2e:null 2f:["$","title",null,{"children":"호기심 해결 in 머신러닝: GPU FP32 학습성능 비교 (pytorch Windows) | Humble Data Miner | Observable"}] 34:T3232,{ const in_pad = 80; const pm = {left:35,top:20,right:30,bottom:10}; const pw = (width-(in_pad*2)-pm.left-pm.right)/3; const ph = pw; const height = pw*2+in_pad+pm.top+pm.bottom; const svg = d3.create("svg").attr("viewBox", [0,0,width,height]); const grid = [ {x:pm.left,y:pm.top,w:pw*2+in_pad,h:ph}, {x:pm.left+pw*2+in_pad*2,y:pm.top,w:pw,h:ph}, {x:pm.left,y:pm.top+ph+in_pad,w:pw,h:ph}, {x:pm.left+pw+in_pad,y:pm.top+ph+in_pad,w:pw,h:ph}, {x:pm.left+pw*2+in_pad*2,y:pm.top+ph+in_pad,w:pw,h:ph}, ]; grid.forEach(c=>{ svg.append("rect") .attr("x",c.x).attr("y",c.y).attr("width",c.w).attr("height",c.h).attr("fill","none").attr("stroke","lightgrey") }); const g_grid = grid.map(c=>{ return svg.append("g").attr("transform",`translate(${c.x},${c.y})`); }); const g_tw = g_grid[0]; // time-watt plot g_tw.append("text") .attr("x",grid[0].w/2).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").text("epoch 100번 학습하는데 걸린시간(s)과 평균전력소모(W)") const tw_x = d3.scaleLinear().domain(ext.ep100).range([0,grid[0].w]); const tw_y = d3.scaleLinear().domain(ext.w).range([grid[0].h,0]); g_tw.append("g").attr("transform",`translate(${0},${ph})`).call(d3.axisBottom(tw_x)) .append("text").attr("x",grid[0].w+2).attr("y",0).attr("text-anchor","start").attr("dominant-baseline","hanging") .attr("fill","black").attr("font-weight","bold").text("s"); g_tw.append("g").attr("transform",`translate(${0},${0})`).call(d3.axisLeft(tw_y)) .append("text").attr("x",0).attr("y",-4).attr("text-anchor","end").attr("dominant-baseline","auto") .attr("fill","black").attr("font-weight","bold").text("W"); g_tw.append("g").selectAll("rect") .data(num) .join("rect") .attr("x",tw_x(0)).attr("y",d=>tw_y(d.gpu_power_consumption)) .attr("width",d=>tw_x(d.epoch100_time)-tw_x(0)).attr("height",d=>tw_y(0)-tw_y(d.gpu_power_consumption)) .attr("fill",d=>d.color).attr("stroke",d=>d.color.darker()).attr("fill-opacity",0.2) .append("title") .text(d=>`${d.device}${d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""} ---\n${d.gpu_power_consumption}W\n${d.epoch100_time}s`); const g_wh = g_grid[1]; // watt-hour bar chart const wh_x = d3.scaleBand().domain(d3.range(num.length)).range([0,grid[1].w]).paddingInner(0.1).paddingOuter(0.1); g_wh.append("g").attr("transform",`translate(${0},${ph})`).call(d3.axisBottom(wh_x).tickSize(0).tickFormat(d=>'')); const wh_y = d3.scaleLinear().domain(ext.wh).range([grid[1].h,0]); g_wh.append("g").attr("transform",`translate(${0},${0})`).call(d3.axisLeft(wh_y)) .append("text").attr("x",0).attr("y",-4).attr("text-anchor","end").attr("dominant-baseline","auto") .attr("fill","black").attr("font-weight","bold").text("Wh"); const is_custom_mode = d=> d.tf32_enabled=="TRUE"||d.xbox_mode=="TRUE"; g_wh.append("g").selectAll("rect") .data(num) .join("rect") .attr("x",d=>wh_x(d.i)).attr("y",d=>wh_y(d.power_in_wh)) .attr("width",d=>wh_x.bandwidth()).attr("height",d=>wh_y(0)-wh_y(d.power_in_wh)) .attr("fill",d=>d.color).attr("stroke","grey") .append("title") .text(d=>`${d.device}${d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""} ---\n${d.power_in_wh.toFixed(2)}Wh`); g_wh.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${wh_x(d.i)+wh_x.bandwidth()/2},${wh_y(0)-4}) rotate(-90)`) .attr("dominant-baseline",d=>is_custom_mode(d)?"auto":"middle").