News Posts matching #Ascend 910C

Return to Keyword Browsing

Huawei Ascend 910C Accelerator's Maturation Allegedly Spurred NVIDIA H200 Export Reversal

Earlier this week, the US government approved exports of NVIDIA H200 data center-grade accelerators to China. This reverses a previous decision, to prevent "full fat" H200 hardware reaching Chinese shores. Team Green engineers were reportedly working on a severely reduced "H20" alternative—whispers of this sanction-conforming design emerged earlier on in 2025. A competing Chinese product—Huawei's Ascend 910C accelerator—is often regarded as a main rival to newly approved past-gen Team Green AI-crunching solutions. According to Bloomberg's inside sources, White House officials were assessing the Ascend 910C's capabilities—most notably in the form of Huawei's super node CloudMatrix 384 system.

Despite lagging behind the H200, in terms of computing power and memory bandwidth, a large array of Ascend 910C accelerators (384 in total) is still a significant force. Additionally, Bloomberg's unnamed sources allege that the current US administration is worried about China rapidly improving its domestic semiconductor production capacities. Anonymous industry figures have whispered about Huawei and foundry partners working towards manufacturing 600,000 units throughout the coming year. Supposedly, the aforementioned White House staffers have considered a slightly different scenario—that Huawei: "would be capable, in 2026, of producing a few million of its Ascend 910C accelerators." According to a related Reuters news article, Chinese regulators are reportedly weighing up methods that will limit local industry's access to NVIDIA H200 systems.

NVIDIA Gets H200 Export License for China, With 25% of Revenue Going to the U.S. Government

The Trump administration has approved a framework allowing NVIDIA to resume exports of its H200 data center accelerators to China, subject to a revenue-sharing condition that allocates 25% of sales proceeds to the U.S. government. Rather than lifting export restrictions outright, the arrangement requires that recipients be vetted through Commerce Department licensing, so shipments will be controlled and limited to approved commercial customers. The decision reopens a major market for high-end accelerators while maintaining a regulatory channel to monitor distribution.

The H200 chip remains a powerful solution for large-scale AI workloads even today. Manufactured using TSMC's N4 process, it combines NVIDIA's "Hopper"-derived architecture with 141 GB of HBM3E memory and 4.8 TB/s memory bandwidth, making it ideal for large model training and dense inference tasks. Although it doesn't match the performance of NVIDIA's current "Blackwell" family and other upcoming accelerators, the H200 is a mature product that has been shipping since spring 2024 and benefits from a well-developed driver and software ecosystem. Its raw compute power is roughly double that of solutions like the Huawei Ascend 910C, making it very competitive even today.

China's AI Chip Ambitions Limited by HBM Memory Supply, Notes Report

According to SemiAnalysis, Chinese AI semiconductor buildout faces an unexpected bottleneck that's more critical than manufacturing constraints: high-bandwidth memory shortages. While domestic foundries like SMIC can produce sufficient AI processors, the lack of HBM supply chains has emerged as the primary limitation throttling China's semiconductor ambitions. Huawei's Ascend 910C illustrates this mismatch. The company has foundry capacity to manufacture 805,000 units annually through TSMC, which is a result of stockpiling before the cutoff, as well as SMIC. However, HBM shortages prevent realizing this production potential. Chinese firms accumulated roughly 11.4 million HBM stacks from Samsung before export controls tightened, but these stockpiles won't sustain long-term AI scaling. The memory crunch effectively neutralizes Chinese manufacturing capabilities, giving Western competitors like NVIDIA and AMD a sustained advantage despite Beijing's substantial semiconductor investments.

The path toward HBM independence runs through domestic memory manufacturers like CXMT and the recent collaboration with YMTC. Converting standard DRAM production to HBM manufacturing requires specialized equipment, which is predominantly supplied by Western suppliers, creating another dependency layer. China could achieve competitive HBM3E production by 2026 if current investment trajectories continue and equipment restrictions remain stable. However, this timeline assumes successful technology transfer and uninterrupted access to critical manufacturing tools. If export controls expand to encompass additional HBM-related equipment, the development window could become considerably longer. YMTC is reportedly ready to enter DRAM production and could begin purchasing equipment by late 2025, utilizing its secret weapon, Xtacking technology, one of the most advanced hybrid bonding techniques available. This matters because HBM requires stacking at least 16 memory chips with incredible precision, something YMTC has been doing for years with their NAND products.

