گوگل
@SemiconductorsX
امین وحدت، مغز متفکر زیرساخت هوش مصنوعی گوگل توی تایوان پرده از یه ماجرای تازه برداشته و گفته چیپ اختصاصی بعدی گوگل احتمالاً دیگه از جنس TPUهای غولپیکر قبلی نباشه

امین وحدت، مغز متفکر زیرساخت هوش مصنوعی گوگل توی تایوان پرده از یه ماجرای تازه برداشته و گفته چیپ اختصاصی بعدی گوگل احتمالاً دیگه از جنس TPUهای غولپیکر قبلی نباشه. حرف حسابش اینه که واسه ایجنتهای هوش مصنوعی، مشکل اصلی دیگه کمبود قدرت پردازش خام نیست؛ گیر اصلی، انتقال سریع دادهها و هماهنگ کردن کارها با کمترین تأخیره که این یه کار کاملاً پردازندهمحوره. خلاصهش اینه که بعد از جدا شدن مرحله آموزش از استنتاج، حالا نوبت یه تراشه کنترلی اختصاصی برای مدیریت ایجنتهاست، چون پردازش الان مفت تموم میشه ولی هماهنگسازی و لنگ نزدن سیستم بدجوری خرج برمیداره. از اون طرفم رابطه گوگل با بر و بچههای زنجیره تأمین تایوان از خرید و فروش ساده گذشته و حالا دارن شونهبهشونه هم این چیپها رو طراحی میکنن.
منبع
Semiconductor Insider@SemiconductorsX
Google’s next custom chip may not be another TPU.
CTO Amin Vahdat says the bottleneck for AI agents isn’t raw compute, it’s moving data and scheduling work at low latency. That’s a CPU problem.
Training and inference already split. The next split: a dedicated control chip for agents. Compute is cheap. Coordination isn’t.
Semiconductor Insider @SemiconductorsX
Google CTO and head of AI infrastructure Amin Vahdat spoke at SEMICON Taiwan 2026 and then gave his first interview to Taiwanese media.
He described Google’s relationship with Taiwan’s supply chain as no longer a buyer–supplier arrangement.
The two sides are now co-designing chips, racks, and end-to-end system architecture.
Vahdat said: “This is no longer a matter of purchasing and partners.”
Without deep co-design with Taiwanese firms, Google could not have built its current infrastructure and compute platforms. Cooperation is accelerating month by month, quarter by quarter, and year by year.
۱۵ شهریور ۱۴۰۵مشاهده در X


نظرها
هنوز نظری نیست.
بخش نظرها فعلاً در دسترس نیست.