Will PAYGO Shake Up How We Pay for Chips?

System builders are used to buying integrated circuits on a simple transactional basis — the chip has a price, and that’s what you pay. But some application spaces may have a wide variety of capabilities that need hardware support, and each feature may not be used for every instance.

Traditionally, one would design different chips for different feature mixes and price points. But a new proposal is out that enables hardware features with incremental charges, a pay-as-you-go (PAYGO) subscription model.

“This could turn out to be a brilliant business model if the optional features are enticing enough for a large part of the customer base to adopt them,” said Guillaume Boillet, director of product management at Arteris IP. “At the same time, it seems like a risky bet, considering the complexity of this multi-variate problem.”

5G is one such application. Handset requirements will be very different from base-station requirements, and even within each of those, implementations will vary. One company is proposing that, instead of investing in multiple chips, overall prices can be reduced by designing a single chip and turning on select features with purchased keys. While this would be done on the honor system for now, that might change in the future if more specific tracking mechanisms are accepted in the industry.

Getting the mix right
It always has been difficult to be sure that a given chip has the right feature mix for a specific application. Too few features, and the chip can’t do the job. Too many, and the price and power make the chip less competitive.

In a wide-ranging application space like 5G, there are some basic capabilities that every stop along the signal path must have. Beyond that, a handset will have different requirements from a base station — and from any other router or other equipment. It might be the size of the antenna array, or it might be the number of processors needed to handle a particular volume of traffic. Historically, each of these applications would have an SoC tuned to its needs.

For 5G, some applications will support a $50 price point, while others will be at a $500 price, according to EdgeQ. “5G is different things to different people,” noted Vinay Ravuri, CEO of EdgeQ. Individually designed chips could be priced accordingly, although the break-even price would reflect the need to cover the development costs of each chip.

Fig 1: One example of how 5G features may need to vary across installations. Source: EdgeQ

Sleeper features
SoC design on leading process nodes is expensive. It becomes tempting to try to do more with fewer chipset variations. In the limit, you could have a single superset piece of silicon with optional features that can be turned on and off.

In a traditional model, those features would be set perhaps at test, where the chip would be configured for its intended mission. Even though they all came from the same base die, each version would look like a separate chip with its own price.

“There are numerous examples of semiconductor companies marketing derivative chips with different features from a single die by enabling or blowing fuses to disconnect various functional blocks on the chip,” noted Andy Jaros, vice president of IP sales and marketing at Flex Logix.

Others agree. “It takes a lot of money to build a chipset, so what they do is they superset the features,” said Ravuri. “And they will average the price across all industries and set a price based on that and say, ‘Hey, a guy who really needs only $50 worth of product from me, you’re going to pay $250. The guy who needs $500 [worth of product], maybe you’ll pay $250, as well, because I don’t know what you need, and you haven’t told me what you need. So therefore, I’m going to charge you somewhere in the middle.’”

But what if the features were individually available for system designers to activate? Then no extra “chip” with its own part number would be necessary. The initial purchase always would be the base chip, with selective upgrades. The only addition would be a hardware mechanism built into the chip that could enable or disable features on demand.

Complex software packages have long been built like this. Above the base set of features, additional capabilities can be unlocked with specific feature lines in the license file. Those lines are added when the feature is purchased.

That’s straightforward to do with software licenses, but it’s not so intuitive for hardware. Obvious differences include the fact that software can be produced (as distinct from being developed) for largely the same cost regardless of features. Hardware, however, must etch those features into silicon, and so more feature options mean a larger die and higher production cost.

The first requirement, then, assuming that the base feature set isn’t intended as a loss leader, is that the price for that base configuration will still need to be profitable — although margin will be slimmer than if more features are used.

At that point, the question becomes one of figuring out how to pay for the additional features. One option is a single one-time upsell payment — a price adder to the base price to account for the additional features. But one-time payments for things are not in fashion, as shareholders prefer instead to get annuities that keep paying. That would suggest a subscription model.

Making it possible to set the features dynamically brings its own benefit. If the amount of traffic a newly developed system will handle over time is uncertain, for example, a designer can provisionally design and price for less traffic. If traffic picks up to the point where latencies are starting to suffer, the system vendor can upgrade the chip in situ to awaken more computing resources to keep up with the increased traffic.

That reduces risk, because it’s possible to start out conservatively and upgrade later. If the traffic growth doesn’t materialize, the system isn’t burdened with the cost of excess capacity. “Some people want to cap it so that, after so many years, there are no more payments,” said Ravuri. “Or some people will say, ‘Well, I don’t mind paying extra per year, because I may want to stop after three years for a certain amount of service just because it never took off the way we thought it was.’”

That’s exactly what EdgeQ is proposing with its new 5G chip. Systems with 5G chips have the capability to receive a key that will turn on a feature. In theory, that feature would be enabled for as long as the subscription is running, although that strategy has been only partly realized, as we’ll see shortly.

