Blackwell Ultra GB300 AI 服务器预计将于 2025 年中期推出，具有“全液体”冷却和更高的性能
2024年12月18日

NVIDIA 的下一代“Blackwell Ultra”GB300 AI 服务器预计将于 2025 年中期上市，其性能将远高于现有产品。NVIDIA 的 GB300 AI 服务器将标志着“AI 潮流”的下一阶段，预计将远远超出当前的选择。

据报道，Team Green 的 Blackwell 架构引起了市场的极大兴趣，受到各主流科技巨头的追捧，为 NVIDIA 带来了巨额收入。

尽管在架构缺陷方面遭遇挫折，但 Blackwell AI 服务器的需求仍然十分强劲，而且看起来 NVIDIA 正在为即将推出的“Blackwell Ultra”系列做准备，据《台湾经济日报》报道，供应链制造商已经开始为下一代 GB300 AI 服务器做准备，该服务器的性能将比目前的产品高得多。

虽然目前有关 NVIDIA“Blackwell Ultra”架构的细节尚未公布，但据透露，这些服务器的功耗预计将比目前的 GB200 同类产品高得多。鉴于此，制造商预计将在服务器中集成“完整”的液冷解决方案，从而刺激对冷却组件的需求，并最终使 Auras Tech 和 Asia Vital Components 等台湾制造商受益。

完全使用液体冷却预计将大幅提高 GB300 AI 服务器的定价。鉴于目前 GB200 NVL72 服务器的价格约为 300 万美元，GB300 顶级配置的定价将大幅提高。如果“Blackwell Ultra”工艺成功推出，这意味着 NVIDIA 的收入将增加，最终让 Team Green 能够更多地利用 AI 潮流。

根据之前的传言，NVIDIA 的 Blackwell Ultra 系列将采用“插座式设计”，允许用户从主板上安装或卸载 GPU。此举将简化制造流程，并可能使多家公司受益，尤其是生产互连组件和插座的台湾制造商。目前的 Blackwell GPU 直接焊接到主板上，随着插座式设计的转变，B300 GPU 可以像 CPU 一样从主板上卸下。在这次产品线调整中，原先的B200 Ultra被升级为B300，GB200 Ultra变为GB300，而B200A Ultra和GB200A Ultra则分别调整为B300A和GB300A。B300系列的发布预计将在2025年第二至第三季度进行，而B200和GB200系列则预计从2024年第四季度开始出货，持续至2025年第一季度。

NVIDIA对Blackwell系列芯片的划分更为细致，旨在满足不同客户群体的需求。B300A主要针对服务器OEM客户，预计在2025年第二季度开始大规模生产。这一策略反映了NVIDIA对市场动态的灵活调整，以及对供应链产能的精准把控。

Team Green 的“B300”阵容预计将于 2025 年中期首次亮相，进入 AI 热潮的下一阶段。

文章来自：https://wccftech.com/nvidia-blackwell-ultra-gb300-ai-servers-to-debut-by-mid-2025/

? END?
【版权声明】

凡本公众平台注明来源或转自的文章，版权归原作者及原出处所有，仅供大家学习参考之用，如有侵权，烦请联系，我们立即删除。

【免责声明】

本公众平台对转载、分享的内容、陈述、观点判断保持中立，不对所包含内容的准确性、可靠性或完善性提供任何明示或暗示的保证，仅供读者参考。

?

微信扫一扫
关注该公众号

NVIDIA’s “Blackwell Ultra” GB300 AI Servers Expected To Debut By Mid-2025, Featuring “Fully-Liquid” Cooling & Much Higher Performance
Dec 2, 2024 at 06:45am EST Muhammad Zuhair

NVIDIA's "Blackwell Ultra" GB300 AI Servers Expected To Debut By Mid-2025, Featuring "Fully-Liquid" Cooling & Much Higher Performant 1
NVIDIA's next-gen "Blackwell Ultra" GB300 AI servers are expected to debut on the market by mid-2025, featuring much higher performance than existing options.

NVIDIA's GB300 AI Servers Will Mark The Next Phase Of The "AI Bandwagon", Expected To Be Much More Than Current Options
Team Green's Blackwell architecture has reportedly seen massive market interest and is sought by every mainstream tech giant, bringing in gigantic revenue for NVIDIA.

RELATED STORY Inno3D Official Confirms CES 2025 To Be The Launch Date For RTX 50 Series GPUs
Despite setbacks in terms of architectural flaws, Blackwell AI server demand is in full force, and it looks like NVIDIA is now preparing for the upcoming "Blackwell Ultra" lineup, as a report by Taiwan Economic Daily states that supply chain manufacturers have started preparing for the next-gen GB300 AI servers, which are set to feature significantly higher performance than current options.

While details surrounding NVIDIA's "Blackwell Ultra" architecture are confined for now, it is revealed that the servers are expected to feature significantly higher power consumption figures compared to their current GB200 counterparts. In light of this, manufacturers are expected to integrate a "complete" liquid-cooled solution with the servers, fueling demand for cooling components and ultimately benefiting Taiwanese manufacturers such as Auras Tech and Asia Vital Components.

The use of liquid cooling entirely is expected to raise the pricing of the GB300 AI servers massively. Given that the current GB200 NVL72 servers cost around $3 million, it is imminent that the pricing of the GB300's top configuration will be much higher. If the "Blackwell Ultra" process is a successful launch, this means more revenue for NVIDIA, ultimately allowing Team Green to capitalize much more on the AI bandwagon.

In terms of previous rumors, NVIDIA's Blackwell Ultra lineup is said to adopt a "socketed design," which will allow the users to install or uninstall the GPUs from the motherboards. This move will be done to simplify the manufacturing complex and could benefit several companies, especially the Taiwan-based manufacturers, which produce interconnect components and sockets. The current Blackwell GPUs are soldered directly to the motherboard, and with the transition to a socket-type design, the B300 GPUs can be removed from the motherboard just like CPUs.

关于我们

北京汉深流体技术有限公司 是丹佛斯中国数据中心签约代理商。产品包括FD83全流量双联锁液冷快换；液冷通用快速接头UQD & UQDB；OCP ORV3盲插快换接头BMQC；EHW194 EPDM液冷软管、电磁阀、压力和温度传感器。在国家数字经济、东数西算、双碳、新基建战略的交汇点，公司聚焦组建高素质、经验丰富的液冷工程师团队，为客户提供卓越的工程设计和强大的客户服务。

公司产品涵盖：丹佛斯液冷流体连接器、EPDM软管、电磁阀、压力和温度传感器及Manifold。
未来公司发展规划：数据中心液冷基础设施解决方案厂家，具备冷量分配单元（CDU）、二次侧管路（SFN）和Manifold的专业研发设计制造能力。

- 针对机架式服务器中Manifold/节点、CDU/主回路等应用场景，提供不同口径及锁紧方式的手动和全自动快速连接器。
- 针对高可用和高密度要求的刀片式机架，可提供带浮动、自动校正不对中误差的盲插连接器。以实现狭小空间的精准对接。
- 基于OCP标准全新打造的液冷通用快速接头UQD & UQDB ;OCP ORV3盲插快换接头BMQC , 支持全球范围内的大批量交付。