¡Ø º» »óǰÀº ¿µ¹® ÀÚ·á·Î Çѱ۰ú ¿µ¹® ¸ñÂ÷¿¡ ºÒÀÏÄ¡ÇÏ´Â ³»¿ëÀÌ ÀÖÀ» °æ¿ì ¿µ¹®À» ¿ì¼±ÇÕ´Ï´Ù. Á¤È®ÇÑ °ËÅ並 À§ÇØ ¿µ¹® ¸ñÂ÷¸¦ Âü°íÇØÁֽñ⠹ٶø´Ï´Ù.
ÁöÁ¤ÇÐÀû Á¦¾à°ú ¼öÀÔ ±ÔÁ¦°¡ Áß±¹¿¡¼ RTX PRO¸¦ ¾ïÁ¦Çϰí ÀÖÀ¸¸ç, Ưº°ÆÇ Ãâ½Ã´Â ±¤¹üÀ§ÇÑ ¼ö¿ä¸¦ ±â´ëÇϰí ÀÖ½À´Ï´Ù. CPX´Â ¿ë·®À» ´Ã¸° GDDR7 ¸Þ¸ð¸®¸¦ äÅÃÇÏ¿© HBM¿¡ ÀÇÁ¸ÇÏ´Â Rubin GPU¸¦ º¸¿ÏÇÏ´Â ÇÑÆí, ´çÀåÀÇ ¸ð¸àÅÒÀº ÇöÀç ¼Ö·ç¼Ç¿¡ ¸Ó¹°·¯ ÀÖ½À´Ï´Ù.
ÁÖ¿ä ÇÏÀ̶óÀÌÆ®
- Áß±¹ ½ÃÀåÀº ÁöÁ¤Çаú ÅëÁ¦ÀÇ Á¦¾àÀ» ¹Þ°í ÀÖÀ¸¸ç, Ưº°ÆÇ Ãâ½Ã´Â Æø³ÐÀº ¼ö¿ä¸¦ ±â´Ù¸®°í ÀÖ½À´Ï´Ù.
- CPX´Â GDDR7 ¸Þ¸ð¸®¸¦ »ç¿ëÇÏ¿© ¿ë·®À» Áõ°Çϰí, HBMÀ» »ç¿ëÇÏ´Â Rubin GPU¸¦ º¸¿ÏÇÕ´Ï´Ù.
- ¼³°èÀÇ ¸ñÇ¥´Â AIÀÇ ÇнÀ°ú Ãß·ÐÀ̸ç, ¾ç»ê ½Ã±â´Â ¹ÌÁ¤ÀÔ´Ï´Ù.
- CSP ·¢ ¹× ÅëÇÕ ½Ã½ºÅÛ¿¡ ´ëÇÑ ¼ö¿ä°¡ Áõ°¡Çϰí ÀÖÀ¸¸ç, °ø±ÞÀÚ´Â º¸´Ù ÅëÇÕÀûÀÎ Á¢±Ù ¹æ½ÄÀ» ¿ä±¸Çϰí ÀÖ½À´Ï´Ù.
- CPX¿Í RubinÀº ¼·Î¸¦ º¸¿ÏÇϰí ÀÖÀ¸¸ç, ¶Ñ·ÇÇÑ Ä«´Ï¹ß¶óÀÌÁ¦À̼ÇÀº º¸ÀÌÁö ¾Ê½À´Ï´Ù.
ÀÌ º¸°í¼¿¡¼´Â ¼¼°è ¹ÝµµÃ¼ »ê¾÷¿¡ ´ëÇÑ Á¶»ç ºÐ¼®°ú ÇÔ²² NVIDIAÀÇ Rubin CPX¿¡ ´ëÇÑ ¹Ì·¡ Àü¸ÁÀ» Á¦°øÇÕ´Ï´Ù.
