Question What role does Broadcom play in semi-custom ASIC (like google TPU)?

[MasterofZen]

Sorry, this thread may seem a little bit off-topic, but I think no other sub-forums are more relevant to this topic.
As I understand Google's TPU (tensor processing unit) is a semi-custom chip made by Broadcom, but it's actually Google that designed the tensor cores, which is very different from other semi-custom chips where the chip vendor has the key IP of the chip (AMD designs and owns key IP in semi-custom chips used in PS and Xbox as an example).

It seems Broadcom's role is very limited in this partnership, so why Google buys chip from Broadcom instead of directly buy wafers from TSMC, this would not only cuts out the middle-man, but also have more control in supply chain (wafer capacity, HBM supply etc). If Google is like Meta, who is just beginning to design their own chip, it would be understandable to have a experienced chip company help out, but Google has been designing TPU for about a decade, why still rely on Broadcom and make it a chip provider instead of a design service provider?

Of course a chip has other components besides processing cores, like DDR/PCI-e peripherals, which Broadcom probably helps out, but can Google just buy these IPs from Cadence/Synopsys etc?

 

[pinku]

Trying to understand the same.

 

[NTMBK]

Answer from GPT4o:

----

Your question about Google’s relationship with Broadcom for TPU manufacturing reflects the complexities of chip design, manufacturing, and supply chain management. Here’s an analysis of why Google might choose Broadcom as a partner rather than directly dealing with TSMC:

1. Expertise Beyond Tensor Cores

While Google designs the tensor cores, the TPU is not just about the processing cores. It includes many components, such as:

• Memory interfaces (e.g., HBM)
• High-speed I/O (e.g., PCIe, networking interfaces)
• Power management
• Packaging technologies (2.5D/3D integration)

Broadcom has extensive expertise in integrating these components into a single, highly optimized package. Even if Google could license IPs from Cadence, Synopsys, or others, integrating and validating them into a coherent, production-ready chip is non-trivial. Broadcom brings decades of experience in this area.

2. Turnkey Manufacturing and Supply Chain Management

Broadcom acts as more than just a “middleman.” It provides:

• Foundry Management: Broadcom can handle wafer procurement and fabrication logistics with TSMC, leveraging its longstanding relationship and volume deals to secure capacity.

• Packaging and Testing: Advanced packaging and chip testing are critical, especially for cutting-edge chips with HBM memory and other high-performance requirements.

• Supply Chain Risk Mitigation: Broadcom manages risks such as supply shortages (e.g., HBM) and ensures timely delivery of fully assembled and tested chips.

For a company like Google, which operates at hyperscale, delegating these responsibilities to Broadcom simplifies operations.

3. Cost-Effectiveness


While going directly to TSMC might seem cheaper, in practice:

• Google would need to invest significantly in building the internal infrastructure to manage the entire chip manufacturing pipeline.
• Broadcom, with its economies of scale and existing partnerships with vendors like TSMC and memory suppliers, can likely negotiate better prices and terms.

Broadcom’s involvement may add a markup, but it offsets Google’s need to build and maintain a massive semiconductor supply chain operation.

4. Risk and Focus

Designing a chip like the TPU involves managing technical, operational, and market risks. By partnering with Broadcom:

• Google reduces the burden of handling manufacturing, packaging, and other non-core functions.
• Google’s chip team can focus on architectural innovation and software integration, which are more aligned with its strengths and strategic priorities.

Moreover, if there are issues in the supply chain or manufacturing process, Broadcom absorbs much of the responsibility and risk, not Google.

5. Time to Market

Broadcom’s established expertise allows for faster development and deployment cycles. Google’s reliance on Broadcom could help meet aggressive timelines, especially when rapid iteration is critical in a competitive AI hardware market.

6. Why Not Meta’s Approach?

Meta’s approach differs because it is a relative newcomer to custom silicon. It works with experienced semiconductor companies as service providers. Google, despite its experience with TPU design, may still find Broadcom’s expertise beneficial in scaling production and managing complexity. Broadcom offers a more comprehensive solution than a pure design service provider.

Conclusion

Broadcom’s role in the TPU production process is far from limited. While Google designs the tensor cores, Broadcom contributes integration expertise, supply chain management, and manufacturing execution. These capabilities allow Google to focus on its core competencies, minimize risk, and accelerate time to market.

Cutting out Broadcom might increase Google’s control but at the expense of added complexity, resource allocation, and risk. In the hyperscale cloud and AI market, partnering with a seasoned player like Broadcom often makes more strategic and economic sense.

Could we please not fill this forum with LLM generated rubbish?

