%200-2l-7-3.4'/%3e%3cpath%20d='m7.4%2015.1-6.3%203.3c-.8.4-.8%201.6%200%202l17.8%208.4c.9.4%201.9.4%202.8%200L39.9%2020c.8-.4.8-1.6%200-2l-7-3.4'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h41.1v38.2H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
高能效比'%3e%3cpath%20d='M27.1%2024c0%203.8-3.1%206.9-6.9%206.9-3.8%200-6.9-3.1-6.9-6.9%200-3.8%203.1-6.9%206.9-6.9%203.8%200%206.9%203.1%206.9%206.9ZM24.4%204.8c0%202.3-1.9%204.2-4.2%204.2C17.9%209%2016%207.1%2016%204.8%2016%202.5%2017.9.6%2020.2.6c2.3%200%204.2%201.8%204.2%204.2ZM9%2034.3c0%202.3-1.9%204.2-4.2%204.2-2.3%200-4.2-1.9-4.2-4.2%200-2.3%201.9-4.2%204.2-4.2%202.3%200%204.2%201.8%204.2%204.2Zm31.1%200c0%202.3-1.9%204.2-4.2%204.2-2.3%200-4.2-1.9-4.2-4.2%200-2.3%201.9-4.2%204.2-4.2%202.3%200%204.2%201.8%204.2%204.2Zm-25.8-6.4L8.5%2032m17.7-4.1L32%2032M20.2%209v8.1'/%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h40.6v39H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
大模型优化
Analysis
架构解析
.
'%3e%3cmask%20id='b'%20width='85'%20height='85'%20x='0'%20y='0'%20maskUnits='userSpaceOnUse'%20style='mask-type:luminance'%3e%3cpath%20fill='%23fff'%20d='M.3%2042.2c0%2023.1%2018.8%2041.9%2041.9%2041.9%2023.1%200%2041.9-18.8%2041.9-41.9C84.1%2019.1%2065.3.3%2042.2.3%2019%20.3.3%2019%20.3%2042.2Z'/%3e%3c/mask%3e%3cg%20mask='url(%23b)'%3e%3cpath%20fill='url(%23c)'%20d='m63.1-36%2057.3%2099.1-99.2%2057.3L-36%2021.2%2063.1-36Z'/%3e%3c/g%3e%3cmask%20id='d'%20width='40'%20height='40'%20x='22'%20y='22'%20maskUnits='userSpaceOnUse'%20style='mask-type:luminance'%3e%3cpath%20fill='%23fff'%20d='M47.1%2050.2h11.4v8.5H47.1v-8.5Zm-1.5-3.1c-.8%200-1.5.7-1.5%201.5v11.5c0%20.8.7%201.6%201.5%201.6H60c.8%200%201.6-.7%201.6-1.6V48.6c0-.8-.7-1.5-1.6-1.5H45.6Zm-19.8-4h11.4v15.5H25.8V43.1ZM24.3%2040c-.8%200-1.6.7-1.6%201.6v18.5c0%20.8.7%201.6%201.6%201.6h14.4c.8%200%201.6-.7%201.6-1.6V41.6c0-.8-.7-1.6-1.6-1.6H24.3Zm1.5-14.2h11.4v8.5H25.8v-8.5Zm-1.5-3.1c-.8%200-1.6.7-1.6%201.5v11.5c0%20.8.7%201.6%201.6%201.6h14.4c.8%200%201.6-.7%201.6-1.6V24.3c0-.8-.7-1.5-1.6-1.5H24.3v-.1Zm22.8%203.1h11.4v15.5H47.1V25.8Zm-1.5-3.1c-.8%200-1.5.7-1.5%201.5v18.5c0%20.8.7%201.6%201.5%201.6H60c.8%200%201.6-.7%201.6-1.6V24.3c0-.8-.7-1