What are the main functions of the KV cache during inference for an AI foundation model?
Select all that apply.
Significantly improve inference efficiency
Improve the utilization of HBM
Improve computing power utilization
Reduce repeated calculations in the inference steps
Exact Extract — Huawei Training Material, Pages 92, 167 and 168
"High memory usage for long sequences"
"The longer the sequence length (N + m, input + output), the higher the memory usage and the greater the inference resource consumption."
"Repeated computation of a large number of historical tokens results in high TTFT"
"Identify and match historical prefix cache by the unique hash code to eliminate repeated computing via querying and improve the inference throughput by 43%"
The inference architecture also distinguishes the prefill stage from incremental decoding, during which previously generated KV-cache information is reused.
Technical Deep Dive:
The Key-Value (KV) cache is a fundamental optimization technique used during autoregressive Transformer inference.
In Transformer attention, each token is represented by query, key, and value vectors. During autoregressive text generation, the model generates tokens sequentially and repeatedly attends to previously processed tokens.
Without KV caching, previously calculated attention key and value representations would need to be recomputed unnecessarily during subsequent decoding steps.
The KV cache stores these representations so that they can be reused.
Option A is correct. Reusing previously calculated attention states reduces redundant computation and improves inference execution efficiency.
Option D is correct. Avoiding repeated calculations is the principal mechanism by which KV caching accelerates autoregressive inference.
Option B is not a direct function of KV caching. HBM provides high-bandwidth memory for model weights, activations, and cached attention states. KV caching consumes HBM capacity rather than inherently increasing its utilization efficiency.
Option C is not a primary guaranteed function. KV caching can improve useful computational throughput, but its effect on hardware utilization depends on memory bandwidth, concurrency, and the inference scheduler.
For long-sequence workloads, the KV cache can become a significant memory bottleneck.
Huawei addresses this through hierarchical caching, memory compression, and cache offloading in its inference acceleration architecture.
The primary functions are therefore reducing repeated computation and improving inference efficiency.
What are the primary features of the liquid-cooling solution in new AI equipment-room scenarios?
Select all that apply.
Improved computing power per unit of electricity consumed
1 week go-live
PUE=1.6
Low Capex investment
Exact Extract — Huawei Training Material, Page 241
Under the new-build liquid-cooling data-center solution, the listed features are:
"Energy saving (PUE ≤ 1.15)"
"Adapting to high-density chip evolution"
Technical Deep Dive:
Liquid cooling is an important data-center infrastructure technology for deploying high-density AI computing systems, particularly clusters containing multiple power-intensive Ascend accelerators.
Conventional air cooling can become inefficient when rack power density increases significantly. Liquid cooling transfers heat using liquid coolant, which generally provides more effective heat removal than relying entirely on air circulation.
This enables increased computing density and can reduce cooling-related energy consumption.
Option A is correct. Improved cooling efficiency allows more computing resources to operate within a given facility power and thermal envelope, increasing effective computing capacity per unit of electricity.
Option B is not supported for the specified new-build scenario. The one-week deployment capability described elsewhere in the training concerns integrated computing and networking cluster rollout, not the complete construction and commissioning of a new liquid-cooled AI equipment room.
Option C is incorrect. A PUE of 1.6 is substantially above the PUE ≤ 1.15 target identified for the liquid-cooling solution.
Option D is not a stated feature of the new-build liquid-cooling scenario. Low initial CAPEX is discussed for other deployment scenarios, particularly leasing and certain reconstruction approaches. New liquid-cooling infrastructure can require significant initial investment in CDUs, liquid distribution, manifolds, and cooling equipment.
The primary verified feature is therefore improved energy efficiency and higher computing density.
Which aspects of intelligent-computing data centers impose high requirements on O & M assurance?
Select all that apply.
Cluster training is interrupted due to a single device fault.
The performance of a single node in the cluster affects the overall performance.
Training of trillion parameter models is a large-scale application for the entire system, involving a large number of components.
An intelligent computing server has eight cards, and a single card failure does not affect the training.
Exact Extract — Huawei Training Material, Pages 57 and 69
"Large scale does not mean strong computing power, and building large computing clusters faces many challenges"
"Cluster availability"
"Frequent failures of core components"
"Difficult cross-domain fault handling"
"Slow recovery of training tasks"
"Cross-domain fault detection and diagnosis for compute, network, and storage"
"High throughput checkpoint access and storage"
"System high availability = as long as possible failure-free running time + as short as possible fault recovery time"
The resumable-training architecture also identifies fault detection, task rescheduling, checkpoint loading, collective communication initialization, and training resumption as essential recovery activities.
Technical Deep Dive:
Large-scale AI training environments contain numerous interconnected components, including Ascend NPUs, Kunpeng CPUs, server memory, network switches, optical transceivers, and distributed storage systems.
