The emerging computer network component in the end may be its moment. in Nvidia’s GTC The event last week in San Jose, the company announced that it will produce a visual network key designed to significantly reduce energy consumption for artificial intelligence data centers. System-Uploaded Optics, CPO, Switch-Terabits per second can be directed from computers in a shelf to computers in another. At the same time, startup Micas networksShe announced that she is in volume production with the CPO key based on Broadcom Technology.
In the data centers today, the network keys in a shelf consist of computers from specialized chips linked to the optical refugees that have reached the system. (Electrical shelf connections are electrical, but many startups hope to change this.) Collecting the assembly and reception devices between lasers, visual circuits, digital signal treatments and other electronics. It makes an electric link for the key and translating data between electronic bits on the side of switching and photons that fly through the data center along the optical fibers.
Uploaded OPTICS is an attempt to increase the display of the frequency and reduce energy consumption by moving the conversion of visual/electrical data as soon as possible to the switch. This simplifies the preparation and provides energy by reducing the number of separate components needed, and electronic signals must travel the distance. The advanced packaging technology allows the chips makers to surround the network chip with many silicon visual transmission and reception tools. The optical fibers are attached directly to the package. Therefore, all ingredients are combined in one package except the laser, which remain external because they are made using non -silicon materials and techniques. (However, CPOS requires only one laser for every eight data links in NVIDIA devices.)
“Artificial Intelligence Computer with 400,000 graphics processing units is actually 24 megawatts laser.” – Back, Nafidia
An attractive technology, as its economies prevented her from publishing. “We were waiting for CPO forever,” he says. Clint ShowOptical Optical Expert and IEEE fellow at the University of California Santa Barbara, who was looking for technology for 20 years. Speaking about NVIDIA’s support for technology, he said that the company “will not do so unless the time is here when [GPU-heavy data centers] Power cannot be spent.
Indeed, NVIDIA referred to energy consumption in upcoming artificial intelligence centers as an incentive. Assembled optics consume “10 percent of the total GPU account” at the AI Ian PakVice President of NVIDIA to computerize Easter and high performance. In the 400,000 GPU factory that translates into 40 megawatts, and more than half that goes only to operate the laser in the assembly transmission and optics. “The artificial intelligence computer with 400,000 graphics processing units is actually 24 megawatts,” says.
Visual rates
One of the basic differences between the Broadcom and NVIDIA chart is the visual change technique that encodes electronic bits on the symptoms of light. In optical silicon, there are two main types of rates –MakhmanderIt is used by Broadcom and is the basis for assembled optics, and the Micorting ring, which chose NVIDIA. Previously, the light that travels through the wave guide is divided into two parallel arms. Each arm can then be modified by an applied electric field, which changes the light stage passes. The weapons then join the formation of one wave guide. Depending on whether the two signs are now in a stage or outside the stage, they will cancel each other or collect. Thus, electronic bits can be encrypted on light.
Micro rates are much more compact. Instead of dividing the light on the path of two parallel tracks, the wave guide is suspended in the form of a loop on the side of the main path of the light. If the light with a wavelength can form a permanent wave in the ring, it will be removed, and filter this wavelength outside the main wave guide. Exactly any wavelength that is frequented with the loop depends on the structure refractive index, which can be processed electronically.
However, the micOring pressure comes at a cost. The fine rates are sensitive to temperature, so each one requires a built -in heating circle, which must be carefully controlled and consumes energy. On the other hand, Mach-Zinder devices are much larger, which leads to more lost lighting and some design problems, says Schow.
Shaw says NVIDIA has managed to market a Micooring silicone engine that is “an amazing engineering achievement.”
NVIDIA CPO keys
According to NVIDIA, the adoption of the CPO keys in a new AI data center will lead to a quarter of the number of lasers, which enhances energy efficiency for trafficking data 3.5-fold, improving the reliability of signals that make it from one computer to another on time by 63-Times, and make Networks more than 10 times the conjunction of extinction, and allowing customers with new new data 30 in 100.
