
ããŒã«ã«LLMã®é床ã¯ã¡ã¢ãªåž¯åã§æ±ºãŸããã¹ããã¯è¡šã®æ£ããèªã¿æ¹
ãªãé«ããã·ã³ã§ãããŒã«ã«LLMãé ãããšãããã®ããé床ã¯ãã¡ã¢ãªåž¯å÷ã¢ãã«ãµã€ãºããšãã1æ¬ã®åŒã§ã»ãŒäºæž¬ã§ããŸãã
ããŒãã®ååã䞊ã¹ãæ¯èŒèšäºã¯å€ãã®ã«ããã®åçããæžããèšäºã¯ã»ãšãã©ãããŸãããåŒã1æ¬èŠããã ãã§ãã©ã®ãã·ã³ã§ã©ã®ã¢ãã«ãã©ããããã®é床ã§åãããè²·ãåã«èªåã§èšç®ã§ããããã«ãªããŸãã
ãéããã·ã³ãã®çŽæãå€ããçç±
ããŒã«ã«LLMçéã§ããããèœèããããŸãã
128GBçµ±åã¡ã¢ãªã®ææ°ãã·ã³ãè²·ãã70Bã¢ãã«ãåããããæ¯ç§æ°ããŒã¯ã³ããåºãªã
CPUãæäžäœã«ããã®ã«ãLLMã®çæé床ãã»ãŒå€ãããªã
ãAIæ§èœâ¯TOPSãããããPCã§ãOllamaãæåŸ å€ãã«é ã
PCã®éãã®çŽæ(CPUäžä»£ãã³ã¢æ°ãã¯ããã¯ãTOPS)ã¯ãLLMæšè«ã«ã¯ã»ãŒéçšããŸãããLLMã«ã¯åºæã®åŸéèŠå ãããããã§ãã
LLMæšè«ã®2ã€ã®ãã§ãŒãº
LLMã®æšè«ã¯ãæ§è³ªã®éã2ã€ã®ãã§ãŒãºã§ã§ããŠããŸãã
ããªãã£ã«(å ¥åã®èªã¿èŸŒã¿): æããããã³ããå šäœãäžæ°ã«åŠçãããèšç®å(GPUã®æŒç®æ§èœ)ãå¹ããé·ãææžã貌ãä»ãããšãã®ãèããŠããæéãã¯ãã
ãã³ãŒã(ããŒã¯ã³çæ): 1ããŒã¯ã³ãã€æç« ãçæãããäœæé床ã¯ãã¡ã
ãã³ãŒãã«ã¯éèŠãªæ§é ããããŸãã1ããŒã¯ã³çæãããã³ã«ãã¢ãã«ã®å šãã©ã¡ãŒã¿ãã¡ã¢ãªããèªã¿åºãã®ã§ãã20GBã®ã¢ãã«ãªãã1ããŒã¯ã³ããšã«20GBã®èªã¿åºããçºçããŸããèšç®ãã®ãã®ã¯è»œããããŒã¿ãéã¶äœæ¥ãæ¯é çã«ãªããŸãã
ã ãããã³ãŒãé床ã¯ãæŒç®æ§èœã§ã¯ãªããã¡ã¢ãªããããŒã¿ãéã¶éãã=ã¡ã¢ãªåž¯åã§é æã¡ã«ãªããŸãã
é床ãäºæž¬ããåŒ
ãã³ãŒãéåºŠã¯æ¬¡ã®åŒã§è¿äŒŒã§ããŸãã
æ¯ç§ããŒã¯ã³æ° â ã¡ã¢ãªåž¯å(GB/s) ÷ ã¢ãã«å æãµã€ãº(GB)
èšç®äŸã3ã€æããŸãã
RTX 4090(垯å1,008GB/s)ã§20GBã®32Bã¢ãã« â çŽ50ããŒã¯ã³/ç§
DGX Spark(垯å273GB/s)ã§43GBã®70Bã¢ãã« â çŽ6ããŒã¯ã³/ç§