Training workloads frequently use synchronous distributed computation, making application performance and availability dependent on the health of multiple devices.
Option A is correct. A single critical hardware failure can interrupt a distributed training job when the affected device participates in required computation or synchronization.
Fault-tolerant training mechanisms can reduce the impact, but uninterrupted execution is not guaranteed.
Option B is correct. Distributed training performance can be constrained by a slow or degraded node.
When synchronization is required, faster nodes may wait for slower participants before continuing to the next training step.
This behavior is commonly known as the straggler effect .
Option C is correct. Trillion-parameter models require large-scale distributed computing infrastructure containing many interdependent hardware and software components.
The resulting complexity creates substantial requirements for fault monitoring, root-cause analysis, and operational coordination.
Option D is incorrect. Failure of a single accelerator card can affect the entire training workload, especially when the device holds a required portion of the model or participates in collective communication.
Huawei addresses these challenges through CCAE cluster management, checkpoint recovery, hardware fault prediction, and automated training resumption.
Therefore, A, B, and C are correct.
How long is typically needed to pre-train a model?
More than 1 month
3 weeks
2 weeks
1 week
Exact Extract — Huawei Training Material, Page 29
"The training duration of a large model is proportional to the number of model parameters and the volume of training data"
The training-duration examples include approximately 49 days for a 175-billion-parameter model, 35 days for a 110-billion-parameter model, and 34 days for a 13-billion-parameter model under their respective hardware and dataset assumptions.
Technical Deep Dive:
Pre-training is the computationally intensive process through which a foundation model learns statistical patterns, language relationships, and general representations from large datasets.
Unlike fine-tuning, which adapts an existing model, pre-training normally begins with randomly initialized model weights and involves repeated forward propagation, backpropagation, and parameter updates over extremely large token datasets.
The total training duration depends primarily on model parameter count, training-token volume, effective accelerator throughput, cluster size, and distributed training efficiency.
Huawei's training sizing methodology expresses the theoretical end-to-end training duration as:
E_t = 8TP/(nX)
Here, T represents the training-token volume, P the parameter count, n the number of accelerator cards, and X the effective computing performance per card, using consistent units.
Training time increases with the computational workload and decreases as effective available computing power increases.
Other factors include network communication overhead, checkpoint frequency, storage throughput, accelerator utilization, and failures requiring recovery.
The examples establish that large-scale pre-training frequently requires more than a month.
Technical qualification: This is the intended answer for the training examples, not a universal minimum. The same material includes a 27-day example for a 65-billion-parameter model. Smaller models or larger compute clusters can complete pre-training substantially faster.
For new medium- and large-sized DC scenarios, Huawei offers PowerPOD, SmartLi, UPS power supplies, and CDU, EHU, and air-cooling-wall cooling solutions.
TRUE
FALSE
Exact Extract — Huawei Training Material, Pages 238 and 241
"AI data center facility: modular, standard, and elastic deployment, enabling quick deployment and elastic evolution"
"Standard POD with cooling, power supply, battery, monitoring, auxiliary room, etc"
"Power supply module"
"Battery room"
"Liquid-cooling module"
"Air-cooling module"
"AI data center solution reference: five scenarios solution to meet low PUE and quick rollout requirements"
These extracts establish the integrated power, battery, air-cooling, and liquid-cooling design principles. They do not enumerate all six named product technologies in the question.
Technical Deep Dive:
Huawei provides integrated data-center infrastructure technologies intended to meet the power supply, cooling, and deployment requirements of medium- and large-scale facilities.
PowerPOD is a prefabricated power supply and distribution architecture. It integrates major electrical infrastructure components and supports modular deployment.
SmartLi provides lithium-battery-based energy storage and backup power support for compatible UPS systems.
UPS systems protect IT equipment from interruptions and disturbances in incoming electrical power.
CDUs , or Coolant Distribution Units, are used in liquid-cooling systems to manage coolant circulation and heat transfer.
EHU refers to indirect evaporative cooling technology within Huawei's data-center cooling portfolio.
Air-cooling walls , also called fan-wall cooling systems, provide large-scale controlled airflow and heat removal for suitable data-center environments.
The combined objective is to provide reliable power delivery, efficient thermal management, and scalable infrastructure deployment.
Huawei's official data-center facility portfolio confirms PowerPOD, SmartLi, UPS systems, and fan-wall cooling. Its infrastructure publications also identify EHU and liquid-cooling components as supported technologies.
However, individual data-center configurations do not necessarily require every listed component. The final design depends on rack power density, power redundancy, facility location, cooling strategy, and customer requirements.
The statement is correct because these are technologies Huawei offers for the stated data-center scenarios.
The TaiShan 200 5280 server supports less storage capacity than the TaiShan 200 2280 server.
