CANN/hixl A3芯片性能数据
HIXL在昇腾A3芯片上部分场景实测性能数据【免费下载链接】hixlHIXLHuawei Xfer Library是一个灵活、高效的昇腾单边通信库面向集群场景提供简单、可靠、高效的点对点数据传输能力。项目地址: https://gitcode.com/cann/hixlHIXL在昇腾A3芯片上存在如下约束条件Atlas A3 训练/推理系列产品该场景下采用HCCS传输协议时不支持Host内存作为远端Cache开启中转内存池时无此限制。所以未给出D2H/H2H场景下采用HCCS传输协议时的实测性能数据。单机场景 (CANN 9.0.0WRITE:传输内存块大小HCCS D2D (GB/s)HCCS D2D BufferPool(GB/s)RDMA D2D(GB/s)RDMA D2D BufferPool(GB/s)16K6.9166.4697.1147.03632K16.05012.5965.7005.52564K28.88222.78110.1269.795128K45.63738.26116.39315.966256K66.66758.22121.54821.051512K81.96776.87621.84221.5961M84.17581.69921.39721.2372M93.07591.30821.53321.4744M110.132110.42422.20222.1878M113.020114.78422.35322.202传输内存块大小HCCS H2D(GB/s)HCCS H2D BufferPool(GB/s)RDMA H2D(GB/s)RDMA H2D BufferPool(GB/s)16K5.2772.7277.0372.65932K11.3095.3475.6835.30864K20.10910.65410.0758.904128K24.60118.96816.41111.379256K27.60624.10821.36013.823512K29.64933.66521.86514.2811M29.86933.65621.35315.0382M30.72033.69321.41115.2184M32.19233.41422.21015.0318M32.31634.05122.26614.471传输内存块大小HCCS D2H(GB/s)HCCS D2H BufferPool(GB/s)RDMA D2H(GB/s)RDMA D2H BufferPool(GB/s)16K——3.2926.7143.16432K——7.9385.6697.60864K——12.30310.09412.951128K——23.24716.35914.516256K——24.75221.51115.214512K——26.50821.81915.2741M——26.65821.23015.4262M——26.61821.33115.5764M——27.25722.17115.4388M——26.37122.27815.371传输内存块大小HCCS H2H(GB/s)HCCS H2H BufferPool(GB/s)RDMA H2H(GB/s)RDMA H2H BufferPool(GB/s)16K——2.5266.1282.36732K——5.3005.7025.06264K——10.31910.1189.386128K——17.47816.46714.760256K——24.72821.51515.218512K——25.23221.86415.3831M——25.82121.37115.3792M——26.08521.54115.3154M——26.14522.22215.3758M——25.63622.31715.419READ:传输内存块大小HCCS D2D(GB/s)HCCS D2D BufferPool(GB/s)RDMA D2D(GB/s)RDMA D2D BufferPool(GB/s)16K7.0756.6506.6655.76832K17.33213.2005.6095.45664K30.55524.2629.9719.694128K49.55440.90315.20615.932256K74.27263.80821.33820.987512K94.98585.09221.70121.6641M95.12990.97521.32721.2512M107.388103.47721.36421.3894M131.441125.88122.13622.1518M134.553134.98922.23022.159传输内存块大小HCCS H2D(GB/s)HCCS H2D BufferPool(GB/s)RDMA H2D(GB/s)RDMA H2D BufferPool(GB/s)16K5.2342.2486.1812.33932K11.6524.4635.6204.26564K19.9688.4999.9497.312128K24.31417.27516.28710.318256K27.75921.59621.18311.774512K30.27425.29321.76914.3451M30.71824.81121.30915.4002M31.64626.96321.40015.6234M33.40530.76522.11615.7118M33.69327.31622.15514.760传输内存块大小HCCS D2H(GB/s)HCCS D2H BufferPool(GB/s)RDMA D2H(GB/s)RDMA D2H BufferPool(GB/s)16K——2.9746.6163.46732K——5.7545.6127.96264K——11.5789.92414.689128K——23.17816.20215.565256K——35.41121.26616.079512K——37.74221.72016.2321M——39.16021.32416.3592M——39.35821.48516.3984M——39.67022.12416.3618M——37.98222.12016.308传输内存块大小HCCS H2H(GB/s)HCCS H2H BufferPool(GB/s)RDMA H2H(GB/s)RDMA H2H BufferPool(GB/s)16K——2.9336.5642.16232K——6.7215.5964.67064K——14.0269.9729.658128K——19.48016.21315.194256K——21.84921.38915.704512K——24.35721.72015.9011M——22.78921.30616.0172M——23.01621.92616.1004M——23.26422.13616.1198M——22.67822.20216.069双机场景 (CANN 