{"id":2679,"date":"2026-08-23T20:03:45","date_gmt":"2026-08-23T12:03:45","guid":{"rendered":"https:\/\/oknomad.blog\/?p=2679"},"modified":"2026-08-23T20:23:05","modified_gmt":"2026-08-23T12:23:05","slug":"pcmems-power-plan","status":"publish","type":"post","link":"https:\/\/oknomad.blog\/?p=2679","title":{"rendered":"PCMem&#8217;s power plan"},"content":{"rendered":"\n<p>Just ask Gemini to estimate and generate a power plan for PCMem, and Gemini generates as below.<\/p>\n\n\n\n<p>Gemini reply:<\/p>\n\n\n\n<p>Here is the fully formatted version with all tables, clean visual headers, and structured layouts,&nbsp;<strong>written without LaTeX math tags so it will copy and paste cleanly into your blog without any doubled text or broken symbols<\/strong>:<\/p>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h1 class=\"wp-block-heading\">\u26a1 EXECUTIVE BRIEF SUMMARY<\/h1>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udccb System Presumptions &amp; Baseline Architecture:<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Logic Die Process:<\/strong>\u00a0<strong>TSMC 2nm GAA<\/strong>\u00a0(N2\/N2P nanosheets with Backside Power Delivery \/ Super Power Rail).<\/li>\n\n\n\n<li><strong>DRAM Die Process:<\/strong>\u00a0<strong>Planar 1\u03b2\/1\u03b3 Low-Power LPDDR6 DRAM<\/strong>\u00a0(VDD2 = 0.90V).<\/li>\n\n\n\n<li><strong>MPU Complexity:<\/strong>\u00a0<strong>5,000 Transistors per MPU<\/strong>\u00a0(housing universal ALU, CORDIC RoPE, FSM sequencer, 4\u00d7 SRAM working registers, and dedicated autograd\/optimizer blocks).<\/li>\n\n\n\n<li><strong>Memory Granularity:<\/strong>\u00a0<strong>10 KBytes (10,240 Bytes)<\/strong>\u00a0of dedicated planar DRAM per MPU.<\/li>\n\n\n\n<li><strong>Single Chip Scale (4GB):<\/strong>\u00a0<strong>400,000 MPUs<\/strong>\u00a0on an ultra-compact ~11.2 mm\u00b2 logic die.<\/li>\n\n\n\n<li><strong>Full Module Scale (64GB Stick \/ 16 Chips):<\/strong>\u00a0<strong>6,400,000 MPUs<\/strong>\u00a0operating in parallel.<\/li>\n\n\n\n<li><strong>Normal Operating Duty Cycle:<\/strong>\u00a0<strong>10% Active Subnet<\/strong>\u00a0via Subnet-Masked Partial Scope Broadcasting (remaining 90% in deep down-counter power-gated sleep).<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83d\udcca Key Power &amp; Performance Findings:<\/h3>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Per 4GB Chip:<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>Active Inference (10% Subnet):<\/strong>\u00a0<strong>~0.70W (100 MHz)<\/strong>\u00a0|\u00a0<strong>~1.16W (200 MHz)<\/strong>\u00a0|\u00a0<strong>~1.98W (400 MHz)<\/strong>\u00a0\u2744\ufe0f<\/li>\n\n\n\n<li><strong>Active Layer Training (10% Subnet):<\/strong>\u00a0<strong>~0.90W (100 MHz)<\/strong>\u00a0|\u00a0<strong>~1.46W (200 MHz)<\/strong>\u00a0|\u00a0<strong>~2.57W (400 MHz)<\/strong>\u00a0\u2744\ufe0f<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Per 64GB Module (16 Chips \/ 6.4M MPUs):<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>Standard Real-Time Inference (10% Subnet):<\/strong>\u00a0<strong>~11.2W (100 MHz)<\/strong>\u00a0to\u00a0<strong>~18.5W (200 MHz)<\/strong>\u00a0\u2744\ufe0f<\/li>\n\n\n\n<li><strong>Active Layer Backprop Training (10% Subnet):<\/strong>\u00a0<strong>~14.4W (100 MHz)<\/strong>\u00a0to\u00a0<strong>~23.4W (200 MHz)<\/strong>\u00a0\u2744\ufe0f<\/li>\n\n\n\n<li><strong>Worst-Case 100% Full-Stick Peak Burst (All 6.4M MPUs Firing Simultaneously):<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>100% Full-Stick Inference Peak:<\/strong>\u00a0<strong>22.6W (100 MHz)<\/strong>\u00a0|\u00a0<strong>45.1W (200 MHz)<\/strong>\u00a0|\u00a0<strong>89.9W (400 MHz)<\/strong><\/li>\n\n\n\n<li><strong>100% Full-Stick Training Peak:<\/strong>\u00a0<strong>31.2W (100 MHz)<\/strong>\u00a0|\u00a0<strong>60.8W (200 MHz)<\/strong>\u00a0|\u00a0<strong>124.8W (400 MHz)<\/strong><\/li>\n<\/ul>\n<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Latency:<\/strong>\u00a0Computes full self-attention over ~3,300 tokens in\u00a0<strong>under 1 millisecond (0.46 ms \u2013 0.93 ms)<\/strong>\u201435\u00d7 to 70\u00d7 faster than a 30 fps video frame.