 

[mikegg]

Could we please not fill this forum with LLM generated rubbish?

It did a good job explaining the benefits Broadcom brings in my opinion.

One thing it missed is that Broadcom has a host of networking solutions to connect all the TPUs together.

 

[GTracing]

It did a good job explaining the benefits Broadcom brings in my opinion.

One thing it missed is that Broadcom has a host of networking solutions to connect all the TPUs together.

The issue is any part of it could be wrong. I don't know much about the subject, but even I can see a lot of its points are questionable.

• It speaks about the deal as a hypothetical.
• In some parts it mentions Google's existing chips, but in other parts it speaks as if Google has no experience in chip design and manufacturing. "Google would need to invest significantly in building the internal infrastructure to manage the entire chip manufacturing pipeline."
• It says that the deal reduces "the burden of handling manufacturing, packaging, and other non-core functions." Google is getting more involved in their tensor mobile chips. Why is this broadcom deal is different than that?
• And in general, a lot it is just word salad. What does "Broadcom offers a more comprehensive solution than a pure design service provider" mean? Google is more similar to a "service provider" than it is to Meta in this example.

In short, it's misleading trash.

 

[mikegg]

The issue is any part of it could be wrong. I don't know much about the subject, but even I can see a lot of its points are questionable.

Anyone writing here could be wrong as well.

GPT4o's rebuttal to your points:

• It speaks about the deal as a hypothetical.

The nature of such discussions is often speculative because details of partnerships between companies like Google and Broadcom are not fully public. That said:
• The analysis is based on common practices in the semiconductor industry and Google’s publicly known reliance on Broadcom for TPUs.
• Google’s partnership with Broadcom has been documented in reports, such as Broadcom being listed as the manufacturing partner for TPUs.

Clarifying: The intent was not to describe the deal as purely hypothetical, but to explain the general rationale behind such a setup.

• In some parts it mentions Google's existing chips, but in other parts it speaks as if Google has no experience in chip design and manufacturing. "Google would need to invest significantly in building the internal infrastructure to manage the entire chip manufacturing pipeline."

It’s true that Google has significant experience in chip design, especially with TPUs, but there’s a difference between design expertise and the operational capability to manage the entire chip supply chain. Designing tensor cores (or even the entire TPU architecture) is distinct from handling:
• Wafer procurement
• Packaging and assembly
• Testing and logistics

Google’s experience in designing mobile tensor chips (used in Pixel devices) is relevant, but those are smaller-scale projects with specific use cases. Cloud-scale TPUs are a different category in terms of complexity and supply chain requirements. Partnering with Broadcom might reduce the burden on Google’s internal teams, allowing them to focus on core innovations rather than operational execution.

• It says that the deal reduces "the burden of handling manufacturing, packaging, and other non-core functions." Google is getting more involved in their tensor mobile chips. Why is this broadcom deal is different than that?

Google’s mobile tensor chips (like the Tensor G-series for Pixel) are a different product category:
• They are fabricated by Samsung Foundry, not TSMC, and use relatively simpler packaging technologies compared to TPUs.
• The scale and performance requirements of cloud TPUs (e.g., integration with HBM, multi-chip modules, extreme thermal/power demands) necessitate more complex manufacturing expertise.

The Broadcom collaboration likely reflects the additional complexity of TPUs compared to mobile SoCs. Google’s decision to involve Broadcom in TPUs but manage mobile chips more directly could stem from resource prioritization or Broadcom’s added value in specific areas.

• And in general, a lot it is just word salad. What does "Broadcom offers a more comprehensive solution than a pure design service provider" mean? Google is more similar to a "service provider" than it is to Meta in this example.

Statements like “Broadcom offers a more comprehensive solution than a pure design service provider” could indeed feel vague. Let’s clarify:
• Pure design service providers (e.g., Cadence, Synopsys) focus on licensing specific IPs or helping with chip design but don’t offer end-to-end manufacturing and supply chain solutions.
• Broadcom, by contrast, provides integration expertise, access to foundry capacity, packaging, and logistics. It’s not just about the tensor cores but the whole chip lifecycle.

Google might rely on Broadcom because it provides not just design assistance but also manages the production complexities, including ensuring wafer and memory supply, testing, and delivering fully validated chips. This makes Broadcom a partner in execution, not just design.

 

[adroc_thurston]

One thing it missed is that Broadcom has a host of networking solutions to connect all the TPUs together

That's literally the reason why they're used by GOOG/etc.

 

[MasterofZen]

Answer from GPT4o:

----

Your question about Google’s relationship with Broadcom for TPU manufacturing reflects the complexities of chip design, manufacturing, and supply chain management. Here’s an analysis of why Google might choose Broadcom as a partner rather than directly dealing with TSMC:

1. Expertise Beyond Tensor Cores

While Google designs the tensor cores, the TPU is not just about the processing cores. It includes many components, such as:

• Memory interfaces (e.g., HBM)
• High-speed I/O (e.g., PCIe, networking interfaces)
• Power management
• Packaging technologies (2.5D/3D integration)

Broadcom has extensive expertise in integrating these components into a single, highly optimized package. Even if Google could license IPs from Cadence, Synopsys, or others, integrating and validating them into a coherent, production-ready chip is non-trivial. Broadcom brings decades of experience in this area.

2. Turnkey Manufacturing and Supply Chain Management

Broadcom acts as more than just a “middleman.” It provides:

• Foundry Management: Broadcom can handle wafer procurement and fabrication logistics with TSMC, leveraging its longstanding relationship and volume deals to secure capacity.

• Packaging and Testing: Advanced packaging and chip testing are critical, especially for cutting-edge chips with HBM memory and other high-performance requirements.

• Supply Chain Risk Mitigation: Broadcom manages risks such as supply shortages (e.g., HBM) and ensures timely delivery of fully assembled and tested chips.

For a company like Google, which operates at hyperscale, delegating these responsibilities to Broadcom simplifies operations.

3. Cost-Effectiveness


While going directly to TSMC might seem cheaper, in practice:

• Google would need to invest significantly in building the internal infrastructure to manage the entire chip manufacturing pipeline.
• Broadcom, with its economies of scale and existing partnerships with vendors like TSMC and memory suppliers, can likely negotiate better prices and terms.

Broadcom’s involvement may add a markup, but it offsets Google’s need to build and maintain a massive semiconductor supply chain operation.

4. Risk and Focus

Designing a chip like the TPU involves managing technical, operational, and market risks. By partnering with Broadcom:

• Google reduces the burden of handling manufacturing, packaging, and other non-core functions.
• Google’s chip team can focus on architectural innovation and software integration, which are more aligned with its strengths and strategic priorities.

Moreover, if there are issues in the supply chain or manufacturing process, Broadcom absorbs much of the responsibility and risk, not Google.

5. Time to Market

Broadcom’s established expertise allows for faster development and deployment cycles. Google’s reliance on Broadcom could help meet aggressive timelines, especially when rapid iteration is critical in a competitive AI hardware market.

6. Why Not Meta’s Approach?

Meta’s approach differs because it is a relative newcomer to custom silicon. It works with experienced semiconductor companies as service providers. Google, despite its experience with TPU design, may still find Broadcom’s expertise beneficial in scaling production and managing complexity. Broadcom offers a more comprehensive solution than a pure design service provider.

Conclusion

Broadcom’s role in the TPU production process is far from limited. While Google designs the tensor cores, Broadcom contributes integration expertise, supply chain management, and manufacturing execution. These capabilities allow Google to focus on its core competencies, minimize risk, and accelerate time to market.

Cutting out Broadcom might increase Google’s control but at the expense of added complexity, resource allocation, and risk. In the hyperscale cloud and AI market, partnering with a seasoned player like Broadcom often makes more strategic and economic sense.

As a LLM, it's doing a good job explaining the pros and cons, but lacks depth and insight, which is kind of understandable.

 

[DAPUNISHER]

Deleted the post as it is exactly the content spambots provide, and there were a number of complaints concerning it. Everyone please provide your own analysis instead of relying on copy pasta. Otherwise the posts will be deleted. Thank you for understanding.

Mod DAPUNISHER

 

[Doug S]

The Broadcom collaboration likely reflects the additional complexity of TPUs compared to mobile SoCs. Google’s decision to involve Broadcom in TPUs but manage mobile chips more directly could stem from resource prioritization or Broadcom’s added value in specific areas.

Apple has also been rumored as working with Broadcom for making their own AI chips, and no one doubts their silicon design chops. The stated reason is Broadcom's expertise in high speed networking - and that's one place where Apple has no experience. There's a big gap between the latest Thunderbolt standard and 800 GbE - and since this is for internal consumption there's no reason to believe Apple would limit its connectivity to what is "standard" today and might use some sort of draft 1.6 TbE if its available to them. And if anyone has such a thing "available" it would be Broadcom.

Google's needs for their TPU cloud would include the similar need for speed, and a similar lack of caring whether or not IEEE has blessed what they're using.

 

[mikegg]

That's literally the reason why they're used by GOOG/etc.

Is there a source?