9.0.0WRITE传输内存块大小HCCS D2D(GB/s)HCCS D2D BufferPool(GB/s)RDMA D2D(GB/s)RDMA D2D BufferPool(GB/s)16K4.0333.8495.7537.18132K8.8228.4825.6895.55164K16.60715.75310.0939.850128K28.84827.92716.38116.156256K46.99245.22421.50721.226512K64.96964.86821.78521.7281M74.85073.79021.32721.2772M86.80685.49921.39721.4594M106.112106.11222.14722.1998M113.020110.42422.28622.305传输内存块大小HCCS H2D(GB/s)HCCS H2D BufferPool(GB/s)RDMA H2D(GB/s)RDMA H2D BufferPool(GB/s)16K3.4062.6906.1592.64832K7.1465.2565.6855.01964K13.32910.40910.0677.738128K15.06018.82616.40210.194256K16.23023.41321.40011.391512K16.89632.64721.79212.8601M16.98033.08621.35713.3192M17.32533.36921.53314.0864M17.98633.87522.20213.5698M18.09532.86922.24212.974传输内存块大小HCCS D2H(GB/s)HCCS D2H BufferPool(GB/s)RDMA D2H(GB/s)RDMA D2H BufferPool(GB/s)16K——3.6006.0183.45232K——8.7185.6917.63064K——15.86310.09613.230128K——23.91416.46314.384256K——25.75721.42214.795512K——26.33821.88814.8321M——27.57621.40415.0952M——27.60621.55215.0264M——27.82122.21015.1948M——27.34622.31715.075传输内存块大小HCCS H2H(GB/s)HCCS H2H BufferPool(GB/s)RDMA H2H(GB/s)RDMA H2H BufferPool(GB/s)16K——2.4126.4672.58632K——4.7935.7035.08464K——9.36310.1129.694128K——16.49116.37214.715256K——25.54721.47015.332512K——26.02021.85315.4781M——26.32721.35315.4762M——26.56221.47815.5514M——26.50022.22215.5848M——26.35522.29815.543READ传输内存块大小HCCS D2D(GB/s)HCCS D2D BufferPool(GB/s)RDMA D2D(GB/s)RDMA D2D BufferPool(GB/s)16K4.1423.9396.8346.06432K9.1768.7895.6075.54864K17.13316.1759.9479.508128K29.84028.87516.13715.650256K49.74147.91121.21220.750512K71.26670.70121.56321.4781M83.33382.39921.15821.1112M97.50497.50421.34921.2984M124,008123.39621.94122.0818M133.120130.48022.06922.187传输内存块大小HCCS H2D(GB/s)HCCS H2D BufferPool(GB/s)RDMA H2D(GB/s)RDMA H2D BufferPool(GB/s)16K3.4292.8636.4322.49932K7.3066.4015.6055.23764K13.26711.1899.9038.161128K15.23818.85116.16710.909256K16.68221.81921.13311.925512K17.48721.30221.66814.4731M17.54423.77321.12615.3052M17.84724.78721.25515.6454M18.38524.65022.07315.6768M18.45821.70122.15114.498传输内存块大小HCCS D2H(GB/s)HCCS D2H BufferPool(GB/s)RDMA D2H(GB/s)RDMA D2H BufferPool(GB/s)16K——2.8386.2712.72432K——5.8445.6075.28564K——10.3519.94410.485128K——16.65316.21715.235256K——30.33221.26616.133512K——37.50421.72016.5041M——37.99421.14316.4732M——38.12121.39316.6404M——38.37922.03816.5108M——37.81022.07316.554传输内存块大小HCCS H2H(GB/s)HCCS H2H BufferPool(GB/s)RDMA H2H(GB/s)RDMA H2H BufferPool(GB/s)16K——2.9146.2341.97632K——6.8725.5974.13864K——12.3339.9508.247128K——20.50516.18315.332256K——20.66121.19016.028512K——26.26121.66416.2551M——21.99921.20416.3462M——26.83021.37816.4824M——22.55922.03016.4568M——21.74722.19116.342【免费下载链接】hixlHIXLHuawei Xfer Library是一个灵活、高效的昇腾单边通信库面向集群场景提供简单、可靠、高效的点对点数据传输能力。项目地址: https://gitcode.com/cann/hixl创作声明:本文部分内容由AI辅助生成(AIGC),仅供参考
本文来自互联网用户投稿,该文观点仅代表作者本人,不代表本站立场。本站仅提供信息存储空间服务,不拥有所有权,不承担相关法律责任。如若转载,请注明出处:http://www.coloradmin.cn/o/2599304.html
如若内容造成侵权/违法违规/事实不符,请联系多彩编程网进行投诉反馈,一经查实,立即删除!