Huawei Pioneers a New Computing Architecture for the AI Era: Making One Million Processors Work as One Computer– Summary of Roundtable Meeting Between Eric Xu and International Media Outlets on September 17

SHANGHAI, CHINA– Media OutReach Newswire — 30 September 2026– Throughout HUAWEI CONNECT on September 17, 2026, Eric Xu (Huawei’s Rotating Chairman) and Dr. Liao Heng (Chief Scientist of HiSilicon) had a Q&A session with reporters of media outlets to talk about Huawei’s Peerium Computing Architecture for the AI period in addition to UnifiedBus(UB).

data-image-height=”0″>

At last year’s HUAWEI CONNECT, Eric Xu offered a keynote speech and launched Ascend chips and their roadmap (consisting of the 950, 960, 970, and SuperPoDs and SuperClusters ). He likewise revealed the open release of the UnifiedBus procedure.

At this year’s HUAWEI CONNECT, Huawei revealed the Ascend 950 and the Atlas 950 SuperPoD. While the market is discussing SuperPoDs or supernodes, the methods of making these items vary extremely. The secret of SuperPoDs has to do with enabling a number of thousand or perhaps 10s of countless processors to work as one computer system.

Today, Huawei revealed a SuperPoD based upon the Ascend 950 chip. The SuperPoD is powered by a brand-new computing architecture that Huawei has actually originated. This architecture now can make one million processors work as one computer system. This architecture is called Peerium, and it can accomplish strong scaling to the million-processor level through embedded parallelism, merged memory dealing with, and peer-to-peer adjoin. It extends the parallelism paradigm of Turing with the intro of Nested BSP (Nested Bulk Synchronous Parallel).

This architecture likewise extends the von Neumann single-machine architecture and overruns the master-slave style paradigm that has actually dominated for years. Considered that the readily available innovations in the market might not support this architecture, Huawei created a crucial adjoin innovation– UnifiedBus. Eventually, one million processors can really work as one bigger computer system.

UnifiedBus is a high-speed bus that utilizes an open procedure to scale without a limitation to link CPUs, NPUs, memory, SSDs, user interface cards, and changes. With UnifiedBus, peer-to-peer adjoin is attained throughout calculate, storage, and networking. This essentially overruns the standard master-slave architecture.

The Atlas 950 SuperPoD is an item built on the Peerium Computing Architecture. A SuperCluster with 256,000 computing cards is presently being released and evaluated.

Q: Dr. Liao, your paper pointed out a SuperPoD that’s comprised of 256,000 cards. Is that the ceiling of your innovation? Can you inform us more about market need for your items?

Liao Heng: 256,000 or approximately 200k, this number is not an approximate number.

Of all, the function of these big clusters is for training. It has a hardware facilities that can support the training of structure designs, which today are currently in the 5T specification variety.

Over the next 2 years, there will be approximately 6 or 7 frontier AI laboratories in China, and all will target at training structure language designs of sizes varying from 10 to 40 trillion criteria. These numbers sort of compare to the memory capability and the size of the SuperPoD. You require such a scale of facilities to train these designs.

The 2nd thing is that in China, a lot of information centers have a nationwide intend on this power grid. 200,000 is a fairly conservative number, offered the style of the power shipment system in a single self-governing zone and single information center school.

As I stated, everyone should be rather knowledgeable about the AI design race that’s occurring around the world. And in China, there are at least 3 or 4 tier-1 gamers that are currently well acknowledged and have actually attained tier-1 status.

Q: What are your strategies to motivate Chinese design service providers to utilize Ascend chips for training?

Eric Xu: For the Ascend 950, we initially introduced the 950PR to the marketplace. It’s mainly utilized for AI reasoning, and presently the overall volume of supply offered in the market is not huge. The SuperPoDs based upon the Ascend 950DT are primarily utilized for AI training. Now, they are under screening, and their massive supply will begin at the end of this year or early next year. We are having substantial discussions and conversations with AI design companies. The screening outcome of utilizing the Ascend 950DT for training is respectable. I think that beginning with next year, a great deal of AI design training will be based upon SuperPoDs that utilize the Ascend 950DT. I believe at the end of the day, it’s not about how tough we are going to press design companies, however about just how much supply capability we will have. We can just provide based upon just how much that’s produced. And we hope the worth chain will broaden its capability much faster to increase supply.

Q: What is Huawei’s take on current calls by leading AI design service providers in the United States to decrease AI advancement?

Eric Xu: We require to take a look at the level and cadence of AI advancement in China and the United States. AI designs in China are mainly open-source designs and their development is fairly transparent. If we take a look at the mainstream design service providers in the United States, they have more calculating power, so possibly just they themselves understand where they remain in regards to the level of elegance of their designs. The marketplace sensation about AI dangers may be weaker in China than in the United States. Design suppliers in China require to accelerate their rate to reach a level where they might likewise feel the dangers from AI advancement. And after that, possibly they will feel the like the leading United States design suppliers. I constantly think that we require to strike a balance in between driving AI advancement and handling AI dangers. At least, the bottom line is we require to make sure AI for excellent, not AI for evil.

Q: Does Huawei have strategies to bring the Atlas 950 SuperPoD to other markets outside China? If yes, which markets are you going to target? Can you show us information about your market share in China? What is your take on China’s push for chip self-sufficiency?

Eric Xu: Considering that we do not have adequate capability to even please the need in China, we do not have strategies to broaden to the worldwide market in a fully-fledged method. Certainly, there are numerous nations that have extremely strong need. We are doing some screening, and we are supplying some supply to those nations, however the volume is rather minimal.

Now, it’s quite difficult to gather information about the market share of NVIDIA in China. Based on the information we have actually gathered, Ascend has actually exceeded NVIDIA.

China’s push for chip self-sufficiency is an unavoidable course forward. China is a nation of 1.4 billion individuals and it’s a commercial powerhouse, so the method individuals work and live is carefully connected to chips. Chinese individuals have an eager sense of seriousness, and will decline a future in which others choose whether we can have access to particular items. Despite the fact that our chips might be less sophisticated, a minimum of their supply is ensured, so that you do not need to fret about chip supply day in and day out. I believe that’s definitely the course forward. For the Chinese federal government, for the domestic market, and for Huawei, the course forward is unquestionably to promote complete self-sufficiency in regards to chips and the whole semiconductor worth chain. We likewise think this will come true eventually.

To check out the complete Q&A session: https://www.apmultimedianewsroom.com/multimedia-newsroom/huawei-pioneers-a-new-computing-architecture-for-the-ai-era-making-one-million-processors-work-as-one-computer-2

Hashtag: #Huawei

The provider is entirely accountable for the material of this statement.