.5-1.6-1.5H45.6v-.1Z'/%3e%3c/mask%3e%3cg%20mask='url(%23d)'%3e%3cpath%20fill='url(%23e)'%20d='M61.6%2022.7H22.7v38.9h38.9V22.7Z'/%3e%3c/g%3e%3c/g%3e%3cdefs%3e%3clinearGradient%20id='c'%20x1='21.122'%20x2='63.025'%20y1='5.966'%20y2='78.544'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23D3B0A1'/%3e%3cstop%20offset='.001'%20stop-color='%23D3B0A1'/%3e%3cstop%20offset='.784'%20stop-color='%23DFC6BB'/%3e%3cstop%20offset='.998'%20stop-color='%23E3CDC3'/%3e%3cstop%20offset='1'%20stop-color='%23E3CDC3'/%3e%3c/linearGradient%3e%3clinearGradient%20id='e'%20x1='22.736'%20x2='61.621'%20y1='42.199'%20y2='42.199'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23fff'/%3e%3cstop%20offset='.664'%20stop-color='%23FEFDFD'/%3e%3cstop%20offset='.906'%20stop-color='%23F9F6F7'/%3e%3cstop%20offset='.999'%20stop-color='%23F5F1F2'/%3e%3cstop%20offset='1'%20stop-color='%23F5F1F2'/%3e%3c/linearGradient%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h84v84H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
GPGPU
通用性
SIMT 通用编程能力
AI主流框架兼容，一行代码实现程序运行
'%3e%3cmask%20id='b'%20width='85'%20height='85'%20x='-1'%20y='-1'%20maskUnits='userSpaceOnUse'%20style='mask-type:luminance'%3e%3cpath%20fill='%23fff'%20d='M-.3%2041.6c0%2023.1%2018.8%2041.9%2041.9%2041.9%2023.1%200%2041.9-18.8%2041.9-41.9C83.5%2018.5%2064.7-.3%2041.6-.3%2018.4-.3-.3%2018.5-.3%2041.6Z'/%3e%3c/mask%3e%3cg%20mask='url(%23b)'%3e%3cpath%20fill='url(%23c)'%20d='m62.498-36.6%2057.2%2099.1-99.1%2057.3-57.2-99.2%2099.1-57.2Z'/%3e%3c/g%3e%3cmask%20id='d'%20width='1920'%20height='6169'%20x='-562'%20y='-1550'%20maskUnits='userSpaceOnUse'%20style='mask-type:luminance'%3e%3cpath%20fill='%23fff'%20d='M1358-1550H-562v6168.7h1920V-1550Z'/%3e%3c/mask%3e%3cg%20mask='url(%23d)'%3e%3cpath%20fill='%23fff'%20d='M40.7%2063.7c-.4%200-.9-.2-1.2-.5-.5-.5-.6-1.2-.3-1.8%202.3-5%201.2-10.9-2.8-14.8-2.6-2.5-6-3.9-9.6-3.9-1.9%200-3.8.4-5.6%201.2-.6.3-1.4.2-1.9-.3s-.6-1.2-.3-1.8c.9-1.8%202-3.5%203.5-4.9%203.2-3.1%207.5-4.9%2012-4.8h.7c2.7-4.7%207-8.3%2012-10.3%202.9-1.1%206-1.7%209.1-1.7%201.9%200%203.8.2%205.6.6.6.1%201.1.6%201.2%201.2%201.9%208-.5%2016.3-6.5%2022-1.6%201.5-3.4%202.8-5.3%203.9.3%206.8-3.7%2013.1-10%2015.8-.1.1-.3.1-.6.1Zm-14-24.2c4.5%200%208.8%201.7%2012%204.8%202.5%202.4%204.1%205.5%204.8%208.9.3%201.7.3%203.5.1%205.2.2-.1.3-.3.5-.4%202.9-2.8%204.4-6.8%203.9-10.9-.1-.7.3-1.4%201-1.6%202-1%203.8-2.2%205.4-3.8%204.8-4.6%207-11.4%205.7-17.9-1.3-.2-2.5-.3-3.8-.3-5.5%200-10.7%202.1-14.7%205.9-1.6%201.5-2.9%203.3-3.9%205.2-.3.6-1%201-1.7.9-.5%200-1-.1-1.5-.1-3.6%200-7.1%201.4-9.6%203.9l-.4.4c.7-.2%201.4-.3%202.2-.2Zm-2.3%2021.9c-.5%200-1%200-1.6-.1-.7-.1-1.3-.7-1.4-1.4%200-.2-.9-6.3%203.4-10.5%202-2%203.7-2.8%205.4-2.8%202%200%203.4%201.4%204.2%202.1.7.7%202.1%202%202.2%203.9.1%201.7-.9%203.4-2.9%205.4-3.1%202.9-7.1%203.4-9.3%203.4Zm5.8-11.5c-.4%200-1.3.2-3%201.9-1.3%201.3-2.2%203.1-2.4%205-.1.5-.1.9-.1%201.4%201.7%200%204.5-.4%206.5-2.4%201.7-1.7%202-2.6%201.9-3-.2-.7-.6-1.3-1.2-1.7-.7-1-1.2-1.2-1.7-1.2Zm21.4-13.7c-2.7-.1-4.9-2.4-4.8-5.1.1-2.7%202.4-4.9%205.1-4.8%202.7.1%204.8%202.3%204.8%205%200%202.8-2.3%205-5.1%204.9Zm0-6.6c-.9%200-1.7.7-1.7%201.6%200%20.9.8%201.6%201.7%201.6.9%200%201.7-.7%201.7-1.6%200-.8-.8-1.6-1.7-1.6Z'/%3e%3c/g%3e%3c/g%3e%3cdefs%3e%3clinearGradient%20id='c'%20x1='20.485'%20x2='62.388'%20y1='5.394'%20y2='77.973'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23B4BADD'/%3e%3cstop%20offset='.001'%20stop-color='%23B4BADD'/%3e%3cstop%20offset='.999'%20stop-color='%23C7D1E4'/%3e%3cstop%20offset='1'%20stop-color='%23C7D1E4'/%3e%3c/linearGradient%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h85v84H0z'/%3e%3c/clipPath%3e%3c/defs%3e%3c/svg%3e)
NPU
高效性
SIMD 矩阵计算密度
NPU级别的计算效率和能效比
近存计算
大带宽、低延时
3D堆叠架构
高效访存架构
10ns级别访存延迟
'%3e%3cmask%20id='b'%20width='85'%20height='85'%20x='-1'%20y='-1'%20maskUnits='userSpaceOnUse'%20style='mask-type:luminance'%3e%3cpath%20fill='%23fff'%20d='M-.3%2041.6c0%2023.1%2018.8%2041.9%2041.9%2041.9%2023.1%200%2041.9-18.8%2041.9-41.9C83.5%2018.5%2064.7-.3%2041.6-.3%2018.4-.3-.3%2018.5-.3%2041.6Z'/%3e%3c/mask%3e%3cg%20mask='url(%23b)'%3e%3cpath%20fill='url(%23c)'%20d='m62.498-36.6%2057.2%2099.1-99.1%2057.3-57.2-99.2%2099.1-57.2Z'/%3e%3c/g%3e%3cg%20fill='%23fff'%20clip-path='url(%23d)'%3e%3cpath%20d='m50.8%2047.2-7.2%204.1v-8.2l7.2-4.3v8.4Zm-8.4-6.4L35%2036.5l7.3-4.2%207.3%204.2-7.2%204.3Zm-1.3%202.3v8.4l-7.4-4.3