Huawei Open-Sources AI Software Stack, Aiming to Rival CUDA

Huawei used its annual developer summit in Shenzhen this week to announce that it will release the Compute Architecture for Neural Networks (CANN) software stack, which powers its Ascend AI processors, as open source before the end of the year. Rotating chairman Eric Xu Zhijun said the goal is to make it easier for developers to get started and build a lively community around Huawei's chips. In the weeks leading up to the public launch, the company has collaborated with Tsinghua University, Peking University, AI startups such as StepFun and Infinigence, and the newly formed Model-Chips Ecosystem Innovation Alliance to agree on contribution rules and establish transparent governance. By putting CANN on public code repositories under a permissive licence, Huawei hopes to capture the same kind of grassroots enthusiasm that drove Linux and RISC-V adoption, only this time for hardware designed in China.

This announcement comes at a sensitive moment on both technical and political fronts. China's regulator has initiated a security review of NVIDIA's H20 GPUs, citing concerns about remote-management features that compromise data sovereignty. At the same time, US authorities have approved limited exports of the same chip. Unlike NVIDIA's CUDA framework, which now blocks third-party translation layers on non-NVIDIA GPUs, CANN will allow forks and cross-platform ports from day one. It will take years for CANN to match CUDA's two decades' worth of optimized libraries and tooling. If Huawei can keep contributors engaged, an open-source CANN may shift the company's role from chip maker to the neutral steward of a unified AI software ecosystem. The first public snapshot is scheduled for the fourth quarter, alongside the release of the Ascend 910C processors and a joint reference server design with Tencent Cloud.

DeepSeek R2 Leak Reveals 512 PetaFLOPS Push on Domestic AI Accelerator Infrastructure

DeepSeek, a company that took the AI world by storm with its R1 model, is preparing a new and reportedly much improved DeepSeek R2 model release, according to a well-known AI insider @iruletheworldmo on X. Powered by Huawei's Ascend 910B chip clusters, a possible Huawei Atlas 900, and DeepSeek's in-house distributed training framework, R2 pushes these accelerators to an impressive 82% utilization, translating to 512 PetaFLOPS of FP16 performance—half an exaFLOP in computing power. According to Huawei lab data, that's roughly 91% of what NVIDIA's older A100 clusters deliver, yet DeepSeek claims it cuts per-unit training costs by a remarkable 97.3%. Behind DeepSeek R2 is a carefully cultivated ecosystem of partners. Tuowei Information, a leading OEM in the Ascend family, manages over half of DeepSeek's supercomputing hardware orders, while Sugon provides liquid-cooled server racks capable of handling up to 40 kW per unit. To keep power consumption in check, Innolight's silicon-photonics transceivers shave off another 35% compared to traditional solutions.

Geographically, operations are split across major hubs: Runjian Shares runs the South China supercomputing center under contracts exceeding ¥5 billion annually, and Zhongbei Communications maintains a 1,500-PetaFLOP reserve in the Northwest for peak demands. On the software side, DeepSeek R2 already supports private deployment and fine-tuning, powering smart-city initiatives in 15 provinces through the Yun Sai Zhilian platform. North China's node, overseen by Hongbo Shares' Yingbo Digital, adds another 3,000 PetaFLOPS to the mix. If computing power is scarce, Huawei is ready to deploy its CloudMatrix 384 system, which is positioned as a domestic alternative to NVIDIA's GB200 NVL72. It features 384 Ascend 910C accelerators to achieve 1.7× the overall petaFLOPS and 3.6× the total HBM capacity of the NVL72 cluster—yet it lags significantly in per-chip performance and consumes nearly four times more power. Nonetheless, the R2 model launch is expected to come smoothly online, and we are waiting for the official launch and benchmarks to see its performance.

Huawei Prepares 6 nm Ascend 920C Accelerator: 900 TeraFLOPS, 4000 GB/s HBM3

Huawei recently revealed that its CloudMatrix 384 AI super node cluster can outperform NVIDIA's GB200 NVL72 in standard benchmarks, even though it consumes more power per performance unit. That system relies on Huawei's current Ascend 910C accelerators, which deliver strong raw compute performance but lag behind in efficiency metrics. To address this gap, Huawei is preparing the Ascend 920 family, with the training-focused Ascend 920C built on SMIC's 6 nm process. According to DigiTimes, each 920C card will deliver more than 900 TeraFLOPS of BF16 half-precision performance. It also upgrades memory to HBM3 modules, providing 4,000 GB/s of bandwidth, which is up from the Ascend 910C's eight HBM2E stacks and 3,200 GB/s.

The existing Ascend 910C peaks at 780 TeraFLOPS in BF16 operations and uses a chip-to-chip interconnect bandwidth of 400 GB/s. Its packaging limits that bandwidth, but it still supports high-speed communication between nodes in ultra-dense AI clusters. Huawei will retain the chiplet-based design in the 920C and refine the tensor acceleration engines for Transformer and Mixture-of-Experts models. Internal projections estimate that overall training efficiency on the 920C will improve by 30-40 percent compared to the 910C. This should narrow the performance-per-watt difference against competitors' solutions. In terms of system integration, the Ascend 920C will support PCIe 5.0 and next-generation high-throughput interconnect protocols. These features aim to improve resource scheduling and reduce latency in super node deployments, where tight node-to-node synchronization is critical. Huawei has not announced a firm release date for the Ascend 920C, but DigiTimes sources claim that it will enter mass production in the second half of 2025, which could mean just a few months from now.