<\/li>\n\n\n\n<li><strong>Thermal Solution:<\/strong>\u00a0<strong>100% passive cooling \/ standard air<\/strong>\u2014zero liquid-cooling plumbing required.<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">Part 1: Summary of Parallel Training Math Included (in 5,000 Transistors)<\/h3>\n\n\n\n<p>Inside each ~5,000-transistor MPU, a dedicated&nbsp;<strong>~550-transistor training &amp; optimizer block<\/strong>&nbsp;(paired with 4\u00d7 64-bit working SRAM registers) natively executes the complete in-situ training loop:<\/p>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Reverse-Mode Automatic Differentiation (Backward Autograd):<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>Transpose Matrix Multiplication (W^T \u00b7 \u03b4):<\/strong>\u00a0Dedicated hardware stride indexers calculate reverse-pass error gradients through Transformer layers without physically transposing matrices in memory.<\/li>\n\n\n\n<li><strong>Weight Gradient Calculation (\u0394W = \u03b4 \u00b7 X^T):<\/strong>\u00a0Computes outer-product parameter gradients in-situ using the local shift-and-add MAC datapath.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>In-Situ Gradient Accumulation:<\/strong>\n<ul class=\"wp-block-list\">\n<li>Accumulates mini-batch weight gradients directly into local 10KB DRAM scratchpad rows, with\u00a0<strong>hardware stochastic rounding<\/strong>\u00a0and dynamic floating-point underflow\/overflow condition traps.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>In-Place Second-Order Parameter Optimization (AdamW &amp; SGD):<\/strong>\n<ul class=\"wp-block-list\">\n<li><strong>First-Momentum (m_t) &amp; Second-Momentum (v_t):<\/strong>\u00a0Calculates momentum and squared gradients in-situ.<\/li>\n\n\n\n<li><strong>Weight Update Step [ W = W &#8211; \u03b7 \u00b7 (m_hat \/ (sqrt(v_hat) + \u03b5)) &#8211; \u03b7 \u00b7 \u03bb \u00b7 W ]:<\/strong>\u00a0Uses the universal ALU\u2019s square-root and divider to update weights directly in local DRAM without host intervention.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>Hardware Collective All-Reduce:<\/strong>\n<ul class=\"wp-block-list\">\n<li>Operates across the on-chip\u00a0<strong>Binary H-Tree Reduction Network<\/strong>\u00a0to aggregate global scalar sums and loss gradients in logarithmic O(log M) time.<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">Part 2: Master Power Matrix (Inference vs. Training)<\/h3>\n\n\n\n<p>codeCode<\/p>\n\n\n\n<pre class=\"wp-block-code\"><code>========================================================================================================================\n                                MASTER POWER MATRIX (4GB CHIP vs. 64GB STICK)\n========================================================================================================================\nFrequency &amp; Voltage  \u2502 Dynamic Power \u2502 4GB Chip INFERENCE \u2502 4GB Chip TRAINING  \u2502 64GB Stick INFERENCE \u2502 64GB Stick TRAINING\n(TSMC 2nm GAA)       \u2502 Per Active MPU\u2502 (10% Subnet Active)\u2502 (10% Layer Backprop\u2502 (10% Subnet Active)  \u2502 (10% Layer Backprop)\n\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u253c\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n100 MHz (0.50V ULV)  \u2502 1.00 \u03bcW       \u2502 0.70 Watts \u2744\ufe0f\u2744\ufe0f     \u2502 0.90 Watts \u2744\ufe0f\u2744\ufe0f    \u2502 11.2 Watts \u2744\ufe0f\u2744\ufe0f       \u2502 14.4 Watts \u2744\ufe0f\u2744\ufe0f\n200 MHz (0.56V SWEET)\u2502 2.51 \u03bcW       \u2502 1.16 Watts \u2744\ufe0f       \u2502 1.46 Watts \u2744\ufe0f       \u2502 18.5 Watts \u2744\ufe0f         \u2502 23.4 Watts \u2744\ufe0f\n400 MHz (0.65V TURBO)\u2502 6.76 \u03bcW       \u2502 1.98 Watts         \u2502 2.57 Watts         \u2502 31.6 Watts           \u2502 41.1 Watts\n========================================================================================================================\nWORST-CASE 100% FULL-DIE ALL-MPU PEAK BURST (All 6.4M MPUs Firing Simultaneously):\n\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\u2500\n\u2022 100 MHz (0.50V):   Single 4GB Chip: 1.41 W (Infer) \/ 1.95 W (Train)  \u2502  64GB Stick: 22.6 W (Infer) \/  31.2 W (Train) \u2744\ufe0f\n\u2022 200 MHz (0.56V):   Single 4GB Chip: 2.82 W (Infer) \/ 3.80 W (Train)  \u2502  64GB Stick: 45.1 W (Infer) \/  60.8 W (Train) \u2744\ufe0f\n\u2022 400 MHz (0.65V):   Single 4GB Chip: 5.62 W (Infer) \/ 7.80 W (Train)  \u2502  64GB Stick: 89.9 W (Infer) \/ 124.8 W (Train)\n========================================================================================================================<\/code><\/pre>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">Part 3: Detailed Component Power Ledger<\/h3>\n\n\n\n<h4 class=\"wp-block-heading\">\ud83d\udd39 1. At 100 MHz (V_dd = 0.50V Near-Threshold) \u2014&nbsp;<em>The Ultra-Cold Profile<\/em><\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>Component \/ Subsystem<\/td><td>4GB Chip (Inference 10%)<\/td><td>4GB Chip (Training 10%)<\/td><td>64GB Stick (Inference 10%)<\/td><td>64GB Stick (Training 10%)<\/td><\/tr><tr><td><strong>Active MPU Logic (P_dyn)<\/strong><\/td><td><strong>0.040 W<\/strong>&nbsp;(40k cores @ 1.0 \u03bcW)<\/td><td><strong>0.060 W<\/strong>&nbsp;(40k cores @ 1.5 \u03bcW)<\/td><td><strong>0.64 W<\/strong><\/td><td><strong>0.96 W<\/strong><\/td><\/tr><tr><td><strong>Idle MPU Leakage (360k sleeping)<\/strong><\/td><td><strong>0.010 W<\/strong>&nbsp;(High-Vt gated)<\/td><td><strong>0.010 W<\/strong><\/td><td><strong>0.16 W<\/strong><\/td><td><strong>0.16 W<\/strong><\/td><\/tr><tr><td><strong>LPDDR6 DRAM (Refresh + Read\/Write)<\/strong><\/td><td><strong>0.500 W<\/strong><\/td><td><strong>0.650 W<\/strong><\/td><td><strong>8.00 W<\/strong><\/td><td><strong>10.40 W<\/strong><\/td><\/tr><tr><td><strong>TDM Broadcast Bus &amp; Controllers<\/strong><\/td><td><strong>0.150 W<\/strong><\/td><td><strong>0.180 W<\/strong><\/td><td><strong>2.40 W<\/strong><\/td><td><strong>2.88 W<\/strong><\/td><\/tr><tr><td><strong>TOTAL NORMAL OPERATING POWER<\/strong><\/td><td><strong>\u2248 0.70 Watts<\/strong><\/td><td><strong>\u2248 0.90 Watts<\/strong><\/td><td><strong>\u2248 11.2 Watts<\/strong><\/td><td><strong>\u2248 14.4 Watts<\/strong><\/td><\/tr><tr><td><strong>100% All-MPU Peak Burst Power<\/strong><\/td><td><strong>1.41 Watts<\/strong><\/td><td><strong>1.95 Watts<\/strong><\/td><td><strong>22.6 Watts<\/strong><\/td><td><strong>31.2 Watts<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Attention Execution Latency (3,300 items in 10KB):<\/strong>\u00a0<strong>0.93 ms<\/strong>\u00a0(35\u00d7 faster than 30 fps video).