Why is it so difficult to design digital baseband (e.g., LTE) chips? | Hacker News Reader

Why is it so difficult to design digital baseband (e.g., LTE) chips?

4 pointsav22977 comments

The LTE digital baseband market was famously dominated by Qualcomm, which maintained a 60+% market share for nearly a decade (2009-2019). MediaTek also successfully entered the market, but not for high end phones for high-income markets (U.S., Europe, etc.). Famously, both Intel and Broadcom (two semiconductor firms known for their ability to design successful chips in other markets, e.g., MPUs, cable modem chips, WiFi chips, Bluetooth chips, etc.) attempted to enter the market and failed. Broadcom exited the market and wound down their LTE business unit. Intel sold its group to Apple, which now has gone on to fail to develop a 5g digital baseband solution (Apple continues to use Qualcomm for its 5g digital baseband solution).

Why (from an engineering perspective) is it so difficult to develop an advanced LTE (or 5G) digital baseband chip such that even semiconductor powerhouses like Intel and Broadcom have struggled to develop solutions which compare to Qualcomm's offerings?

Note - answers from semiconductor design engineers appreciated :)

Direct involvement with drafting communications standards. Understanding Communication theory, advanced mathematics & algorithms. Computer Science involved with Hardware acceleration of the DSP algorithms. Quality of Results measurement and testing involving ridiculously expensive test equipment. (Like one piece of test equipment will cost as much as a house) Large scale chip manufacturing will require direct involvement with Semiconductor fabrication plants. the baseband comms chips are basically driving the latest small-scale wafer technology and requires the fab to evolve new techniques and new machinery. A 'winner takes all' environment, where the winning company will either win the mutli-million dollar contract, or take a huge hit. Because the baseband side is inherently centralized, it's not realistic to have many separate deployed solutions. I worked on comms for many years. It's interesting because so much human effort has been spent finding very sophisticated ways for processing large numbers of independent signals, and correcting errors. There are very intersting problems like digital pre-distortion, adaptive equalization etc. and crazy linear algebra problems. But the down side, is that you would most likely be working on a small piece of the large puzzle, and there is very little recognition for the amount of work required. There are very few 'famous' baseband engineers. The feeling is very much like a cog in the machine you you don't really see your direct contribution.

It's not the semiconductor design that is the most difficult. It is implementing the signal processing requirements.

Encoding of digital data stream to RF signal is difficult. Decoding a complex RF signal from amongst numerous conflicting signals, EMF, noise sources, etc is extremely difficult. To create a successful LTE chip requires GHz RF engineering skills (which are very rare) and then massive amounts of DSP programming (which are also pretty rare).

The only way to acquire the required skills is to work in the field at an established vendor's R&D lab. There is no way that a hacker in a spare room can acquire the required skills. Besides you need a couple of $M for test equipment and suitable environment. As for startups, they would need to poach talent from an established LTE manufacturer and pretty much start from scratch. VCs don't like those sorts of risks.

Both the RF engineering and DSP programming sound immensely interesting. As a software dev with a EE background I’m thirsty for such opportunities. Where does one look for these jobs?

Could you explain the difference between the "semiconductor design" and "implementing the signal processing requirements" - not sure I followed the difference between these. The standard in question (e.g., HSPDA) defines the signal processing techniques a chip needs to be able perform. You must then design a chip that implements the standard's signal processing techniques. So those two phrases seem identical to me in this context? Of course, there is a segmentation between the RF Front End (often made by a separate manufacturer) and the digital baseband ASIC (which implements the DSP techniques specifically).

>There is no way that a hacker in a spare room can acquire the required skills

https://bellard.org/lte/

To be fair, you can't use Fabrice Bellard as a counterexample because he isn't human -- he is a time-traveling AI from the future endowed with alien superpowers.

But, other than that, you pretty much stomped all over GP.

Patents form a large part of the enterprise issue.

If you're going to have to pay Qualcomm for the patents, you don't have much margin as a separate vendor and you might not save enough money and gain enough control to justify it as an in-house design. If you're Apple, you also don't want to put your chip out there and see it fail to connect to some towers than Qualcomm chips do which makes iPhone look bad; you really want your product in the testing matrix the base stations use, so if it fails to work, it's a base station bug, not a handset bug.

Reply on news.ycombinator.com