Why is the AI era of forcing the entire computing

For decades, has been due to the efficiency and effectiveness of the nose’s law, the efficiency of the nose’s law and the gentle equipment and space equipped with the scale and the scale. This architecture provided online education to the billions and put them to all people to our fingers.

However, the next counting revolution requires even more. Performance AI’s promise Requires step to restrict excessive opportunities. To achieve this, we must revive some of the innovative funds innovations to achieve this industry and jointly in order to jointly and connect the technology. Let us examine this architecture forces to drive this architecture.

Specialized data to the executive equipment

For decades, in calculating, the democratization of the counting of the counting through the scale of almost the same, commodity servers, has been democratization. This uniform has allowed the placement of a flexible workload and use effective resources. Requirements like youThis can cancel the trend than on mathematical operations on mass databases.

We are now the determination of specialized hardware, including ISICS, GPUS, and tensorigues (TPU), and the beginning of the beginning and the beginning of the start of the following starts, the beginning and the beginning of the start of the following. Domain’s specific estimates are important to spread this summer, optimized, optimized on combatic issues, AI continuous achievement.

In addition to Ethernet: Increasing Specialized Interaction

These specialized systems need to be infected with the fact that they are often associated with “everyone, ‘, and closer to the TABIT-second transmitters and speed of local memory. Today’s industries are mainly based on goods flows and TCP / IP protocols, which is equipped to solve very difficult requirements.

As a result of scale gene Ai load In most specialized acceleration, we see specialized international relations, such as specialized means of instances, such as TPUS and NVLINK, such as ICI. The networks built on this purpose prefer to share transfers on the memory, preferred to share the transfers on the memory, efficient, efficiently, and are efficiently shared, efficient, efficient, and the special equipment to share information.

It is important to submit to the bonds in this connection, submissive towards contacts in Cittle Cenrus, and move to the next generation of AI.

Disposal of memory

The effectiveness of calculating for decades has survived the capacity of memory. Technovations such as fled and tramps such as Srievs such as partially softens the AI’s intense character increases the problem.

Waiting for the high-quality capacity of the units of seizure of guaranteed units (HBM), it will be reduced to a direct translation package to the working package. However, it faces the depths of HBM: restricts common information flows on the perimeter of the physical chip: creates external information flows on the perimeter of the physical chip and creates external dates of databases and creates externally restrictions.

These restrictions emphasize that the lack of convocals and processing architecture of the processing architecture. Without these innovations, you will be calculated sharp calculated for strong restrictions, efficiency and scale.

Up to the high density of the server field

Studying today (ml) studying (ml) This is a stubborn syncion, which is stubborn at the microsection level, makes new requirements. Unlike systems embraced, ML calculations require homogeneous elements; Mixing joints are faster. Contacts should be in advance and higher efficiency, because the delay of one element can stop the whole process.

Coordination and these special requirements must be used to calculate the unprecedented density. Due to processors between processors, it is delayed delay and reducing electricity consumption, leaving the new class of ultra-density AI systems.

This disk requires optimal design for infrastructure for infrastructure for the most closely coordinated calculations, and requires the infrastructure to increase the physical schemes and strength, strengthening physical schemes and dynamic management.

A new approach to harm

A traditional injury is based on the most essentials between the most connected systems to achieve a trip. ML Recount requires another method.

First, the counting scale must be very expensive. Secondly, model training is a synchonic process that uniforms can remain in cascade in cascade in cascade. Finally, ML HARDWARE often motivates the boundary of current technology, and leads to a high level of failure.

Instead, the developing strategy is to register – in real registration – real-time monitoring, rapid distribution of spare resources and quick reorganization. The bed equipment and network design should be detected and replacement of SWIFT failure to save the performance.

Stable approach to power

To get acquainted with the power of the power today and strength You have calculated. We need to be aimed at the final design of the traditional system, when the traditional system is directed to maximum implementation of the Chip. Бул мамиле – бул тутумдун бардык компоненттерин – эсептөө, тармак, эстутум, коңгуроо, муздатуу, муздатуу, муздатуу, муздатуучу жана кемчиликсиз иш-аракеттерди жүргүзүү үчүн биргелешип иштөө. The operation optimization components limit the effectiveness of the general system.

As we motivate our great performance, personal chips are more powerful, and often require that in traditional air centers of information centers. This requires more energy, but it will require effective, fluid refrigeration solutions and key development infrastructure.

In addition to cools, they create uniform sources, significantly finance costs, and create significant financial costs, except for cooling natural communal and diesel generators. Instead, we need to unite different electrical sources, but we need to be multi-energy sources and preservation on a large microbian controller. AI should be flexibility and geographical distribution of flexibility and geographical distribution, and we may not be required only hours in the system of expensive systems.

This developing electricity allows the accessibility of electricity to real time – due to lack of deficiencies such as wildlights, such as wildlights. All this is currently available at the level of real time requires telemetry and activation.

Security and Privacy: Baked is not baked

Critical lesson in the Internet Age is security and privacy is not effective in current architecture. Wonderful actors threats are only complex, user information and m.Ls of intellectual property Requires protection for infrastructure. One important observation is Ai, finally, the attacker will improve. This, in turn, means that we should ensure that we should absorb our defense at the same time.

This includes technical information magazines, technical information magazines, equipment and technical safety boundaries to protect the latest information, technical calculations, technical calculations, technical calculations, and complex basic management systems. The use of these guarantees for this security is important to protect users and maintain their confidence. The fact that telecom and routes can be a price / second, which can be a real-time monitoring, and the key to neutralization, including those from the dangers of insider.

Speed as a strategic imperative

The rhythm of hardware renewal changed dramatically. Dissemination of superkompeses, unlike traditional infrastructure machine, requires a completely different way. This is because ML does not work easily to heteogenic distribution; The comparison should be tailored to every new device generation every new device generation completely to exercise their capabilities and compilate. In one part of two or more activities, the new equipment, two or more actions that have reached new equipment can be used to deliver two or more activities.

Therefore, instead of mass and synchronons of one homogeneous equipment, it often was simultaneously synchononed by the center of the whole center. This refreshes the improvement of the most weakness of its hardware, and it is a quick solution to the ability to resist an AI engines.

The goal is 100,000 plus the expansion of chips from design from design to increase the effectiveness of algorithm disruption. This requires each key to demand a radical acceleration and automation and modeling model for industrial infrastructure. One step one step should be automated for each step to monitor and repair architecture.

At present: Collective effort for the next Gen InfratRuture

The gene Ai Not only evolution but also a revolution that requires the counting infrastructure to recover. Difficulties – specialized equipment, interacted equipment, and sustainable operations are important, but also AI transformer.

It is easy to know that our computation infrastructure has not existed in several years, so we can not only improve grain. Instead, we should try to study the study, industrial industry, industrial industry, the first principles, to create a new plan for basic global infrastructure. This, in turn, leads to business to business into business on business and efficiency as an unprecedented scale.

Amin Wahat is a student study, systems and cloud AI GM and GM Google cloud.