11-19 更新 实时
首页/ 算力/ 正文

Bull 在 SC26 展示 Intel Xeon Diamond Rapids 封装

算力 来源:ServeTheHome 2026-11-19 01:21 0 阅读 机器翻译
Bull 在 SC26 展示 Intel Xeon Diamond Rapids 封装

Bull SC26 芝加哥-Bull Sequana XH31600 Intel Diamond Rapids 封装 3

Intel 正在对其下一代服务器 CPU 进行一次小型发布(或者说在 SC26 上算是差不多的一次发布)。展馆现场有许多系统,但其中一台 Bull Sequana XH31600 真正展示了 Diamond Rapids 封装。

Bull 在 SC26 上展示 Intel Xeon Diamond Rapids 封装

先来看这块 Bull 计算刀片:后部是盲插供电与液冷接口,前部是网络接口,正面则有四个 Intel Xeon Diamond Rapids 封装。

Bull SC26 芝加哥-Bull Sequana XH31600 2

需要说明的是,你看到的是两个搭载 Intel Diamond Rapids 的双路计算节点,每个节点配备 16 通道内存。

Bull SC26 芝加哥-Bull Sequana XH31600 1

这是该封装的视图。我们此前已在 OCP Summit 2026 上、ASRock Rack DMRD16-2L+ 中见过 LGA9324 插槽。

Bull SC26 芝加哥——Bull Sequana XH31600 Intel Diamond Rapids 封装 4

可以看到,在 Bull 平台上,CPU 位于相当宽的主板上,因为主板两侧都布满了内存。

Bull SC26 芝加哥——Bull Sequana XH31600 9

其设计思路是这些刀片采用液冷,因此可以看到所有部件都配有热管和液冷块。

Bull SC26 芝加哥——Bull Sequana XH31600 2

另一个节点同样配有液冷块,以及通往网络的线缆连接。

Bull SC26 芝加哥——Bull Sequana XH31600 5

在网络方面,Bull 采用其 BXI 技术实现互连。在展台上,Bull 展示了一台配备 64 个 800G OSFP 端口的 BXI 交换机,但 BXI 网卡标注的却是 QSFP-DD 和 PCIe Gen5 x16。实际上,Bull 的做法看起来与我们在 NVIDIA ConnectX-8 C8240 800G Dual 400G NIC Review 中所做的类似——这是一款 SuperNIC,当时我们设法用两条 PCIe Gen5 x16 链路从 ConnectX-8 上获得了 800Gbps。

Bull SC26 Chicago-Bull Sequana XH31600 12

稍微引人注意的是,其中还有一块 OCP NIC 3.0 规格的 NVIDIA ConnectX-8 SuperNIC。有意思的是,Bull 拥有自己的互连方案,在一套搭载 Gen6 CPU 的平台上采用了 51.2T 交换机和 PCIe Gen5 网卡,但紧挨着它的却是 NVIDIA 大批量出货的 PCIe Gen6 网卡竞品。

Bull SC26 Chicago-Bull Sequana XH31600 13

这个平台非常酷的一点在于,它是高度定制的。

Bull SC26 Chicago-Bull Sequana XH31600 14

多年来,Bull 的刀片设计一直算是外观较为出色的一类,主要原因是他们把一切都做了液冷。不过,或许值得思考的是,NVIDIA 正以 ORv3 推动整个行业前行,而 AMD 现在也在推双宽(double-wide)方案。给人的感觉是,如果你想获得供电和散热能力,行业里已经有了一套面向高密度机架的大批量标准。也许更有意思的一点是,Bull 的设计在机身后部没有采用铜缆背板。

Bull SC26 Chicago-Bull Sequana XH31600 15

总体而言,这次看到的东西非常值得一看。

SC26 上令人感兴趣的一点是,这类平台在几年前还会让人惊叹不已。如今,随着 NVIDIA Vera Rubin 设计进入 ORv3 机架,Bull 系统已融入更大批量的 AI 基础设施组件之列。不过,Bull 的看点在于,SiPearl、Intel Xeon Diamond Rapids、NVIDIA 和 AMD Venice 节点以统一的外形规格同台展示。

Bull SC26 芝加哥——Bull Sequana XH31600 Intel Diamond Rapids 包装 1

许多其他厂商也在 SC26 上展示设计精巧的 Diamond Rapids 系统,包括每插槽 32DIMM 的 2DPC 配置。就 Bull 而言,它是我们发现的唯一一家有实物包装的厂商。Diamond Rapids 似乎正处于送样阶段。最合理的猜测是,我们或许会在 2027 年第二季度开始看到它发布,但如果这一时间推迟到 6 月之后(或者变成 1P 平台与 2P 平台分批发布),Intel 下一个真正的发布窗口就是 2027 年 9 月初。

译文由英文原文机器翻译,术语按行业惯例统一,可在文末展开原文核对。

查看英文原文(Bull Shows Intel Xeon Diamond Rapids Package at SC26)

Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 3Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 3

Intel is doing a mini launch of its next-generation server CPU (or something thereabouts at SC26). There are many systems on the floor, but one, the Bull Sequana XH31600, actually showed Diamond Rapids packages.

