英伟达已经调整下一代Vera Rubin人工智能计算平台的部分硬件规格，包括HBM4高带宽显存带宽以及NVLink高速互连带宽。与此前公布或流传的指标相比，新规格有所下降，但横向扩展网络带宽则出现提升。这一变化可能意味着，英伟达正在根据实际量产情况重新平衡芯片性能、供应能力和系统成本。
根据英伟达目前公布的Vera Rubin平台规格，Rubin GPU配备288GB HBM4显存，显存带宽为19.2TB/s。此前，英伟达的技术资料曾给出最高22TB/s的显存带宽指标。两者相比，当前公布的数值低约12.7%。需要注意的是，英伟达此前使用的是“最高可达”指标，因此这既反映了规格数字的变化，也不一定意味着所有产品的实际性能都出现了同等幅度的下降。
HBM4是新一代高带宽显存技术，采用更宽的内存接口，为人工智能模型训练和推理提供更高的数据吞吐能力。随着模型规模不断扩大，GPU不仅需要更强的计算能力，还必须持续获得大量模型参数和中间数据。如果显存带宽不足，即使计算单元拥有很高的理论算力，也可能因等待数据而无法充分发挥性能。
Rubin平台此前追求的高显存带宽，也给HBM4供应链带来了较高的技术要求。显存制造商需要在较高的数据传输速率下，同时满足功耗、散热、信号完整性和长期可靠性要求。带宽指标越高，设计验证与量产过程面临的挑战通常也越大。
如果英伟达降低显存带宽要求，符合产品标准的HBM4芯片范围可能随之扩大，供应商也有机会提高合格产品的产出比例。这并不意味着显存制造商必须重新设计整套产品，而是可能让部分无法达到此前最高性能目标、但仍满足新要求的产品进入供应体系，从而改善整体供货能力。
除了显存带宽，英伟达还调整了Vera Rubin超级芯片的NVLink互连规格。根据目前的产品资料，Vera Rubin超级芯片的NVLink带宽从此前的7.2TB/s调整为6TB/s，降幅约16.7%。NVLink是英伟达用于GPU之间高速通信的互连技术，对于多GPU系统中的数据交换、模型并行训练以及复杂人工智能推理任务具有重要作用。
在大型人工智能集群中，GPU之间需要频繁交换梯度、激活值以及其他中间数据。对于采用专家混合架构的大型模型而言，不同计算单元之间的通信尤其重要。因此，降低NVLink带宽指标可能影响某些通信密集型任务的性能，但最终影响程度仍取决于模型架构、软件调度方式以及实际负载，不能简单地按照带宽下降比例推算整体算力损失。
与此同时，英伟达提高了部分横向扩展网络的带宽指标。现有资料显示，Vera Rubin超级芯片的横向扩展网络带宽从此前的0.8TB/s提高至0.9TB/s。这类网络主要用于连接不同计算节点，使数据中心能够将大量GPU组织成统一的计算集群。提高网络带宽有助于改善节点之间的数据传输能力，尤其是在计算任务需要跨服务器协同完成时。
这种一降一升的调整，体现了人工智能计算平台设计中的权衡关系：GPU内部及机柜内的通信能力固然重要，但随着集群规模扩大，跨节点通信也可能成为限制整体性能的因素。英伟达可能正在根据实际应用需求，重新分配系统设计中的性能预算。
不过，规格变化本身并不能证明英伟达已经确认存在显存短缺，也不能单独证明降低指标就是为了提高产量。相关解释仍需要结合供应商量产进度、产品认证情况以及英伟达后续披露的信息进行判断。
对于HBM4供应链而言，规格调整可能改变供应商之间的竞争重点。此前，制造商需要争取达到更高的数据传输速度；如果英伟达降低量产产品的带宽要求，竞争可能更加集中于良率、供货规模、功耗、稳定性和成本。能够大规模交付合格产品的供应商，可能因此获得更多机会。
从整个平台来看，Vera Rubin并非单纯依靠显存带宽决定性能。GPU计算能力、显存容量、芯片间互连、网络架构以及软件优化共同决定了系统表现。英伟达官方公布的Vera Rubin NVL72机柜配置包含72颗Rubin GPU，并通过第六代NVLink连接多个GPU，以支持大规模人工智能训练和推理任务。
因此，即使部分硬件指标低于早期目标，Vera Rubin仍可能凭借计算架构、显存容量和系统级优化取得较高性能。真正值得关注的是，英伟达能否在规格调整后按计划扩大出货规模，以及实际运行中的模型训练和推理性能是否符合客户预期。
本文由 AI图鉴自动抓取并翻译，版权归原作者所有；如需查看原文请点下方链接。
Young Weida has adapted some of the hardware specifications for the next generation of Vera Rubin's artificial intelligence computing platform, including HBM4 high bandwidth VRAM bandwidth and NVLink high-speed bandwidth. Compared to previously published or disseminated indicators, the new specifications have decreased, but the horizontal extension of network bandwidth has increased. This change may mean that Weeda is rebalancing chip performance, supply capacity and system costs according to actual production.