ããŒãPC(垯å135GB/s)ã§6.6GBã®9Bã¢ãã« â çŽ20ããŒã¯ã³/ç§
å®éã«ã¯KVãã£ãã·ã¥ã®èªã¿åºããªã©ãå ãããããå®å¹å€ã¯çè«å€ã®7ã8å²çšåºŠã«èœã¡ãŸããããã§ããæ¡ãå€ããªãã粟床ã§äºæž¬ã§ããè³Œå ¥å€æã«ã¯ååã§ãã
åé ã®èœèäŸã®çš®æããããã®åŒã§ãã128GBçµ±åã¡ã¢ãªæ©ã¯ã容éãã¯ç«æŽŸã§ãã垯åãã273GB/sçšåºŠãããªãããã70Bã®é«å¯åºŠã¢ãã«ã§ã¯æ¯ç§4ã5ããŒã¯ã³ã«ãããªããŸããã
äž»èŠããŒãã®ã¡ã¢ãªåž¯åäžèЧ(2026幎7æ)
ãã®åŒã«ä»£å ¥ããããã®åž¯åäžèЧã§ããã«ã¿ãã°ã®ãã¡ã¢ãªåž¯åå¹ ãã®è¡ã ãèŠãã°ããã·ã³ã®LLM驿§ã¯èªããŸãã

èŠç¹ãæç« ã§ãæŒãããŸããã³ã³ã·ã¥ãŒãGPUã®äžäœã¯1TB/sçŽãAppleã®äžäœãããã¯500ã800GB/sçŽã話é¡ã®çµ±åã¡ã¢ãªæ©(DGX Sparkã»Strix Halo)ã¯250ã270GB/sçŽãæ®éã®ããŒãPCãšCPUå®è¡ã¯90ã140GB/sçŽããã®4ã€ã®éçŽãèŠããŠããã°ãã¹ããã¯è¡šãèŠãç¬éã«ããããã®éåºŠãæµ®ãã³ãŸãã
ãªããã©ã®ã¢ãã«ãã¡ã¢ãªã«èŒãããã¯å®¹éã®åé¡ãªã®ã§ãå¥éã[2026幎7ææŽæ°] ããŒã«ã«LLMå¿ èŠVRAMæ©èŠè¡šãããªãã®PCã§åãã¢ãã«ããšæãåãããŠãã ããã容éã§ãèŒãããã垯åã§ãéãããããã®2軞ã§ããŒãéžã³ã¯å®çµããŸãã
MoEãé床åé¡ãå€ãã
2026幎ã®ã¢ãã«äºæ ã«ã¯ããã®åŒã®èªã¿æ¿ãã1ã€å¿ èŠã§ããäž»æµã«ãªã£ãMoE(Mixture of Experts)åã¢ãã«ã¯ãç·ãã©ã¡ãŒã¿ã®ãã¡äžéš(掻æ§ãã©ã¡ãŒã¿)ã ããæ¯ããŒã¯ã³èªã¿åºããŸãã
MoEã®é床 â ã¡ã¢ãªåž¯å ÷ 掻æ§ãã©ã¡ãŒã¿åã®ãµã€ãº
gpt-oss 120B(ç·117Bã»æŽ»æ§5.1B): ã¡ã¢ãªã¯65GBå¿ èŠã ããæ¯ããŒã¯ã³ã®èªã¿åºãã¯æ°GBãDGX Sparkã§ãæ¯ç§30ããŒã¯ã³çŽãçãã
Qwen3.5 122B-A10B(掻æ§10B)ã35B-A3B(掻æ§3B)ãåãæ§é
ãããã2026å¹Žã®æ°ããããŒãéžã³ã®å®ç³ãçãŸããŸããã倧容éã»äœåž¯åã®ãã·ã³(DGX SparkãStrix Haloã128GB Mac)ã¯ãé«å¯åºŠ70Bã§ã¯ãªã100Bè¶ MoEãåããããã®ç®±ãšèãããšãæ§èœãæå€§éåŒãåºããŸããéã«é«å¯åºŠã¢ãã«çããªãã容éãã垯å(=GPU)ã«æè³ãã¹ãã§ãã
