Huawei CloudMatrix AI efficiency has achieved what the corporate claims is a big milestone, with inner testing displaying its new knowledge centre structure outperforming Nvidia’s H800 graphics processing models in working DeepSeek’s superior R1 synthetic intelligence mannequin, in response to a completetechnical paperlaunched this week by Huawei researchers.
The analysis, carried out by Huawei Applied sciences in collaboration with Chinese language AI infrastructure startup SiliconFlow, offers what seems to be the primary detailed public disclosure of efficiency metrics for CloudMatrix384.
Nevertheless, it’s essential to notice that the benchmarks had been carried out by Huawei on its methods, elevating questions on unbiased verification of the claimed efficiency benefits over established business requirements.
The paper describes CloudMatrix384 as a “next-generation AI datacentre structure that embodies Huawei’s imaginative and prescient for reshaping the muse of AI infrastructure.” Whereas the technical achievements outlined seem spectacular, the shortage of third-party validation means outcomes must be considered within the context of Huawei’s persevering with efforts to exhibit technological competitiveness exterior of US sanctions.
The CloudMatrix384 structure
CloudMatrix384 integrates 384 Ascend 910C NPUs and 192 Kunpeng CPUs in a supernode, related by an ultra-high-bandwidth, low-latency Unified Bus (UB).
Not like conventional hierarchical designs, a peer-to-peer structure allows what Huawei calls “direct all-to-all communication,” permitting compute, reminiscence, and community sources to be pooled dynamically and scaled independently.
The system’s design addresses notable challenges in creating trendy AI infrastructure, notably for mixture-of-experts (MoE) architectures and distributed key-value cache entry, thought-about important for big language mannequin operations.
Efficiency claims: The numbers in context
The Huawei CloudMatrix AI efficiency outcomes, whereas carried out internally, current spectacular metrics on the system’s capabilities. To grasp the numbers, it’s useful to consider AI processing like a dialog: the “prefill” part is when an AI reads and ‘understands’ a query, whereas the “decode” part is when it generates its response, phrase by phrase.
In response to the corporate’s testing, CloudMatrix-Infer achieves a prefill throughput of 6,688 tokens per second per processing unit, and 1,943 tokens per second when producing a response.
Consider tokens as particular person items of textual content – roughly equal to phrases or elements of phrases that the AI processes. For context, this implies the system can course of 1000’s of phrases per second on every chip.
The “TPOT” measurement (time-per-output-token) of beneath 50 milliseconds means the system generates every phrase in its response in lower than a twentieth of a second – creating remarkably quick response occasions.
Extra considerably, Huawei’s outcomes correspond to what it claims are superior effectivity rankings in contrast with competing methods. The corporate measures this by way of “compute effectivity” – basically, how a lot helpful work every chip accomplishes relative to its theoretical most processing energy.
Huawei claims its system achieves 4.45 tokens per second per TFLOPS for studying questions and 1.29 tokens per second per TFLOPS for producing solutions. In perspective, TFLOPS (trillion floating-point operations per second) measures uncooked computational energy – akin to the horsepower score of a automotive.
Huawei’s effectivity claims recommend its system does extra helpful AI work per unit of computational horsepower than Nvidia’s competing H100 and H800 processors.
The corporate reviews sustaining 538 tokens per second beneath the stricter timing necessities of sub-15 milliseconds per phrase.
Nevertheless, the spectacular numbers lack unbiased verification from third-parties, commonplace follow for validating efficiency claims within the expertise business.
Technical improvements behind the claims
The reported Huawei CloudMatrix AI efficiency metrics stem from a number of technical particulars quoted within the analysis paper. The system implements what Huawei describes as a “peer-to-peer serving structure” that disaggregates the inference workflow into three subsystems: prefill, decode, and caching, enabling every part to scale based mostly on workload calls for.
The paper posits three improvements: a peer-to-peer serving structure with disaggregated useful resource swimming pools, large-scale skilled parallelism supporting as much as EP320 configuration the place every NPU die hosts one skilled, and hardware-aware optimisations together with optimised operators, microbatch-based pipelining, and INT8 quantisation.
Geopolitical context and strategic implications
The efficiency claims emerge towards the backdrop of intensifying US-China tech tensions. Huawei founder Ren Zhengfei acknowledged lately that the corporate’s chips nonetheless lag behind US rivals “by a technology,” however stated clustering strategies can obtain comparable efficiency to the world’s most superior methods.
Nvidia CEO Jensen Huang appeared to validate this throughout a current CNBC interview, stating: “AI is a parallel drawback, so if every one of many computer systems is just not succesful… simply add extra computer systems… in China, [where] they’ve loads of vitality, they’ll simply use extra chips.”
Lead researcher Zuo Pengfei, a part of Huawei’s “Genius Youth” program, framed the analysis’s strategic significance, writing that the paper goals “to construct confidence within the home expertise ecosystem in utilizing Chinese language-developed NPUs to outperform Nvidia’s GPUs.”
Questions of verification and business impression
Past the efficiency metrics, Huawei reviews that INT8 quantisation maintains mannequin accuracy similar to the official DeepSeek-R1 API in 16 benchmarks in inner, unverified assessments.
The AI and expertise industries will seemingly await unbiased verification of Huawei’s CloudMatrix AI efficiency earlier than drawing definitive conclusions.
However, the technical approaches described recommend real innovation in AI infrastructure design, providing insights for the business, whatever the particular efficiency numbers.
Huawei’s claims – whether or not validated or not – spotlight the depth of competitors in AI {hardware} and the various approaches corporations take to attain computational effectivity.
(Picture by Shutterstock )
See additionally: From cloud to collaboration: Huawei maps out AI future in APAC
<figurewp-block-image”>
Wish to be taught extra about cybersecurity and the cloud from business leaders? Take a look at Cyber Safety & Cloud Expo happening in Amsterdam, California, and London.
Discover different upcoming enterprise expertise occasions and webinars powered by TechForge right here.