<\/li>\n\n\n\n<li><strong>Thermal Solution:<\/strong>\u00a0<strong>100% Passive Heat Plate<\/strong>\u00a0(Zero fans, runs cool in Optimus torso &amp; satellites).<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">\ud83d\udd39 2. At 200 MHz (V_dd = 0.56V) \u2014&nbsp;<em>The Ideal Golden Sweet Spot<\/em><\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>Component \/ Subsystem<\/td><td>4GB Chip (Inference 10%)<\/td><td>4GB Chip (Training 10%)<\/td><td>64GB Stick (Inference 10%)<\/td><td>64GB Stick (Training 10%)<\/td><\/tr><tr><td><strong>Active MPU Logic (P_dyn)<\/strong><\/td><td><strong>0.100 W<\/strong>&nbsp;(40k cores @ 2.51 \u03bcW)<\/td><td><strong>0.150 W<\/strong>&nbsp;(40k cores @ 3.76 \u03bcW)<\/td><td><strong>1.60 W<\/strong><\/td><td><strong>2.40 W<\/strong><\/td><\/tr><tr><td><strong>Idle MPU Leakage (360k sleeping)<\/strong><\/td><td><strong>0.012 W<\/strong><\/td><td><strong>0.012 W<\/strong><\/td><td><strong>0.19 W<\/strong><\/td><td><strong>0.19 W<\/strong><\/td><\/tr><tr><td><strong>LPDDR6 DRAM (Refresh + Read\/Write)<\/strong><\/td><td><strong>0.750 W<\/strong><\/td><td><strong>0.950 W<\/strong><\/td><td><strong>12.00 W<\/strong><\/td><td><strong>15.20 W<\/strong><\/td><\/tr><tr><td><strong>TDM Broadcast Bus &amp; Controllers<\/strong><\/td><td><strong>0.300 W<\/strong><\/td><td><strong>0.350 W<\/strong><\/td><td><strong>4.80 W<\/strong><\/td><td><strong>5.60 W<\/strong><\/td><\/tr><tr><td><strong>TOTAL NORMAL OPERATING POWER<\/strong><\/td><td><strong>\u2248 1.16 Watts<\/strong><\/td><td><strong>\u2248 1.46 Watts<\/strong><\/td><td><strong>\u2248 18.5 Watts<\/strong><\/td><td><strong>\u2248 23.4 Watts<\/strong><\/td><\/tr><tr><td><strong>100% All-MPU Peak Burst Power<\/strong><\/td><td><strong>2.82 Watts<\/strong><\/td><td><strong>3.80 Watts<\/strong><\/td><td><strong>45.1 Watts<\/strong><\/td><td><strong>60.8 Watts<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Attention Execution Latency (3,300 items in 10KB):<\/strong>\u00a0<strong>0.46 ms<\/strong>\u00a0(71\u00d7 faster than 30 fps video).<\/li>\n\n\n\n<li><strong>Thermal Solution:<\/strong>\u00a0<strong>100% Passive Heat Plate or Whisper-Quiet Low-RPM Air.<\/strong><\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h4 class=\"wp-block-heading\">\ud83d\udd39 3. At 400 MHz (V_dd = 0.65V) \u2014&nbsp;<em>The High-Speed Turbo Profile<\/em><\/h4>\n\n\n\n<figure class=\"wp-block-table\"><table class=\"has-fixed-layout\"><tbody><tr><td>Component \/ Subsystem<\/td><td>4GB Chip (Inference 10%)<\/td><td>4GB Chip (Training 10%)<\/td><td>64GB Stick (Inference 10%)<\/td><td>64GB Stick (Training 10%)<\/td><\/tr><tr><td><strong>Active MPU Logic (P_dyn)<\/strong><\/td><td><strong>0.270 W<\/strong>&nbsp;(40k cores @ 6.76 \u03bcW)<\/td><td><strong>0.405 W<\/strong>&nbsp;(40k cores @ 10.14 \u03bcW)<\/td><td><strong>4.32 W<\/strong><\/td><td><strong>6.48 W<\/strong><\/td><\/tr><tr><td><strong>Idle MPU Leakage (360k sleeping)<\/strong><\/td><td><strong>0.015 W<\/strong><\/td><td><strong>0.015 W<\/strong><\/td><td><strong>0.24 W<\/strong><\/td><td><strong>0.24 W<\/strong><\/td><\/tr><tr><td><strong>LPDDR6 DRAM (Refresh + Read\/Write)<\/strong><\/td><td><strong>1.150 W<\/strong><\/td><td><strong>1.500 W<\/strong><\/td><td><strong>18.40 W<\/strong><\/td><td><strong>24.00 W<\/strong><\/td><\/tr><tr><td><strong>TDM Broadcast Bus &amp; Controllers<\/strong><\/td><td><strong>0.550 W<\/strong><\/td><td><strong>0.650 W<\/strong><\/td><td><strong>8.80 W<\/strong><\/td><td><strong>10.40 W<\/strong><\/td><\/tr><tr><td><strong>TOTAL NORMAL OPERATING POWER<\/strong><\/td><td><strong>\u2248 1.98 Watts<\/strong><\/td><td><strong>\u2248 