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${d.device}`); g_wh.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${wh_x(d.i)+wh_x.bandwidth()/2+4},${wh_y(0)-4}) rotate(-90)`) .attr("dominant-baseline","hanging").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${is_custom_mode(d)?(d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""):""}`); g_wh.append("text") .attr("x",grid[1].w/2).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").text("같은 학습량에 대한 총 전력소모량(Wh)") const g_em = g_grid[2]; // #epochs per minute const em_x = d3.scaleBand().domain(d3.range(num.length)).range([0,grid[2].w]).paddingInner(0.1).paddingOuter(0.1); g_em.append("g").attr("transform",`translate(${0},${ph})`).call(d3.axisBottom(em_x).tickSize(0).tickFormat(d=>'')); const em_y = d3.scaleLinear().domain(ext.epsmin).range([grid[2].h,0]); g_em.append("g").attr("transform",`translate(${0},${0})`).call(d3.axisLeft(em_y)) .append("text").attr("x",0).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").attr("font-weight","bold").text("#epochs/min"); g_em.append("g").selectAll("rect") .data(num) .join("rect") .attr("x",d=>em_x(d.i)).attr("y",d=>em_y(d.epochs_per_min)) .attr("width",d=>em_x.bandwidth()).attr("height",d=>em_y(0)-em_y(d.epochs_per_min)) .attr("fill",d=>d.color).attr("stroke","grey") .append("title") .text(d=>`${d.device}${d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""} ---\n${d.epochs_per_min.toFixed(2)} epochs/minute`); g_em.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${em_x(d.i)+em_x.bandwidth()/2},${em_y(0)-4}) rotate(-90)`) .attr("dominant-baseline",d=>is_custom_mode(d)?"auto":"middle").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${d.device}`); g_em.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${em_x(d.i)+em_x.bandwidth()/2+4},${em_y(0)-4}) rotate(-90)`) .attr("dominant-baseline","hanging").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${is_custom_mode(d)?(d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""):""}`); g_em.append("text") .attr("x",grid[2].w/2+10).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").text("분당 epoch처리 수 (비용무관 성능)") const g_tc = g_grid[3]; const tc_x = d3.scaleBand().domain(d3.range(num.length)).range([0,grid[3].w]).paddingInner(0.1).paddingOuter(0.1); g_tc.append("g").attr("transform",`translate(${0},${ph})`).call(d3.axisBottom(tc_x).tickSize(0).tickFormat(d=>'')); const tc_y = d3.scaleLinear().domain(ext.tfp32).range([grid[3].h,0]); g_tc.append("g").attr("transform",`translate(${0},${0})`).call(d3.axisLeft(tc_y)) .append("text").attr("x",0).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").attr("font-weight","bold").text("TFLOPS"); g_tc.append("g").selectAll("rect") .data(num) .join("rect") .attr("x",d=>tc_x(d.i)).attr("y",d=>tc_y(d.theoretical_compute_fp32)) .attr("width",d=>tc_x.bandwidth()).attr("height",d=>tc_y(0)-tc_y(d.theoretical_compute_fp32)) .attr("fill",d=>d.color).attr("stroke","grey") .append("title") .text(d=>`${d.device}${d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""} ---\n${d.theoretical_compute_fp32.toFixed(2)} TFLOPS`); g_tc.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${tc_x(d.i)+em_x.bandwidth()/2},${tc_y(0)-4}) rotate(-90)`) .attr("dominant-baseline",d=>is_custom_mode(d)?"auto":"middle").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${d.device}`); g_tc.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${tc_x(d.i)+tc_x.bandwidth()/2+4},${tc_y(0)-4}) rotate(-90)`) .attr("dominant-baseline","hanging").