“By combining silicon optical lights into the keys, the old NVIDIA restriction of the track networks and institutional networks and the opening of the gate to the GPU AI”.
https://www.youtube.com/watch?YouTubeYoutu.be
The company plans two categories of switching, Spectrum-X and Quantum-X. Quantum-X, which the company says will be available later this year, depends on Inviniband Network technology, which is a more high -performance network scheme. It provides 800 GB/s from each of 144 outlets, and CPO chips have a liquid refrigerator instead of air cooling, as well as an increasing part of the new artificial intelligence centers. The ASIC NVIDIA’s network includes FP8 sharp Technology, which allows central processing units and graphics processing units to empty some tasks to the network chip.
Special-X is an ETIRNET key that can make a total frequency display of about 100 Terabits per second from either 128 outlets or 512 ports and 400 terabytes/s from 512 or 2048 ports. Spectrum-X is expected to have ready in 2026.
NVIDIA works on basic optical technology for years. But it took cooperation with 11 partners – including TSMCand CorningAnd Foxconn– To get a switch to a commercial state.
Ashkan Sidi, director of visual communication products at NVIDIA, stressed how important it is that the technologies that these partners brought to the table were improved to meet the needs of the Artificial Intelligence Data Center instead of collecting simply from the current technologies of these partners.
“Innovations and energy saving from CPO are closely related to your packaging scheme, your packaging partners, and your packaging flow,” says Sidi. “The grandmother is not only in the visual ingredients directly, but it is how to fill it in a high -return way, it can be tested that you can manage at a good cost.”
The test is especially important, because the system is an integration of many expensive ingredients. For example, there are 18 silicone optical pieces in each of the CPU-X CPU. Each of these should call a laser and 16 optical fibers. Sidi says that the team had to develop many new test procedures to obtain them properly and follow where the mistakes were crawling.
Micas Networks keys
Micas Networks has already been produced with a key -based key in Broadcom.Micas Network
Broadcom has chosen Mach-Zinder Baillly CPO SwitchThis is partially explained because it is a more uniform technique, which may facilitate integration with assembly future infrastructure, and explains Robert HanaFirst Director for Product Marketing at the Broadcom Visual Systems Department.
Mikas system One CPO component, which consists of Broadcom’s Tomahawk 5 ETHERNET switch chip is surrounded by eight optical silicon optical engines 6.4 TB/s. Air cooler devices are in full production now, which provides them before CPO keys in NVIDIA.
Hannia is called NVIDIA, in support of Micas and Broadcom timing. He says: “Several years ago, we made a decision to ski on the whereabouts of a pile,” he says. Mitch GalbrithThe chief operational officer in Mikas. With the database operators scrambled to run their infrastructure, the CPO time appears to have come, he says.
The new key is 40 % for energy saving against the systems filled with transmission and standard reception devices. but, Charlie isThe Vice President of Corporate Strategy at Micas says that CPO’s high reliability is less important. “Linger“The term transient failure of homogeneous visual links is one of the perpetrators responsible for prolonging the long -term artificial intelligence training, he says. CPO is expected to have a lower link because there are less ingredients in the signal path, among other reasons.
CPO in the future
Schow suggests that the large data centers to save energy are often used from CPO. After that, “I think it will be the new normal.” However, the improvements to other advantages of the electronics will allow CPO makers to continue to increase the frequency range – for at least a period.
Schow doubts the individual silicone rates – which runs at a speed of 200 GB/second in NVIDIA optical engines – you will be able to exceed more than 400 GB/s. However, other substances, such as lithium Newptom and Andium phosphate should be able to overcome this. The trick will be integrated with the ingredients OpenLight It works on, among other groups.
Meanwhile, homogeneous optics are not standing. Broadcom has unveiled this week a new digital signal processor that can reduce energy by more than 20 percent for the 1.6 TB/reception device, partly due to a more advanced silicone process.
And startups like Aveccinaand Ayar LabsAnd Lightmatter It brings visual interconnection all the way to the graphics processing unit itself. The former two developed chips aimed at going inside the package itself as GPU or another processor. Lightmatter runs a step further, which makes the optical silicone motor on the substrate that is recognized on 3D chips.
From your site articles
Related articles about the web
adxpro.online