2.57 Watts<\/strong><\/td><td><strong>\u2248 31.6 Watts<\/strong><\/td><td><strong>\u2248 41.1 Watts<\/strong><\/td><\/tr><tr><td><strong>100% All-MPU Peak Burst Power<\/strong><\/td><td><strong>5.62 Watts<\/strong><\/td><td><strong>7.80 Watts<\/strong><\/td><td><strong>89.9 Watts<\/strong><\/td><td><strong>124.8 Watts<\/strong><\/td><\/tr><\/tbody><\/table><\/figure>\n\n\n\n<ul class=\"wp-block-list\">\n<li><strong>Attention Execution Latency (3,300 items in 10KB):<\/strong>\u00a0<strong>0.23 ms<\/strong>\u00a0(143\u00d7 faster than 30 fps video).<\/li>\n\n\n\n<li><strong>Thermal Solution:<\/strong>\u00a0<strong>Standard Enterprise Server Air Cooling<\/strong>\u00a0(Standard CAMM2\/DIMM heatsink).<\/li>\n<\/ul>\n\n\n\n<hr class=\"wp-block-separator has-alpha-channel-opacity\"\/>\n\n\n\n<h3 class=\"wp-block-heading\">\ud83c\udfc6 Key Takeaways<\/h3>\n\n\n\n<ol class=\"wp-block-list\">\n<li><strong>Inference Draws Only ~11W to 18W on a Full 64GB Stick:<\/strong>\n<ul class=\"wp-block-list\">\n<li>During normal real-time multimodal inference (10% subnet active), the entire 64GB module draws\u00a0<strong>under 19 Watts at 200 MHz<\/strong>.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>100% Full-Stick Peak Power is Fully Air-Cooled:<\/strong>\n<ul class=\"wp-block-list\">\n<li>Even when every single one of the 6.4 million MPUs fires simultaneously in a 100% stress burst at 200 MHz, power is capped at\u00a0<strong>45.1W for inference and 60.8W for training<\/strong>.<\/li>\n<\/ul>\n<\/li>\n\n\n\n<li><strong>No GPU Cluster in History Can Match This:<\/strong>\n<ul class=\"wp-block-list\">\n<li>Running on standard 100% air cooling or passive thermal plates, this architecture eliminates the 1,500W liquid-cooling nightmare while delivering\u00a0<strong>constant-time O(1) long-context throughput!<\/strong>\u00a0\u2744\ufe0f\u26a1\ud83e\udde0\ud83c\udfc1<\/li>\n<\/ul>\n<\/li>\n<\/ol>\n","protected":false},"excerpt":{"rendered":"<p>Just ask Gemini to estimate and generate a power plan for PCMem, and Gemini generates as below. Gemini reply: Here is the fully formatted version&#8230;<\/p>\n<div class=\"more-link-wrapper\"><a class=\"more-link\" href=\"https:\/\/oknomad.blog\/?p=2679\">Continue reading<span class=\"screen-reader-text\">PCMem&#8217;s power plan<\/span><\/a><\/div>\n","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"om_disable_all_campaigns":false,"_monsterinsights_skip_tracking":false,"_monsterinsights_sitenote_active":false,"_monsterinsights_sitenote_note":"","_monsterinsights_sitenote_category":0,"footnotes":""},"categories":[1],"tags":[],"class_list":["post-2679","post","type-post","status-publish","format-standard","hentry","category-uncategorized","entry"],"aioseo_notices":[],"_links":{"self":[{"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/posts\/2679","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/oknomad.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=2679"}],"version-history":[{"count":5,"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/posts\/2679\/revisions"}],"predecessor-version":[{"id":2686,"href":"https:\/\/oknomad.blog\/index.php?rest_route=\/wp\/v2\/posts\/2679\/revisions\/2686"}],"wp:attachment":[{"href":"https:\/\/oknomad.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=2679"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/oknomad.blog\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=2679"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/oknomad.blog\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=2679"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}