TRUE
FALSE
Exact Extract — Huawei Training Material, Page 123
"TaiShan 200 Server Models Overview"
"Balanced — 2280 — 2U2-way — Big data, distributed storage, etc — Balanced computing and network scalability"
"Storage — 5280 — 4U2-way 40 disks — Distributed Storage — 560 TB local storage capacity"
"Storage — 5290 — 4U72 disks — Distributed Warm and Cold Storage (Video Storage and Computing) — Local storage capacity exceeding 1 PB"
Technical Deep Dive:
The Huawei TaiShan 200 server family offers different configurations optimized for computing performance, storage density, and workload requirements.
The TaiShan 200 2280 is a balanced 2U, dual-processor server intended for workloads requiring an appropriate combination of computing resources, memory capacity, storage, and network connectivity.
The TaiShan 200 5280, however, is specifically designed for storage-intensive applications and provides substantially greater local drive capacity.
According to Huawei's server specifications, the TaiShan 200 5280 supports up to 40 3.5-inch SAS/SATA HDDs or SSDs, together with four 2.5-inch NVMe SSDs.
By comparison, the TaiShan 200 2280 supports configurations of up to 16 3.5-inch drives, 27 2.5-inch SAS/SATA drives, or 16 NVMe SSDs.
The higher drive density of the 5280 makes it suitable for distributed storage platforms, software-defined storage, large-scale datasets, and other capacity-intensive applications.
Its 4U chassis provides the physical space required for the larger drive configuration.
Although actual usable capacity depends on installed drive sizes, RAID or erasure-coding configuration, and storage overhead, the 5280 is the storage-optimized model with greater supported local drive capacity.
The statement is therefore false. The product specifications are also consistent with Huawei's published TaiShan Server Technical Specifications.
The TaiShan 200 2280 is a 2U2P server built for balanced performance.
TRUE
FALSE
Exact Extract — Huawei Training Material, Page 123
The TaiShan 200 Server Models Overview identifies the following specifications for the 2280 configuration:
Type: Balanced
Model: 2280
Specifications: 2U2-way
Scenario: Big data, distributed storage, etc.
Features: Balanced computing and network scalability
The product family also emphasizes reliable design, high-speed interconnection, and carrier-class components.
Technical Deep Dive:
Huawei TaiShan 200 2280 is a two-rack-unit, two-processor server intended for balanced enterprise computing workloads.
The designation 2U2P describes two important hardware characteristics.
2U indicates that the server occupies two rack units in a standard server cabinet.
2P indicates support for two physical processor sockets.
The dual-processor architecture enables the system to provide substantial computing resources while maintaining a compact rack-mount form factor.
The TaiShan 200 2280 is positioned for workloads requiring balanced processor performance, memory capacity, storage accessibility, and network scalability.
Examples include big-data processing platforms, distributed storage services, enterprise applications, and virtualization environments.
A balanced server design must deliver appropriate performance across multiple subsystems rather than concentrating exclusively on computing density or storage capacity.
Memory bandwidth, PCIe connectivity, network interfaces, and processor efficiency all influence overall application performance.
TaiShan servers based on Kunpeng processors also benefit from ARM-compatible operating systems and optimized application software.
When planning deployments, engineers should validate application requirements against the exact server configuration, including processor generation, memory capacity, disk configuration, expansion options, and software support.
The training material explicitly identifies the TaiShan 200 2280 as a balanced 2U, two-way server.
Therefore, the statement is true.
Which of the following statements about the Atlas 800I A2 inference server are accurate?
Select all that apply.
The overall inference throughput is higher, and the user experience of latency is better.
Recommended for non-latency-sensitive scenarios
8 NPUs per server, full-mesh high-bandwidth interconnection
Supports models of hundreds of billions of parameters and long-sequence inference.
Exact Extract — Huawei Training Material, Page 90
"Latency-sensitive scenarios: All-connect module server products meet performance requirements"
"8 devices per server, full-mesh interconnection, and high bandwidth"
"Supports models of hundreds of billions of parameters and long-sequence inference."
"The overall inference throughput is higher, and the user experience in terms of latency is better."
The same product comparison identifies the Atlas 800I A2 as a 4U, eight-module inference server and illustrates an interconnection bandwidth of 392 GB/s.
Technical Deep Dive:
The Huawei Atlas 800I A2 is an AI inference server designed for demanding generative AI and large language model workloads.
Its architecture integrates eight Ascend AI processing modules and provides high-bandwidth communication between the accelerators.
This design enables models to be distributed across multiple processing devices while minimizing the communication overhead associated with tensor parallelism and other distributed inference techniques.
Option A is correct. High-bandwidth interconnection and optimized inference execution improve aggregate token-processing throughput and help reduce response latency under suitable workloads.