v-8.3l7.4%204.2ZM26%2051.5c-.2-.3-.6-.4-.9-.3l-1.1.7c-.1.1-.2.2-.2.3V33.1l7.3%204.2V48H31c-.3.2-.4.6-.2.9.1.2.3.3.6.3.1%200%20.2%200%20.3-.1l.1-.1%209.2%205.4v8.4L24%2053c.2.1.4.1.6%200l1.1-.6c.4-.2.5-.6.3-.9Zm15.8-30.4c.1.1.3.2.5.2s.3-.1.4-.2l16.6%209.6-7.1%204.3-9.3-5.4c-.4-.2-.9-.2-1.3%200L32.3%2035%2025%2030.8l16.8-9.7Zm21.6%2017v-7.4c0-.5-.3-.9-.7-1.1L42.9%2018.2c-.4-.2-.9-.2-1.3%200L21.8%2029.6c-.6.3-.9.9-.7%201.5v22.5c0%20.5.3.9.7%201.1l19.8%2011.5h.1c.1.1.3.1.5.1h.1c.2%200%20.4-.1.6-.2l19.9-11.4c.4-.2.7-.7.7-1.1l-.1-15.5h-2.6v14.2c0-.1-.1-.2-.2-.3l-1.1-.7c-.3-.2-.7-.2-.9.1-.2.3-.2.7.1.9.1.1.1.1.2.1l1.1.7c.1.1.2.1.3.1l-16.6%209.6v-8.4l9.2-5.3.2.1c.3.3.7.2.9%200%20.3-.3.2-.7%200-.9-.1-.1-.2-.1-.3-.2l-.3-.1V37.3l7.3-4.3.1%205.2h2.6v-.1Z'/%3e%3cpath%20d='M42.4%2031.3c.3%200%20.6-.3.6-.6v-1.3c0-.3-.3-.6-.6-.6s-.6.3-.6.6v1.2c-.1.3.2.6.6.7Zm0%203.7c.3%200%20.6-.3.6-.6v-1.2c0-.3-.3-.6-.6-.6s-.6.3-.6.6v1.2c-.1.3.2.6.6.6Zm12%2019.1c-.2.3-.1.7.2.8l1%20.6c.3.2.7.1.9-.2.1-.1.1-.3.1-.5s-.1-.3-.3-.4l-1-.6c-.4-.1-.8%200-.9.3Zm-25.7%201.5c.1%200%20.2%200%20.3-.1l1-.6c.3-.2.4-.6.2-.9-.1-.1-.2-.2-.4-.3-.2%200-.3%200-.5.1l-1%20.6c-.3%200-.6.3-.6.7%200%20.3.3.5.6.6.2%200%20.3-.1.4-.1Z'/%3e%3c/g%3e%3c/g%3e%3cdefs%3e%3cclipPath%20id='a'%3e%3cpath%20fill='%23fff'%20d='M0%200h85v84H0z'/%3e%3c/clipPath%3e%3cclipPath%20id='d'%3e%3cpath%20fill='%23fff'%20d='M0%200h42.4v48.4H0z'%20transform='translate(21%2018)'/%3e%3c/clipPath%3e%3clinearGradient%20id='c'%20x1='20.485'%20x2='62.388'%20y1='5.394'%20y2='77.973'%20gradientUnits='userSpaceOnUse'%3e%3cstop%20stop-color='%23B4BADD'/%3e%3cstop%20offset='.001'%20stop-color='%23B4BADD'/%3e%3cstop%20offset='.999'%20stop-color='%23C7D1E4'/%3e%3cstop%20offset='1'%20stop-color='%23C7D1E4'/%3e%3c/linearGradient%3e%3c/defs%3e%3c/svg%3e)
算力积木
超节点
Scale up 算力扩展
满足10万亿级别的 MoE大模型推理需要
通用性
统一线程执行，真正承接主流开发生态
以SIMT通用执行模型为底座，兼容AI主流框架与软件生态，支持复杂控制流、动态调度与完整
推理流程承载。
高效性
通用架构也能实现 NPU 级效率
通过 SIMD、张量融合计算、混合精度执行与片上算存协同，提升有效算力、吞吐表现与单位功耗产出。
大宽带
以3D堆叠重构数据供给架构
通过3D堆叠存储与分层带宽池化，提升数据带宽与访问效率，支撑长上下文与MoE等高数据吞吐场景。
低延时
面向实时智能的响应式架构
依托近核热数据驻留、低跳数互连与解码优先调度，显著改善TTFT、TPOT与尾时延表现。
超节点
从单芯片走向超节点推理平台
通过统一内存语义的高速互连与系统级扩展架构，支撑从单卡、整机到超节点的持续扩展。
DOMESTIC TECHNOLOGY
IFWA软件栈
