It has been more than six months since the ODSA Workgroup started working together on an informal basis. The Workgroup has grown, released a very detailed technical white paper and is heading toward its first workshop on January 29.
For those of you who are not familiar with the ODSA – it stands for Open Domain-Specific Architecture. The charter of this Workgroup is to define an architecture that enables the mixing and matching of available silicon die from vendors to develop SoCs (system on a chip) for data center applications. The interface specifications being contemplated or developed are geared toward making a Lego Blocks approach possible where each Block is a best-of-breed chiplet made available by silicon vendors.
I have had informal discussions with many of you in the industry to solicit feedback on everything that has been announced as related to the work done to date by the ODSA Workgroup. My objective has been to gauge the pulse in the industry. It is throbbing! The interest level is extremely high. However, apprehensions and skepticism exist as well, which is a natural occurrence with anything new that challenges the accepted norms. I will lay those out in this blog.
I met with industry colleagues from three groups of companies that play in the world of SoCs and domain-specific accelerators (co-processors for networking, security, storage, machine learning and inferencing). I realized that how the value of ODSA gets assessed or perceived is different depending on the group. The first group includes hyperscalers designing and/or building their own SoCs. An example is Google and its TPU (Tensor Processing Unit). The second group comprises very large silicon suppliers (like Broadcom and Intel) that build and sell their own SoCs, and also build custom SoC silicon for others such as hyperscalers who supply some of the IP included in the SoC. In the third group are smaller silicon companies with domain-specific expertise that also design and build SoCs. The myriad of machine learning/inferencing silicon startups fall in this category, as do smaller FPGA, GPU, security, networking and storage-related SoC vendors.
Let me start with the hyperscalers. It is well known that many of them are building their own specialized IP (intellectual property) to make their data center operations more efficient. For example, they are designing IP related to domain-specific functions for networking, security or machine learning and inferencing, and deploying their own SoCs built using such IP. Such monolithic SoCs not only include domain-specific logic, but also other general purpose logic and circuits such as Arm for general purpose processing, MAC and SerDes for network and host interfaces, and DDR or HBM for memory access. The general-purpose logic and circuits are deemed commodity IP and can be sourced from multiple vendors. Hyperscalers care about speed of innovation and deployment, and the ability to source commodity silicon IP from best-of-breed vendors. When they are forced into a silicon manufacturing process node or fab due to dependencies on the commodity list of IP, there are undesired delays in time to deployment – even with the IP they develop and own. These are challenges they want addressed and they see the ODSA as a potential and viable path. While they support the ODSA goals, concerns from this camp relate to what aspects of ODSA-related offerings from vendors will be open and if the interfaces that promise the Lego Blocks approach will be adopted by chiplet vendors. Will the ODSA specification and contributed open IP be sufficient and practical for both the vendors and the users in terms of actual products development and deployment, and how will such open IP be governed?
Most of the large silicon suppliers are incumbents in the data center SoC market. These companies typically have established central engineering teams that own most, if not all, of the commodity IP mentioned previously. Domain-specific expertise exist in specific business units and they design and manufacture SoCs utilizing the commodity IP available from the central engineering team. Some of these companies have established custom silicon offerings that address this market. Most of the apprehensions and skepticism I have heard are from this group and perhaps rightfully so. After all, they are serving this market, have easy access to needed IP, enjoy the benefits of economies of scale and the status quo is favorable. Specifically, one company pointed out that the efforts of the ODSA will likely trail technology trends. Another one believes in the significant business value of the initiative but thinks the disruptions to current silicon packaging and verification workflows will be too high, causing the initiative to fail. The third piece of feedback I captured is skepticism about the use of the word “open” in the initiative, because silicon vendors are, by nature, very protective of their IP. These are valid concerns that the ODSA community will need to address to garner valuable support from these leading silicon suppliers.
The third group of smaller silicon companies has been the most vocal of their support of ODSA because the economic benefits are very obvious to them. A machine learning startup summarized the value very aptly saying, “We have raised $30M in funding for developing our SoC. About $5M will be used for our own R&D for the domain-specific IP, our differentiation. The irony is that about $20M is stipulated for licensing other general purpose IP such as a high performance CPU, and high-speed host and memory interfaces.” Other companies building monolithic chips using licensed IP are often forced into smaller manufacturing process nodes that are significantly more expensive. Many of them are beginning to lean toward the Open RISC-V architecture because it is open. An innovative startup called SiFive has created commercially deployable versions of CPUs based on that architecture. And their bold business model requires payment of a significantly lower upfront fee compared to the incumbents. The chiplet approach promoted by the ODSA addresses these issues that are extremely important to the smaller silicon companies. The initiative promises to significantly reduce cost of all relevant general purpose IP, including CPUs, as needed in SoCs.
There is good momentum behind the ODSA effort. However, as with any new initiative that disturbs the status quo, there are many challenges too. They need to be addressed by the ODSA Workgroup to ensure success. So far, no one that I have met has disputed the substantial commercial benefits of this initiative and this is very encouraging. To the general industry, a thriving ecosystem of domain-specific SoC silicon suppliers in a fast-growing market is needed more now than ever before.