Option B is incorrect. The Atlas 800I A2 is specifically positioned for latency-sensitive applications requiring high-performance multi-device inference. Standard inference cards are positioned for more cost-sensitive, less latency-critical deployments.
Option C is correct after terminology correction. The original question uses the term GPUs , but the Atlas 800I A2 uses Huawei Ascend AI processors, commonly referred to as NPUs. These devices are interconnected using high-bandwidth communication technology, including HCCS-based interconnection in supported configurations.
Option D is correct. The multi-module architecture provides substantial aggregate computing and memory resources, supporting deployment of large language models containing hundreds of billions of parameters and long input sequences.
Such deployments require careful consideration of model parallelism, KV-cache memory, device communication, and serving concurrency.
Huawei's MindIE inference engine can provide additional optimization through model scheduling, continuous batching, and inference acceleration.
Terminology correction: Option C originally states “8 GPUs per server.” This has been corrected to “8 NPUs per server” to match the Ascend hardware architecture.
Large scale does not necessarily equate to high computing power. What challenges arise in constructing high-computing-power clusters?
Select all that apply.
Cluster scale
Development Ease of Use
Cluster Availability
Cluster performance
Exact Extract — Huawei Training Material, Page 57
"Large scale does not mean strong computing power, and building large computing clusters faces many challenges"
The four challenge categories identified are cluster size, cluster performance, cluster availability, and development ease of use.
Technical Deep Dive:
Building a high-performance AI computing cluster requires more than increasing the number of accelerators or servers.
Option A is correct — Cluster scale. Large AI training environments may require thousands or tens of thousands of accelerators. Expanding cluster capacity introduces networking, power distribution, storage scalability, and infrastructure management challenges.
Option B is correct — Development ease of use. Developers require efficient frameworks, debugging tools, operator libraries, and performance optimization capabilities. Complex programming requirements can delay application deployment and prevent full utilization of computing resources.
Option C is correct — Cluster availability. Large clusters contain substantial numbers of interconnected components, increasing the probability that an individual component will fail during a long-running training job.
Fault tolerance, checkpoint recovery, monitoring, and cross-domain troubleshooting are therefore essential to maintaining training continuity.
Option D is correct — Cluster performance. The theoretical computing capacity of an accelerator cluster does not automatically translate into equivalent application performance.
Communication overhead, load imbalance, storage bottlenecks, synchronization delays, and inefficient parallelization can significantly reduce effective throughput.
Huawei addresses these challenges through high-performance Ascend processors, distributed acceleration tools such as MindSpeed, high-bandwidth networking, optimized AI storage, and centralized CCAE management.
Accordingly, all four options represent fundamental challenges in high-computing-power cluster construction.
Which statement about the new TaiShan 200 model server is correct?
Up to 10 PCIe slots
Supports 920 processors
Does not support NVMe SSDs
Temperature range: 5°C to 35°C
Exact Extract — Huawei Training Material, Page 125
"2U2P new model server specifications"
"Form Factor — 2U2-way"
"CPU — 2 x 920 processors, up to 64 cores and 2.9 GHz"
"Memory — 32 x DDR5 DIMMs, up to 4800 MT/s, 4 TB maximum memory (128 GB memory configuration)"
"PCIe Expansion — Up to 10 PCIe slots"
"Temperature Range — 5°C to 35°C"
The storage configurations additionally include 16 or 24 2.5-inch NVMe drives, subject to the specified compatible SAS/SATA drive configurations.
Technical Deep Dive:
Huawei's new-generation TaiShan 200 2U2P server combines Kunpeng computing resources with high-capacity memory, flexible storage, and expanded I/O connectivity.
Option A is correct. The specified 2U2P model supports up to 10 PCIe expansion slots. PCIe connectivity enables the installation of networking, storage, and other compatible expansion devices, allowing the server to be adapted for different workloads.
Option B is correct. The server supports two Kunpeng 920 processors. These ARM-based processors are intended for enterprise computing applications requiring multi-core processing, scalability, and energy-efficient operation.
Option C is incorrect. NVMe SSD support is explicitly included. NVMe storage uses PCIe connectivity to deliver low-latency and high-throughput storage access, making it particularly valuable for databases and data-intensive applications.
Option D is correct. The specified model has an operating temperature range of 5°C to 35°C.
This range is important when planning cooling capacity and selecting an appropriate data-center deployment environment.
The server also supports DDR5 memory, providing a higher-bandwidth memory platform than earlier DDR4-based configurations.
Critical question-quality correction: The original question is framed as single-choice, but A, B, and D are all explicitly correct for the 2U2P model on page 125. It must be converted into a multiple-answer question to be technically valid. Selecting D alone, as implied by the original question explanation, would incorrectly exclude two supported specifications.