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${is_custom_mode(d)?(d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""):""}`); g_tc.append("text") .attr("x",grid[3].w/2+10).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").text("이론상 FP32 성능 (한계 성능)") const g_pr = g_grid[4]; const pr_x = d3.scaleBand().domain(d3.range(num.length)).range([0,grid[4].w]).paddingInner(0.1).paddingOuter(0.1); g_pr.append("g").attr("transform",`translate(${0},${ph})`).call(d3.axisBottom(pr_x).tickSize(0).tickFormat(d=>'')); const pr_y = d3.scaleLinear().domain(ext.pratio).range([grid[4].h,0]); g_pr.append("g").attr("transform",`translate(${0},${0})`).call(d3.axisLeft(pr_y)) .append("text").attr("x",0).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").attr("font-weight","bold").text("(#epochs/min)/(TFLOPS)"); g_pr.append("g").selectAll("rect") .data(num) .join("rect") .attr("x",d=>pr_x(d.i)).attr("y",d=>pr_y(d.compute_efficiency)) .attr("width",d=>pr_x.bandwidth()).attr("height",d=>pr_y(0)-pr_y(d.compute_efficiency)) .attr("fill",d=>d.color).attr("stroke","grey") .append("title") .text(d=>`${d.device}${d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""} ---\n${d.compute_efficiency.toFixed(2)} (#epochs/min)/(TFLOPS)`); g_pr.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${pr_x(d.i)+em_x.bandwidth()/2},${pr_y(0)-4}) rotate(-90)`) .attr("dominant-baseline",d=>is_custom_mode(d)?"auto":"middle").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${d.device}`); g_pr.append("g").selectAll("text") .data(num) .join("text") .attr("transform",d=>`translate(${pr_x(d.i)+pr_x.bandwidth()/2+4},${pr_y(0)-4}) rotate(-90)`) .attr("dominant-baseline","hanging").attr("font-size",14).attr("font-weight","bold") .attr("fill","white").attr("stroke","black").attr("paint-order","stroke") .text(d=>`${is_custom_mode(d)?(d.tf32_enabled=="TRUE"?" (TF32 precision)":d.xbox_mode=="TRUE"?" (Xbox mode on)":""):""}`); g_pr.append("text") .attr("x",grid[3].w/2+36).attr("y",-4).attr("text-anchor","middle").attr("dominant-baseline","auto") .attr("fill","black").text("한계성능 대비 실제성능 (실효율성)") svg.append("defs").append("marker") .attr("id", "arrowhead") .attr("viewBox", "0 -5 10 10") // Coordinates for the arrowhead .attr("refX", 0) // Adjust to fit line width or node radius .attr("refY", 0) .attr("markerWidth", 6) .attr("markerHeight", 6) .attr("orient", "auto") .append("path") .attr("d", "M0,-5L10,0L0,5") // Path of the triangle .attr("fill", "black"); // Color of the arrow const g_integral = svg.append("g").attr("transform",`translate(${width-pm.right-pw-in_pad/2-10},${ph/2})`); const g_time = svg.append("g").attr("transform",`translate(${pm.left+pw/2},${ph+in_pad/2+14})`); const g_divide = svg.append("g").attr("transform",`translate(${pm.left+pw+in_pad/2-10},${height-ph/2})`); const g_equal = svg.append("g").attr("transform",`translate(${pm.left+pw+in_pad+pw+in_pad/2-10},${height-ph/2})`); function draw_arrow(g,sp,ep,use_marker=true) { const path = g.append("path") .attr("d", `M ${sp[0]} ${sp[1]} L ${ep[0]} ${ep[1]}`) .attr("fill","none").attr("stroke","black").attr("stroke-width",2); if(use_marker) path.attr("marker-end", "url(#arrowhead)"); } draw_arrow(g_integral,[-30,0],[20,0]); draw_arrow(g_time,[0,-30],[0,20]); draw_arrow(g_divide,[-30,0],[30,0],false); draw_arrow(g_equal,[-30,0],[20,0]); function draw_circle(g,text,fs) { const g_sym = g.append("g"); g_sym.append("circle") .attr("r",16).attr("fill","white").attr("stroke","black").attr("stroke-width",2); g_sym.append("text") .attr("dy",+4) .attr("font-size",fs).attr("font-weight","bold") .attr("text-anchor","middle").attr("dominant-baseline","middle").text(text); return g_sym; } draw_circle(g_integral, "∫",26); draw_circle(g_time, "t",26); draw_circle(g_divide, "/",26); draw_circle(g_equal, "=",26); return svg.node(); }35:[{"id":"87d328a957f01e9754063d2fe86253664b6f58887c6de297ee016631f8b42a039ccc42614f1a054e21171d66e708ded2c0e7ec219ee21476cc965d093fcd2d96","href":"https://static.observableusercontent.com/files/87d328a957f01e9754063d2fe86253664b6f58887c6de297ee016631f8b42a039ccc42614f1a054e21171d66e708ded2c0e7ec219ee21476cc965d093fcd2d96","name":"image@4.png","type":"image/png","lastModified":1781755158643.336,"size":3687},{"id":"209cbb26b77d8b1e3ec56fed135a65cbe12f92144b42bbefb8de2750dd5dffacce7bf895289898fd8c93172f9b411cd4e0f7fa15cdb369b9ffaf6f2f525359b2","href":"https://static.observableusercontent.com/files/209cbb26b77d8b1e3ec56fed135a65cbe12f92144b42bbefb8de2750dd5dffacce7bf895289898fd8c93172f9b411cd4e0f7fa15cdb369b9ffaf6f2f525359b2","name":"image@5.png","type":"image/png","lastModified":1781755301950.11,"size":4698},{"id":"9e38ad447b62ab36d28f72d6dc37ef4dbddc6d2578e2f979f26b37d073afec2f4c549e7ade6a57ca537e346700d4671f76025abe64e2e30daa39f3b52fc89633","href":"https://static.observableusercontent.com/files/9e38ad447b62ab36d28f72d6dc37ef4dbddc6d2578e2f979f26b37d073afec2f4c549e7ade6a57ca537e346700d4671f76025abe64e2e30daa39f3b52fc89633","name":"image@3.png","type":"image/png","lastModified":1781755062448.396,"size":11350},{"id":"a1635c964e73b011a3b1156701f3b4e91e5308f4ebff764cefd1b03b114e0aab42b8dc4bd09a821bd63a132894528004f55f5126663dfb08444c7335669166d2","href":"https://static.observableusercontent.com/files/a1635c964e73b011a3b1156701f3b4e91e5308f4ebff764cefd1b03b114e0aab42b8dc4bd09a821bd63a132894528004f55f5126663dfb08444c7335669166d2","name":"RTX3080Ti_RX6800_RyzenZ2E.tsv","type":"text/tab-separated-values","lastModified":1781750317233.039,"size":486}] 32:{"id":"97b7defbcdcc6a40","slug":"gpu-fp32-performance-benchmark-on-pytorch-windows","title":"호기심 해결 in 머신러닝: GPU FP32 학습성능 비교 (pytorch Windows)","thumbnail":"2f74137902fc78858ad045b03591390acd50afedf7ba140656c5a88e3c07ebc4","access_level":"public","create_time":"$D2026-06-18T02:35:34.148Z","update_time":"$D2026-06-19T08:38:46.306Z","view_time":null,"first_public_time":"$D2026-06-18T22:10:43.605Z","has_unpublished_edits":null,"fork":false,"stars":0,"starred":false,"owner":{"type":"team","id":"50faa4da52a9ff9b","login":"humble-data-miner","name":"Humble Data Miner","avatar_url":"https://avatars-next.observableusercontent.com/3046b1bea86c5e825baea7c0c403d90ad9690b5396b11b5d02b3a0839fbcb946","tier":"starter_2024"},"version":526,"publish_version":null,"role":null,"authored":false,"authors":[{"type":"individual","id":"4e2fc41e06361cc3","login":"humble-data-miner","name":"Humble Data Miner","avatar_url":"https://avatars-next.observableusercontent.com/3046b1bea86c5e825baea7c0c403d90ad9690b5396b11b5d02b3a0839fbcb946","tier":null}],"members":[{"type":"individual","id":"4e2fc41e06361cc3","login":"humble-data-miner","name":"Humble Data Miner","avatar_url":"https://avatars-next.observableusercontent.com/3046b1bea86c5e825baea7c0c403d90ad9690b5396b11b5d02b3a0839fbcb946","tier":null,"role":"editor"},{"type":"team","id":"50faa4da52a9ff9b","login":"humble-data-miner","name":"Humble