Report Suggests Huawei Ascend 910C AI Accelerator's Utilization of Foreign Parts; Investigators Find 7 nm TSMC Dies

Earlier today, TechPowerUp covered the alleged performance prowess of Huawei's CloudMatrix 384 system super node. According to SemiAnalysis opinion, the system's Ascend 910C AI accelerators are a generation behind—in terms of chip performance—when compared to NVIDIA's GB200 "Blackwell" AI GPU design. SMIC seemed to be in the picture, as Huawei's main fabrication partner—possibly with an in-progress 5 nm node process. Instead, SemiAnalysis has surmised that the Ascend 910C is based on plenty of non-native technologies. Huawei's (current and prior) "aggressive skirting of export controls" has likely enabled the new-gen AI chip's better than expected performance stats. SemiAnalysis documented the early sample's origins: "while the Ascend chip can be fabricated at SMIC, we note that this is a global chip that has HBM from Korea (Samsung), primary wafer production from TSMC (Taiwan), and is fabricated by 10s of billions of wafer fabrication equipment from the US, Netherlands, and Japan...One common misconception is that Huawei's 910C is made in China. It is entirely designed there, but China still relies heavily on foreign production."

Despite China's premiere foundry business making pleasing in-roads with a theorized "7 nm N+2" manufacturing test line, Huawei has seemingly grown impatient with native immature production options. Today's SemiAnalysis article presents a decent dose of inside knowledge: "while SMIC does have 7 nm, the vast majority of Ascend 910B and 910C are made with TSMC's 7 nm. In fact, the US Government, TechInsights, and others have acquired Ascend 910B and 910C and every single one used TSMC dies. Huawei was able to circumvent the sanctions on them against TSMC by purchasing ~$500 million of 7 nm wafers through another company, Sophgo...It is rumored Huawei continues to receive wafers from TSMC via another 3rd party firm, but we cannot verify this rumor." Another (fabless) Chinese chip design firm—Xiaomi—appears to still have direct/unrestricted access to TSMC manufacturing lines, albeit not for enterprise-grade AI products.

Huawei CloudMatrix 384 System Outperforms NVIDIA GB200 NVL72

Huawei announced its CloudMatrix 384 system super node, which the company touts as its own domestic alternative to NVIDIA's GB200 NVL72 system, with more overall system performance but worse per-chip performance and higher power consumption. While NVIDIA's GB200 NVL72 uses 36 Grace CPUs paired with 72 "Blackwell" GB200 GPUs, the Huawei CloudMatrix 384 system employs 384 Huawei Ascend 910C accelerators to beat NVIDIA's GB200 NVL72 system. It takes roughly five times more Ascend 910C accelerators to deliver nearly twice the GB200 NVL system performance, which is not good on per-accelerator bias, but excellent on per-system level of deployment. SemiAnalysis argues that Huawei is a generation behind in chip performance but ahead of NVIDIA in scale-up system design and deployment.

When you look at individual chips, NVIDIA's GB200 NVL72 clearly outshines Huawei's Ascend 910C, delivering over three times the BF16 performance (2,500 TeraFLOPS vs. 780 TeraFLOPS), more on-chip memory (192 GB vs. 128 GB), and faster bandwidth (8 TB/s vs. 3.2 TB/s). In other words, NVIDIA has the raw power and efficiency advantage at the chip level. But flip the switch to the system level, and Huawei's CloudMatrix CM384 takes the lead. It cranks out 1.7× the overall PetaFLOPS, packs in 3.6× more total HBM capacity, and supports over five times the number of GPUs and the associated bandwidth of NVIDIA's NVL72 cluster. However, that scalability does come with a trade-off, as Huawei's setup draws nearly four times more total power. A single GB200 NVL72 draws 145 kW of power, while a single Huawei CloudMatrix 384 draws ~560 kW. So, NVIDIA is your go-to if you need peak efficiency in a single GPU. If you're building a massive AI supercluster where total throughput and interconnect speed matter most, Huawei's solution actually makes a lot of sense. Thanks to its all-to-all topology, Huawei has delivered an AI training and inference system worth purchasing. When SMIC, the maker of Huawei's chips, gets to a more advanced manufacturing node, the efficiency of these systems will also increase.
Return to Keyword Browsing
Nov 19th, 2026 06:14 PST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts