返回博客稿

Strata:恢复性能寄存器后与解锁端对照

恢复端使用本卡最初日志中的 SS0=0x02200115、SS1=0x4;PLM仍开放。FP16/BF16矩阵吞吐已恢复原基线,六组完整输出哈希与原始未解锁端一致。本次衡量性能选择寄存器恢复前后的速度差,恢复端并非完整物理冷启动。

同一IQ2_XS、Strata0.1.43/同引擎SHA、同350W、PCIe Gen3×16、同输入及专家画像;关闭前缀/对话缓存、尾部跳过和CPU prefill分担。每种输入/请求组合两轮;配置间速度差与解锁收益分别计算。

Prefill

每输入只生成1 token。吞吐列为引擎prompt阶段中位数;提速=解锁/恢复−1,TTFT缩短=1−解锁/恢复。

配置 输入 token 恢复 token/s 解锁 token/s 提速 恢复 TTFT s 解锁 TTFT s TTFT缩短
pf2048 1007 769.6 798.0 +3.7% 1.330 1.283 +3.5%
pf2048 4089 1299.2 1330.4 +2.4% 3.171 3.098 +2.3%
pf2048 16401 1297.3 1333.4 +2.8% 12.675 12.334 +2.7%
pf2048 28667 1315.9 1355.9 +3.0% 21.824 21.181 +2.9%
pf18432 1007 767.0 797.0 +3.9% 1.334 1.285 +3.7%
pf18432 4089 1829.0 1993.3 +9.0% 2.259 2.075 +8.1%
pf18432 16401 2565.9 2922.5 +13.9% 6.425 5.646 +12.1%
pf18432 28667 2543.4 2883.2 +13.4% 11.311 9.983 +11.7%
pf32768 1007 763.5 791.6 +3.7% 1.341 1.294 +3.5%
pf32768 4089 1823.8 1981.2 +8.6% 2.267 2.088 +7.9%
pf32768 16401 2571.9 2917.9 +13.4% 6.414 5.656 +11.8%
pf32768 28667 2690.0 3047.5 +13.3% 10.697 9.448 +11.7%
pf32768-fp16 1007 715.6 753.4 +5.3% 1.429 1.358 +4.9%
pf32768-fp16 4089 1569.9 1749.8 +11.5% 2.629 2.361 +10.2%
pf32768-fp16 16401 2300.2 2738.8 +19.1% 7.164 6.022 +15.9%
pf32768-fp16 28667 2479.6 2961.1 +19.4% 11.602 9.725 +16.2%

28K prefill 同配置对照

Decode与并发

每批同样四个请求,各输出1024 token;同类/混合各两批。整批吞吐含准入和客户端排队;TPOT为每请求(E2E−TTFT)/(N−1)的中位数。中央解码窗口另见summary.json。

配置 组合 恢复整批 token/s 解锁整批 token/s 吞吐提速 恢复 TPOT ms 解锁 TPOT ms TPOT缩短
decode-solo shared 130.03 121.92 -6.2% 6.93 7.28 -5.0%
decode-solo mixed 78.81 75.32 -4.4% 11.57 12.30 -6.4%
decode-solo-spec8 shared 111.58 111.76 +0.2% 8.03 7.89 +1.8%
decode-solo-spec8 mixed 74.88 73.01 -2.5% 11.91 11.27 +5.3%
decode-solo-fullcache shared 131.98 130.59 -1.1% 6.78 6.85 -1.0%
decode-solo-fullcache mixed 89.63 86.94 -3.0% 9.79 10.09 -3.1%
decode-solo-fullcache-spec8 shared 116.06 123.32 +6.3% 7.74 7.61 +1.7%
decode-solo-fullcache-spec8 mixed 83.20 81.69 -1.8% 10.67 10.61 +0.6%
decode-two shared 100.17 102.31 +2.1% 19.50 18.84 +3.4%
decode-two mixed 68.60 65.31 -4.8% 28.05 28.95 -3.2%
decode-two-mtp shared 109.18 111.75 +2.4% 17.28 17.39 -0.7%
decode-two-mtp mixed 74.07 74.63 +0.8% 26.45 24.40 +7.8%
decode-four shared 127.55 128.40 +0.7% 29.61 29.30 +1.1%
decode-four mixed 74.07 72.22 -2.5% 52.07 53.53 -2.8%
decode-four-mtp shared 122.05 123.01 +0.8% 30.94 30.93 +0.0%
decode-four-mtp mixed 75.89 73.77 -2.8% 48.58 49.92 -2.8%

Decode 同配置对照

观测与范围

固定缓存decode各组实际11546个native槽、15843MiB;单流auto缓存另列。18432 MMQ两端均有6072/6072个借出槽在RAM中备份,23.16GiB专家内存全锁页;FP16 32768两端均9126/9126、27.24GiB全锁页。MMQ auto:32768借出10652槽,解锁端覆盖10598、恢复端10592,RAM边缘存在6槽差异,因此以18432组作为更干净的大分块对照;raw文件保留。两端按先后顺序运行,重复次数有限,几个百分点的差异需结合范围和资源采样解释。

逐请求CSV、恢复端资源窗口、解锁端资源窗口。