In its latest research report, LightCounting mentioned that at the Taipei conference of the Open Compute Project (OCP) held on August 5th and 6th, the ongoing debate over Scale-Up solutions—a hot topic in the current AI hardware architecture field—once again took center stage.
LightCounting stated that one of the most noteworthy panel discussions at the Taipei conference revolved around "scale-up networking." Key participants in this discussion included Kurtis Bowman (Chairman of the UALink Alliance Board of Directors and Director of Architecture and Strategy at AMD) and Ram Velaga (Senior Vice President of Broadcom and General Manager of the Core Switch Business Group).
Interestingly, Ram Velaga from Broadcom—also a co-founder of UALink—referred to both UALink and NVLink as competitors during his speech. On one hand, he refused to acknowledge UALink technology in any way; on the other hand, he continued to liken NVLink Fusion to a "fox guarding the henhouse" (implying that NVIDIA excludes other players from this domain it dominates). What Broadcom truly focuses on is its Ethernet standards: the company is considered a "gorilla" (i.e., an industry leader) in the Ethernet switch sector, and it is certainly keeping a close eye on the Scale-Up Ethernet (SUE) market. Broadcom’s Tomahawk Ultra switch, launched last month, marks its official product entry into the SUE field. LightCounting predicts that Broadcom will release a version of the Tomahawk Ultra equipped with Co-Packaged Optics (CPO) technology as early as the beginning of next year (or even sooner).
For details on the discussions at the Taipei forum, refer to reports from Business Today. Ram Velaga, a former Cisco executive, stated, "NVIDIA is an excellent GPU company, but I don’t believe it possesses strong networking technology." He cited an example: over two decades ago, HP and Dell promoted closed blade server racks, but cloud service providers ultimately opted for white-box server racks. "This shows that a company may not excel at everything. Even if it reluctantly attempts vertical integration, the market will eventually move toward an open ecosystem." He criticized NVLink Fusion for not being truly open, arguing, "NVIDIA can decide to revoke authorization at any time, and it also requires connections to its own switch chips, GPUs, or CPUs."
Kurtis Bowman questioned Broadcom’s repeated claim that Ethernet outperforms all competing technologies—including those supported by the UALink Alliance—in terms of latency in the Scale-Up domain. He stated, "SUE is not a standard at all; it is merely a 'whitewashing' of Broadcom’s own standards and specifications. Currently, SUE provides no specifications for data transmission latency."
In response, Ram Velaga directly presented samples of Broadcom’s products to counter the argument. He bluntly noted that no UALink solutions had yet emerged in the market, suggesting that UALink was "not ready." Addressing the latency standard issue, he retorted, "Raising this question shows you don’t understand Ethernet at all. Ethernet has never defined latency in its standards—customers choose low-latency switch chips based on their preferences... The only reason you want to set a latency standard is because you don’t know how to design low-latency switch chips, right?" Ram also questioned the underlying specification of UALink—PCIe—stating that it had traditionally been used for short-range chip transmission and would struggle to support long-range communication between thousands of processors like Ethernet. He challenged, "Is there any company today whose PCIe switch chips have sufficient buffer memory to support long-range transmission? Good luck with that."
In the face of Ram’s aggressive remarks, Kurtis Bowman refused to back down, stating that he looked forward to the launch of UALink solutions in nine months, at which point both sides could compete based on tangible technological results. In a post-conference exclusive interview with DIGITIMES, Kurtis Bowman acknowledged that Ethernet does perform exceptionally well in the Scale-Out domain. This is because the core goal of Scale-Out is to physically transmit data—a requirement that aligns perfectly with Ethernet’s inherent characteristics. However, he emphasized that Scale-Up is not about data transmission; instead, it focuses on connecting and synchronizing the operations of "different computing chips within the same rack"—an area where Ethernet is not strong. For this reason, he expressed skepticism about Broadcom’s "Ethernet technology promotion." More importantly, Kurtis Bowman argued that based solely on the current specifications provided by the UEC (Ultra Ethernet Consortium), few Ethernet chip manufacturers or custom chip developers could match Broadcom’s capabilities. He noted that Broadcom’s decades of accumulated technical know-how in Ethernet chips were the real secret to its success. Given Broadcom’s overwhelming advantages and control over high-barrier technology domains, he questioned whether an environment where "openness and free competition" are loudly advocated can truly be considered "open."
Kurtis’s criticism is not without merit. While the dispute appears to be about technical standards, it is ultimately driven by competing interests. When it comes to Broadcom and NVIDIA—who is the "fox," or are both of them "foxes"?
Launched in May 2024 by a consortium of companies including AMD, Broadcom, Google, HPE, Intel, and Microsoft, UALink (Ultra Accelerator Link) is an open standard designed for AI and high-performance computing (HPC) scenarios. Its goal is to enable high-bandwidth, low-latency communication between accelerators, CPUs, and switch chips. Compared to traditional solutions such as PCIe, CCIX, or NVLink, UALink places greater emphasis on openness and scalability, facilitating broader collaboration among vendors. For instance, the UALink 1.0 version offers a bidirectional transmission rate of up to 50GB/s per lane and supports interconnection configurations for up to 1,024 accelerators, with its design tailored to meet the training needs of ultra-large-scale models.
In its research report, LightCounting pointed out that the competition among the three major technical routes—SUE, UALink, and NVLink Fusion—is expected to be resolved by the market by the end of this decade (2020-2030). During this period, all parties will engage in fierce competition to gain industry leadership, presenting significant opportunities for switch suppliers.
Key Term Explanations
Scale-Up Networking：A network architecture that enhances performance by upgrading single-node hardware (e.g., increasing chip capacity) rather than adding more nodes.
Scale-Up Ethernet (SUE)：Broadcom-proposed Ethernet-based solution for Scale-Up scenarios, aiming to compete with UALink/NVLink.
Co-Packaged Optics (CPO)：A next-gen optical interconnect technology that integrates optical components with switch chips to reduce latency and power consumption.
Advanced Micro Devices (AMD)：A leading semiconductor company, competing with NVIDIA in GPUs and x86 CPUs.
White-Box Server：Generic, unbranded servers (often customized by cloud providers) to reduce costs, as opposed to branded servers (e.g., Dell/HP).
Open Compute Project (OCP)：A collaborative initiative to design open, efficient hardware for data centers and cloud computing.