news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Jin Lei from Aofei Temple
Quantum Bit | Public Account QbitAI

Intel"Light", breaking through the thorny computing power problem in the era of large models——

Launched the industry's first fully integratedOCI(Optical Computing Interconnect) chip.

△Image source: Intel

You should know that at the moment when large AI models are developing in accordance with the Scaling Law, in order to achieve better results, either the model scale or the data scale is moving towards a larger trend.

This will result in large AI models placing higher demands on computing, storage, and intermediate I/O communications at the computing power level.

Intel's breakthrough this time isI/O Communication：

In CPU and GPU,Use optical I/O to replace electrical I/O for data transmission.

What is the use?

In a nutshell,Data transmission distance is much longer，The quantity is large，Lower power consumption——More suitable for the "physique" of large AI models.

△Image source: Intel

So why does Intel use "light"? How does it achieve this?

Using "light" to transform a horse-drawn carriage into a truck

The traditional electrical I/O method (copper wire connection) certainly has its advantages, such as supporting high bandwidth density and low power consumption, but the fatal problem isThe transmission distance is relatively short (less than 1 meter)。

This is not a problem if it is placed in a rack, but the computing power of large AI models is often at the level of server clusters.

Not only does it occupy a large area, it also spans multiple racks, and the cables need to be tens or even hundreds of meters long, and the power consumption is quite high; it will consume all the power supplied to the racks, so that there is not enough electricity to perform computing and reading and writing operations on the storage chips.

besides,Storage-to-Compute RatioOn the other hand, it is precisely because of the "big" characteristic of the large model that the ratio of reading once and doing hundreds of calculations has now become close to 1:1.

△Image source: Intel

This requires a new approach that can increase computing power and storage density while reducing power consumption and size, so that more computing and storage can be accommodated in a limited space.

With optical I/O, the problem is solved:

It can support 64 32Gbps channels in one direction over a fiber length of up to 100 meters.

A vivid metaphor is that it is like usingcarriage(Limited capacity and distance) to useCars and TrucksTo deliver goods (larger quantities and longer distances).

Moreover, even in a relatively short distance, OCI can be used to complete some higher-density and more flexible data transmission tasks.motorcycle, faster and more flexible.

It is worth mentioning that this OCI method is not just theoretical.

According to Intel, they have used proven silicon photonics technology to integrate silicon photonic integrated circuits (PICs) containing on-chip lasers, optical amplifiers, and electronic integrated circuits.

In addition, it has previously demonstrated OCI chips packaged with its own CPU, which can also be integrated with next-generation CPUs, GPUs, IPUs and other SOCs (system-on-chips).

That’s not all. Intel has also shipped more than 8 million silicon photonic integrated circuits, including more than 32 million lasers that are now in use.

△Image source: Intel

So the next question is:

How was Intel's OCI created?

Vice President of Intel Labs and Director of Intel China LabsSong JiqiangDuring the communication process, he made an in-depth analysis and interpretation of this issue.

△Song Jiqiang, Vice President of Intel Research Institute and Director of Intel China Research Institute

Silicon photonics technology combines two of the most important inventions of the 20th century: silicon integrated circuits and semiconductor lasers.

It supports faster data transfer over longer distances than traditional electronics while leveraging efficiencies in Intel’s high-volume silicon manufacturing.

The silicon photonic integration technology released by Intel this time, the OCI chip has reached the level of optoelectronic co-packaging.

This optoelectronic co-packaging places a silicon photonic integrated circuit (PIC) and an electronic integrated circuit (EIC) on a substrate to form an OCI core particle as an integrated connected component.

This means that xPU, including CPU and future GPU can be packaged together with OCI chip.

The OCI chip converts all electrical I/O signals from the data center's CPU into light, and transmits them between nodes or systems in two data centers through optical fibers.

The current bidirectional data transmission speed has reached 4Tbps. Its upper-layer transmission protocol is compatible with PCIe 5.0 and supports 64 32Gbps channels in one direction, which is sufficient for current data centers:

It uses eight pairs of optical fibers and consumes only 5 picojoules (pJ) per bit, or 10-12 joules, which is three times lower than the power consumption of pluggable optical transceiver modules (the latter is 15 pJ per bit).

△Image source: Intel

In an optical transmission channel, there are actually 8 different bands, and the frequency interval of each band is 200GHz, occupying a total of 1.6THz spectrum spacing for transmission.

From visible light to invisible light, the spectrum width of light is actually very wide. Starting from THz, it is close to optical communication.

So in what fields will OCI chips be used in the future?

In this regard, Song Jiqiang said:

One is that it can be used to achieve communication, and it can also be packaged together with computing chips such as CPU and GPU, so that computing and communication are packaged very tightly together.
Through silicon photonics integration and advanced packaging technology, Intel also has many different advanced packaging technologies, we can achieve higher-density I/O chiplets, and then combine them with other xPUs. In the future, based on chiplets, we will form many different types of computing and interconnection chips, which will have very good application prospects.

According to the performance evolution roadmap of OCI I/O interface core particles, it can currently achieve a technical solution of 32Tbps transmission speed, mainly relying on iterative and steady improvement of three indicators, namely:

• There are 8 stable bands in one optical fiber
• The optical data transmission rate of each band is 32Gbps
• Can pull 8 pairs of optical fibers at the same time without affecting each other

The multiplication of these three indicators means that the current data transmission speed is 2Tbps in one direction, and 4Tbps in two directions. In the future, it can continue to evolve upward and gradually improve bandwidth capabilities.

△Image source: Intel

Finally, Song Jiqiang also explained Intel's differentiation in silicon photonics integration technology:

Mainly, we made the high-frequency laser transmitter on the wafer and integrated the silicon optical amplifier. These are two core technologies, and both are manufactured at the wafer level.
Next, we can mass-produce such highly integrated lasers, because the advantage of this on-chip laser is that it can be transmitted using ordinary optical fibers.
And in terms of stability, it is almostAn error may occur once in 10 billion hours。

So what do you think of Intel's pick for "light"? Feel free to leave a comment in the comment section to discuss.

Reference Links:
[1]https://mp.weixin.qq.com/s/ozx_ficqlxjEPKa5AlBdfA
[2]https://community.intel.com/t5/Blogs/Tech-Innovation/Artificial-Intelligence-AI/Intel-Shows-OCI-Optical-I-O-Chiplet-Co-packaged-with-CPU-at/post/1582541
[3]https://www.youtube.com/watch?v=Fml3yuPR2AU