Data Miner","avatar_url":"https://avatars-next.observableusercontent.com/3046b1bea86c5e825baea7c0c403d90ad9690b5396b11b5d02b3a0839fbcb946","tier":"starter_2024","role":"editor"}],"comments":[],"custom_thumbnail":true,"trashed":false,"lapsed":null,"subscribed":null,"forks":0,"collection":{"id":"ffed4a92f8068ba7","slug":"myth-buster-in-person","title":"Myth Buster in Person","owner":{"id":"50faa4da52a9ff9b","login":"humble-data-miner","name":"Humble Data Miner"}},"type":"notebook","body":{"title":"호기심 해결 in 머신러닝: GPU FP32 학습성능 비교 (pytorch Windows)","cells":[{"id":0,"value":"# 호기심 해결 in 머신러닝:
GPU FP32 학습성능 비교 (pytorch Windows)","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":365,"value":"---\n## 동기 (의식의 흐름)\n1. 충동적으로 **Xbox ROG Ally X**를 구입하게 됐고\n2. 매주 주식시장 예측모델을 학습할때 저전력 혹은 학습시간상에 쓸만할까 궁금해짐\n * 메모리가 24GB로 UMPC치고 넉넉한편, 24GB내에서 VRAM 사이즈를 조절가능\n4. 마침 2025년 9월에 AMD의 **PyTorch on Windows (native)** 지원 발표가 있었고\n5. 현시점에 [AMD 공식 PyTorch Windows 설치 가이드](https://rocm.docs.amd.com/projects/radeon-ryzen/en/latest/docs/install/installryz/windows/install-pytorch.html) 따라 문제없이 설치, 성능 측정\n * 성능 비교대상: Geforce RTX 3080 Ti, Radeon RX 6800\n","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":383,"value":"

\n## 제한 사항\n* 단순 sequential layout만을 다룸\n * 이미지나 언어 모델같은 다른 레이아웃은 다루지 않음\n* 학습과만 관련된 성능(시간,전력소모)을 측정\n * 예를들어 **Xbox ROG Ally X**의 APU인 **Ryzen Z2 Extreme**은 개별 NPU를 가지고 있어 추론 과정에 활용 가능해 보이나 여기서 추론 성능은 다루지 않음\n* 성능 비교대상 중의 하나인 **Radeon RX 6800** 은 공식 지원기기에 미포함\n * [rocm nightly build repository](https://rocm.nightlies.amd.com/) 상의 타겟(gfx1030) 실험빌드로 설치 후 성능 측정","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":400,"value":"

\n## 사전 지식\n* TF32(TensorFloat-32)\n * NVIDIA가 FP32계산 부담을 줄이고자 도입한 19비트 실수 포맷\n * Ampere 아키텍처 (Geforce RTX 3000 series) 부터 사용 가능\n * (계산 정확도 손실) 살을 약간 내어주고 (성능 향상) 뼈를 취하는 전략\n * 정확도 손실이 일어나더라도 행렬곱셈과 같은 선형계산의 결과 차이 미미\n * 연산의 내부에서만 TF32를 사용하고 최종결과는 FP32로 변환해 반환\n* Xbox mode\n * Microsoft가 도입한 게임 친화적 Windows 구동 모드\n * 게임기에 가까운 컨트롤러 기반 UI\n * 불필요한 일부 서비스를 중단하고 자원을 반환","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":425,"value":"

\n## 네트워크 레이아웃","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":428,"value":"mermaid`\nflowchart LR\n input[(표본수:2M\\n특성:107차원\\n결과:1차원)]\n Linear1[\"Linear+RELU\\n107×512\"]\n Linear2[\"Linear+RELU\\n512×512\"]\n Linear3[\"Linear+RELU\\n512×512\"]\n Linear4[\"Linear+RELU\\n512×512\"]\n Linear5[\"Linear+RELU\\n512×512\"]\n LinearOut[\"Linear\\n512×1\"]\n Loss[\"Loss function\\nMAE\"]\n Result[\"Estimated\\nrelative\\nyield\"]\n\n LinearOut -- \"2M×1\" --> Result\n input-- \"2M×107\" -->Linear1 -- \"2M×512\" --> Linear2 -- \"2M×512\" --> Linear3 -- \"2M×512\" --> Linear4 -- \"2M×512\" --> Linear5 -- \"2M×512\" --> LinearOut -- \"2M×1\" --> Loss\n input-- \"2M×1\" --> Loss\n`;","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":481,"value":"