å®è·µ: è³Œå ¥åãã§ãã¯ã®4ã¹ããã
æ©æãè²·ãåã«ããã®é çªã§èšç®ããŠãã ããã
䜿ãããã¢ãã«ã®å æãµã€ãºã確èªãã(éåååŸã®å®ãµã€ãºãVRAMæ©èŠè¡šåç §)
åè£ãã·ã³ã®ã¡ã¢ãªåž¯åãã«ã¿ãã°ã§ç¢ºèªãã
垯å÷ãµã€ãºã§çè«é床ãåºãã0.7ã0.8ãæãã
ãã®é床ãçšéã«å¯ŸããŠèš±å®¹ã§ããã倿ãã
蚱容ã©ã€ã³ã®ç®å®ã¯çšéã«ãã£ãŠå€ãããŸãã
ãã£ãã察話: æ¯ç§10ããŒã¯ã³ä»¥äžããã°äŒè©±ãšããŠã¹ãã¬ã¹ãå°ãªã
ã³ãŒãã£ã³ã°è£å®: æ¯ç§30ããŒã¯ã³ä»¥äžã»ãããåŸ ã¡æéãäœæ¥ãªãºã ãå£ããã
å€éãããåŠç(èŠçŽã»åé¡ãªã©): æ¯ç§æ°ããŒã¯ã³ã§ãæç«ãããé床ãã容éãåªå ããŠãã
ãææ°ã§é«ã=éããã§ã¯ãªãããèªåã®ã¢ãã«ãµã€ãºãšåž¯åã®å²ãç®ãã§æ±ºãŸãããããããŒã«ã«LLMã®æ©æéžã³ã§å¯äžèŠããã¹ãåçã§ãã
ãŸãšã
LLMã®äœæé床ã¯ãã³ãŒãã§æ±ºãŸãããã³ãŒãã¯æ¯ããŒã¯ã³ãã¢ãã«å šäœãã¡ã¢ãªããèªã¿åºã
ã ããé床 â ã¡ã¢ãªåž¯å ÷ ã¢ãã«å æãµã€ãºãå®å¹ã¯ãã®7ã8å²
ããŒãã¯4éçŽ: GPUäžäœ1TB/sçŽ / Appleäžäœ500ã800GB/sçŽ / çµ±åã¡ã¢ãªæ©250GB/sçŽ / ããŒãPCã»CPU 90ã140GB/sçŽ
MoEã¢ãã«ã¯æŽ»æ§ãã©ã¡ãŒã¿ã§èªã¿æ¿ããã倧容éã»äœåž¯åãã·ã³ã¯MoEåã
è²·ãåã«ã垯å÷ã¢ãã«ãµã€ãºããèšç®ããçšéå¥ã®èš±å®¹ã©ã€ã³(察話10ãè£å®30ããŒã¯ã³/ç§)ãšæ¯ã¹ã
çšéã«åãæ©ææ§æã確ããããæ¹ãž
TodoONadaã§ã¯ãããŒã«ã«LLMã»çæAIã®å°å ¥å€æããPoCã»æ§ç¯ãŸã§ãæ¯æŽããŠããŸãã
çšéãšäºç®ã«åãããæ©æã»ã¢ãã«æ§æã®èšèš
ããã®æ§æã§æ¬åœã«å¿ èŠãªé床ãåºãããã®äºåæ€èšŒ
瀟å ããŒã¿ãæ±ãæ§æã®ã»ãã¥ãªãã£èšèš
ãŸãã¯30äžåã»2é±éèŠæš¡ã®ããæ€èšŒãããé«äŸ¡ãªæ©æãè²·ãåã«ãçãã®æ§æã§é床ãåºããã確ããããšããããã§ããçžè«ãã ããããåãåããã¯ãã¡ãã