¸ñÂ÷
Á¦1Àå ¼·Ð
- NVIDIA°¡ »õ·Ó°Ô ¹ßÇ¥ÇÑ Rubin CPX´Â AI ºÐ»ê Ãß·Ð ¾ÆÅ°ÅØÃ³ ÄÁÅØ½ºÆ® ½ºÅ×ÀÌÁö ÄÄÇ»ÆÃÀ» °ü¸®ÇÕ´Ï´Ù.
- VR NVL 144 CPX ÄÄǻƮ Æ®·¹ÀÌ¿¡´Â AI Ãß·Ð ±â´ÉÀ» °ÈÇÏ´Â 8°³ÀÇ Ãß°¡ Rubin CPX GPU°¡ Æ÷ÇԵǾî ÀÖ½À´Ï´Ù.
- VR NVL144 CPXÀÇ ÃßÁ¤ ¼ÒºñÀü·ÂÀº ·¢´ç ¾à 400kWÀÔ´Ï´Ù.
Á¦2Àå NVIDIAÀÇ GDDR7 äÅÃÀº 2026³â¿¡µµ RTXC PRO 6000¿¡ ÁÖ·ÂÇϸç, Rubin CPX ½Ã½ºÅÛÀÇ ¾ç»êÀº 2026³â¸»À̳ª 2027³â¿¡ °³½ÃµÉ Àü¸Á
- GB200/GB300 ·¢ÀÇ ¼ö¿ä(ÃâÇÏ´ë¼ö) ºÐÆ÷ ¿¹Ãø : ÁÖ¿ä CSP(°í°´)º°(2025³â)
Á¦3Àå ÁöÁ¤ÇÐÀû ¿äÀο¡ ÀÇÇØ H20Àº Áß±¹¿¡¼ ÈÄÅðÇßÀ¸³ª, ½ÃÀå¿¡¼´Â RTX PRO 6000ÀÇ ±âȸµµ âÃâµÇ¸ç 6000DÀÇ Æ¯º°ÆÇÀº 2025³â 4ºÐ±â ÀÌÈÄ¿¡ Áõ»êµÉ Àü¸Á
Á¦4Àå NVIDIA CPX°¡ GDDR7ÀÇ ÇâÈÄ ¼ºÀåÀ» ÃËÁøÇÒ °ÍÀ¸·Î ¿¹Ãø
KSA
Geopolitical constraints and import controls curb RTX PRO in China, with a special edition awaiting broader demand; CPX adopts GDDR7 memory with increased capacity, complementing Rubin GPUs that rely on HBMs, while near-term momentum remains with current solutions.
Key Highlights:
- China market constrained by geopolitics and controls; the special edition release is awaiting broader demand.
- CPX uses GDDR7 memory to boost capacity and complements Rubin GPUs using HBMs.
- Design targets AI training and inference; mass production timing remains to be clarified.
- Demand for CSP racks and integrated systems is rising; providers seek more integrated approaches.
- CPX and Rubin complement each other, with no clear cannibalization in sight.
Table of Contents
1. Introduction
- Newly Unveiled Rubin CPX from NVIDIA Manages Context-Stage Computing of AI Disaggregated Inference Architecture
- VR NVL 144 CPX Compute Tray Contains 8 Additional Rubin CPX GPUs to Enhance AI Inference Capabilities
- Estimated Power Consumption of VR NVL144 CPX Is Near 400kW per Rack
2. NVIDIA's Adoption of GDDR7 Will Still Focus on RTXC PRO 6000 in 2026; Ramp-Up of Rubin CPX System Is Expect to Start Either by End of 2026 or in 2027
- Estimated Distribution of Demand (Shipments) of GB200 and GB300 Racks by Major CSP (Customer) in 2025
3. Geopolitical Factors Have Caused Setbacks for H20 in China but Also Created Opportunities for RTX PRO 6000 in the Same Market; Ramp-Up of Special 6000D Version Is Expected to Occur After 4Q25
4. NVIDIA CPX Is Expected to Drive Future Growth for GDDR7