According to the specifications of the Vera Rubin platform, currently published in Weida, Rubin GPU is equipped with 288GB HBM4Z82Z and VRAM bandwidth of 19.2TB/s. Previously, technical information from In Weida had given VRAM bandwidth indicators up to 22 TB/s. Compared to the two, the currently published value is about 12.7 per cent lower. It is important to note that Weeda had previously used the “highest achievable” indicator, thus reflecting both a change in the number of specifications and not necessarily an equivalent decline in the performance of all products.
HBM4 is a new generation of high-bandwidth VRAM technology that uses a wider memory interface to provide higher data-admission capability for artificial intelligence model training and reasoning. As the size of the model expands, the GPU needs not only greater computing capacity, but also continued access to a large amount of model parameters and intermediate data. If VRAM bandwidth is insufficient, even if the computing unit has high theoretical strength, it may be unable to fully perform because it waits for data.
The high VRAM bandwidth previously sought by the Rubin platform also brought higher technical requirements to the HBM4 supply chain. VRAM manufacturers need to meet the requirements of power consumption, heat dispersion, signal integrity and long-term reliability at a higher data transmission rate. The higher the bandwidth indicator, the greater the challenges associated with the design validation and volume process.
If Young Weida lowers the VRAM bandwidth requirements, the scope of HBM4 chips that meet product standards may be expanded, and suppliers may also have the opportunity to increase the proportion of output from qualified products. This does not mean that the manufacturer of VRAM has to redesign the product package, but may bring into the supply system parts of the product that were unable to meet the previous maximum performance targets but still meet new requirements, thereby improving overall supply capacity.
In addition to the VRAM bandwidth, the NVLink interconnectivity specification of the Vera Rubin Superchip was adjusted by In Weida. According to current product information, the NVLink bandwidth of the Vera Rubin Superchip was adjusted from the previous 7.2TB/s to 6TB/s, a decrease of about 16.7 per cent. NVLink is an interconnective technology for high-speed communication between GPUs in Weidar and plays an important role in data exchange in multi-GPU systems, in parallel training in models and in complex artificial intelligence reasoning missions.
In large artificial intelligence clusters, frequent exchange of gradients, activation values and other intermediate data is required between GPUs. Communication between different computing units is particularly important for large models using a mixed structure of experts. Thus, the reduction of the NVLink bandwidth indicator may affect the performance of some communications-intensive missions, but the ultimate impact will depend on the model structure, the way the software is deployed and the actual load, and the overall calculation loss cannot be simply calculated by applying the rate of bandwidth reduction.
At the same time, Weida has raised the bandwidth indicator for a partial horizontal extension of the network. Available information indicates that the horizontal extension network bandwidth of the Vera Rubin Superchip increased from the previous 0.8 TB/s to 0.9 TB/s. Such networks are primarily used to connect different computational nodes, enabling data centres to organize large numbers of GPUs into uniform computing clusters. Increased network bandwidth will help improve data transmission between nodes, especially when computing tasks require inter-server teamwork.
This one-litre adjustment reflects the trade-off in the design of artificially intelligent computing platforms: While communication capacity within the GPU and in the cabinet is important, cross-node communication can also be a limiting factor to overall performance as clusters expand. Weeda may be reallocating the performance budget in the design of the system based on actual application needs.
However, the change in specifications does not in itself prove that Young Wida has identified a shortage of VRAM, nor does it prove alone that the lower target was intended to increase production. The explanations still need to be judged in the light of the supplier's production schedule, product certification and subsequent disclosures by Yvette.
For the HBM4 supply chain, specification adjustments may change the focus of competition among suppliers. Previously, manufacturers had to strive for a higher rate of data transmission; competition might be more concentrated on good rates, the size of the supply, the amount of effort, stability and costs if British Wida reduced bandwidth requirements for the mass products. Suppliers capable of delivering qualified products on a large scale may thus have more opportunities.
From the platform as a whole, Vera Rubin does not rely solely on VRAM bandwidth for decisive energy. System performance is determined jointly by GPU computing capabilities, VRAM capacity, chip interconnection, network architecture and software optimization. The official Vera Rubin NVL72 cabinet set up with 72 Rubin GPUs and linked several GPUs through the sixth generation of NVLink to support large-scale artificial intelligence training and reasoning missions.
Thus, even if some of the hardware indicators fall short of the early targets, Vera Rubin may still achieve higher performances through computing architecture, VRAM capacity and system-level optimization. There is a real concern about whether Ying Wei Da will be able to scale up deliveries as planned after the specifications have been adjusted, and whether model training and reasoning performance in practice meet client expectations.
Auto-fetched and translated by AI Atlas. All rights belong to the original author.