Bull Shows Intel Xeon Diamond Rapids Package at SC26

Starting off, here is the Bull compute blade with blind mate power and liquid at the rear, networking in front, and four Intel Xeon Diamond Rapids packages up front.

Bull SC26 Chicago-Bull Sequana XH31600 2Bull SC26 Chicago-Bull Sequana XH31600 2

Just ot be clear about what you are seeing, these are two dual-socket compute nodes with Intel Diamond Rapids, each with 16-channel memory.

Bull SC26 Chicago-Bull Sequana XH31600 1Bull SC26 Chicago-Bull Sequana XH31600 1

Here is a view of the package. We have already seen the LGA9324 socket covered in the ASRock Rack DMRD16-2L+ from OCP Summit 2026.

Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 4Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 4

You can see that in the Bull platform, the CPUs are on fairly wide motherboards because of all of the memory on either side.

Bull SC26 Chicago-Bull Sequana XH31600 9Bull SC26 Chicago-Bull Sequana XH31600 9

The idea is that these blades are liquid cooled, so you can see how everything has heatpipes and liquid cooling blocks.

Bull SC26 Chicago-Bull Sequana XH31600 2Bull SC26 Chicago-Bull Sequana XH31600 2

The other node had the liquid cooling blocks as well as the cabled connections to the networking.

Bull SC26 Chicago-Bull Sequana XH31600 5Bull SC26 Chicago-Bull Sequana XH31600 5

On the networking, Bull had its BXI technology for interconnect. In its booth, Bull showed a BXI switch with 64x 800G OSFP ports, but the BXI NICs were shown as QSFP-DD and PCIe Gen5 x16. It actually looks like Bull is doing something similar to what we did with the NVIDIA ConnectX-8 C8240 800G Dual 400G NIC Review This is a SuperNIC when we managed to use two PCIe Gen5 x16 links to get 800Gbps from the ConnectX-8.

Bull SC26 Chicago-Bull Sequana XH31600 12Bull SC26 Chicago-Bull Sequana XH31600 12

What stood out a bit is that there was an NVIDIA ConnectX-8 SuperNIC as well in an OCP NIC 3.0 form factor. It’s interesting that Bull has its own interconnect, with a 51.2T switch and a PCIe Gen5 NIC in a platform with a Gen6 CPU, but right next to it is NVIDIA’s high-volume PCIe Gen6 NIC competitor.

Bull SC26 Chicago-Bull Sequana XH31600 13Bull SC26 Chicago-Bull Sequana XH31600 13

What is very cool about this platform is that it is incredibly bespoke.

Bull SC26 Chicago-Bull Sequana XH31600 14Bull SC26 Chicago-Bull Sequana XH31600 14

For years, Bull has had one of the better-looking blade designs, largely because they liquid-cool everything. Probably worth pondering, however, is that with NVIDIA pushing the industry with ORv3, and AMD now with the double-wide. It feels like there is a high-volume standard for high-density racks in the industry if you want power and cooling. Maybe the more interesting part is that Bull’s design does not have a copper-cabled backplane in the rear.

Bull SC26 Chicago-Bull Sequana XH31600 15Bull SC26 Chicago-Bull Sequana XH31600 15

Overall, this was very neat to see.

What has been interesting at SC26 is that this type of platform would have been mind-blowing years ago. Now, with the NVIDIA Vera Rubin designs in ORv3 racks, the Bull systems fit in with the higher-volume AI infrastructure bits. What Bull has, however, is SiPearl, Intel Xeon Diamond Rapids, NVIDIA, and AMD Venice nodes on display in a common form factor.

Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 1Bull SC26 Chicago-Bull Sequana XH31600 Intel Diamond Rapids Package 1

Many other vendors are showing off Diamond Rapids systems with neat designs, including 32DIMM/ socket 2DPC configurations at SC26. For its part, Bull is the only one we have found that has a package. Diamond Rapids seems to be in the sampling phase. The best guess is maybe we start to see a launch in Q2 2027, but if that slips past June (or it becomes a split launch with 1P platforms and 2P platforms), the next real launch window for Intel is early September 2027.

原文为英文,正文由自动抓取并保留原文。

本文由采集器自动同步

原文发布于 ServeTheHome(Bull Shows Intel Xeon Diamond Rapids Package at SC26),标题、摘要与正文由机器翻译,术语按行业惯例统一,未经人工逐句校订。

阅读原文 ↗