\n## 실험 대상","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":20,"value":"{\n const lh = 32;\n return Inputs.table(num,{\n columns:[\"vendor\",\"device\",\"code\",\"theoretical_compute_fp32\",\"torch_version\",\"custom_setting\"],\n header:{vendor:\"제조사\", device:\"GPU 장치 이름\", code:\"AMD gfx 코드\", theoretical_compute_fp32:\"이론상 FP32 성능\", torch_version:\"pytorch 버전\",custom_setting:\"설정의 특이사항\"},\n format:{\n vendor:(d,i)=>d==\"NVIDIA\" ? html`${d3.select(nvidia_logo).attr(\"height\",lh).clone(true).remove().node()}`:\n d==\"AMD\" && num[i].device.startsWith(\"RX\") ? html`${d3.select(radeon_logo).attr(\"height\",lh).clone(true).remove().node()}`:\n d==\"AMD\" && num[i].device.startsWith(\"Ryzen\") ? html`${d3.select(ryzen_logo).attr(\"height\",lh).clone(true).remove().node()}` :\"\",\n device: (d,i)=>html` ${d}`,\n theoretical_compute_fp32:x=>`${x.toFixed(2)} TFLOPS`,\n },\n width:{vendor:55,device:160,code:90,theoretical_compute_fp32:100,torch_version:200},\n maxWidth:900\n });\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":483,"value":"

\n## 실험 결과","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":128,"value":"$34","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":485,"value":"

\n## 결과 요약\n* **Ryzen Z2 Extreme**의 기대이하의 학습능력\n * 전력소비량 2배 이상, 학습시간의 실효율이 절반 이하\n * 대부분의 일반 학습용도로 부적합\n * 다만 미리 학습된 모델에 대량으로 추론하는 경우 NPU의 활용 가능성이 핵심\n * 추정되는 성능저하의 요인\n * 전용 VRAM의 부재 (시스템 메모리를 떼어쓰는 형태)\n * 35W로 제한되어있는 전력사용량과 협소한 쿨링 시스템\n* Xbox 모드의 성능향상 없음\n * 학습에 필요한 모든 데이터를 device에 올려놓고 쓰기 때문에 큰 의미 없음\n* 순수 FP32성능만으로는 **RX 6800**이 전력소모량이나 실효율면에서 우위\n * 앞으로 RoCM의 지속적인 개선을 기대\n * RX 6000 series이상의 기기를 보유중인 경우 활용을 시도해볼만함","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":525,"value":"---\n## 부연설명 동영상\n\n","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":22,"value":"---\n## Appendix","pinned":false,"hidden":false,"mode":"md","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":24,"value":"","pinned":false,"hidden":false,"mode":"html","author":"$undefined","summary":"$undefined","output":"","database":"$undefined"},{"id":16,"value":"raw_data = FileAttachment(\"RTX3080Ti_RX6800_RyzenZ2E.tsv\").tsv({typed:true})","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":27,"value":"num = {\n const data = JSON.parse(JSON.stringify(raw_data));\n data.forEach((d,i)=>{\n d.epoch_in_sec = d.epoch100_time/100;\n d.epochs_per_min = 60/d.epoch_in_sec;\n d.power_in_wh = d.epoch100_time*d.gpu_power_consumption/3600;\n d.compute_efficiency = d.epochs_per_min/d.theoretical_compute_fp32;\n d.custom_setting = d.tf32_enabled == \"TRUE\" ? \"FP32 정확도 = TF32\": d.xbox_mode == \"TRUE\" ? \"XBOX 모드 적용\":\"\";\n d.color = d.vendor == \"NVIDIA\" ? (d.tf32_enabled == \"TRUE\" ? d3.color(\"green\"):d3.color(\"green\").brighter(1.6)) :\n d.device.startsWith(\"RX\") ? d3.color(\"crimson\") : (d.xbox_mode == \"TRUE\" ? d3.color(\"orange\") : d3.color(\"orange\").brighter());\n d.i = i;\n });\n return data;\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":143,"value":"ext = {\n return {\n ep100:[0,d3.max(num.map(d=>d.epoch100_time))],\n w:[0,d3.max(num.map(d=>d.gpu_power_consumption))],\n epsmin:[0,d3.max(num.map(d=>d.epochs_per_min))],\n wh:[0,d3.max(num.map(d=>d.power_in_wh))],\n tfp32:[0,d3.max(num.map(d=>d.theoretical_compute_fp32))],\n pratio:[0,d3.max(num.map(d=>d.compute_efficiency))]\n };\n}","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":66,"value":"radeon_logo = FileAttachment(\"image@3.png\").image()","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":78,"value":"ryzen_logo = FileAttachment(\"image@5.png\").image()","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"},{"id":82,"value":"nvidia_logo = FileAttachment(\"image@4.png\").image()","pinned":false,"hidden":false,"mode":"ojs","author":"$undefined","summary":"$undefined","output":"$undefined","database":"$undefined"}],"files":"$W35","stdlib":"1","titling":"auto"}} 33:"$32:body"