Conceptually, features could be enabled one-at-a-time for maximal configurability. This brings up the multivariate problem that Boillet mentioned. “Which feature should be offered as an option?” he asked. “What’s the corresponding silicon area? For how long will the upgrade last? How should the tier-based pricing be structured? What would be the alternative cost for the customer to upgrade its hardware?”

Allowing fine-grained choices would be complicated, because each such line would need its own key, and that key would need to be unique to that chip. The result would be significant key multiplication, which may well be manageable, but it bogs down logistics.

An alternative is to have different feature bundles. This is how EdgeQ is approaching the market. Those bundles could correspond to the different individual chips that would have been made in the past. That limits the number of options — a designer would not be able to pick and choose features for differentiation beyond the bundle. The key count would be greatly reduced, however.

But bundling also comes with risk. “That would probably require reducing the spatial and temporal granularity at which the upgrades are subscribed, thus probably reducing its value for the customer,” cautioned Boillet.

Delivering the features
EdgeQ will deliver keys over-the-air. Even though the delivered key is encrypted in transit, some people may still perceive a risk of the key being hacked. An alternative approach might be to implement optional features in an embedded FPGA (eFPGA). This assumes one or more of them could be architected in an efficient way — given the wide variety of features that might be available all across the chip — without compromising performance or power.

In that case, instead of downloading a key, the actual hardware code for the added features could be downloaded and implemented in the eFPGA. “Activating or de-activating certain functions by software provides more flexibility, but can be hacked,” observed Jaros. “This is a good use case for using an on-chip eFPGA, which can add another layer of security where the paid-for function is programmed into the eFPGA. Or, the eFPGA can be used as a multiplexer to wire up the paid-for functional blocks using encrypted bitstreams, which are very difficult to hack.”

Unactivated features unavoidably will affect the die size, with implications for cost. EdgeQ says the extra cost is still less than what the fully amortized cost of a dedicated design would be. Even if implemented with one or more eFPGAs, the silicon space for those arrays would remain in place regardless of the actual feature set they implemented. This alternative still burdens the base chip with extra area, and would need to be considered in the pricing model.

It’s also important that those unused features draw no power so that, operationally, it’s competitive with any other dedicated chip. EdgeQ says that its processor arrays, for example, revert to low power when not in use. That would apply whether the cores have been paid for but traffic is low at the moment, or the cores were extras that hadn’t been activated.

Keeping track
It’s one thing to have a technical feature that enables features to be activated. It’s another to set up the logistics so feature privileges can be enforced. After all, how does a chip buried somewhere down inside a system know the status of the subscription?

It’s technically possible for the chip to communicate with the outside world to check on such things. “The chip has the capability to call home,” said Ravuri. “The problem is when the chip goes into a box, there’s a security concern.” This is why prior attempts at such a strategy have been pared back.

Those kinds of concern are amplified when a radical new payment model is being proposed. So EdgeQ is going in partway, relying on trust for the time being. If a model like this becomes well accepted in the industry, further discussions of enforcement might be taken on. “[For a] business model like this, you have to start with an honor system,” observed Ravuri.

The practical implications of this are at least twofold. First, as implemented by EdgeQ, features can be enabled by a key, but there’s a catch for EdgeQ if the subscription lapses. “That key enables or disables the feature,” explained Ravuri. “But once it’s enabled, I cannot disable it from the cloud.”

While this could be an invitation to abuse, it also depends on which companies are purchasing the chips. Focusing on larger companies, while avoiding what might appear to be fly-by-night operations, could help, since such companies have reputations and goodwill to protect. Being known for feature fraud is likely not something that would be worth the long-term cost to such a company.

The other implication comes from reporting. How will the buyer and seller know how much is owed for a given period? The base chip price is easy, since the total is set by the number of chips sold, and it’s a one-time thing. It’s the ongoing part that’s harder.

Imagine, for instance, a maker of a variety of 5G systems at varying price points. The maker could purchase the same base chip and, for some systems it ships, enable no new features. For another system it enables a particular feature set, while for yet another system it enables a different set.

The total owed for a month would include nothing for the first system, which has no upgrades, as well as the subscription for the second system’s features times the number of chips used in those systems, and the third feature’s upgrade subscription times its volume.

In the call-home model, each of these chips would check in with their maker periodically to report which features were enabled. The chipmaker would then tally all of that and send an invoice. If a system was taken out of use, that chip would stop calling home and would no longer be charged.

Because no call-home capability is in place now, the roles are reversed. In this honor system, it’s up to the customer to report on how many chips have been used with each feature set. It’s not, “Here’s what you owe us” from the vendor. It’s, “Here’s what we owe you” from the customer.

Longer term, it’s unlikely that the current model can suffice. There’s too much temptation to violate honor systems once you go beyond big-name customers. But if a call-home capability was instituted, these chips always would need a communication channel.

A chip focusing on the 5G market is pretty much guaranteed to have such a channel. But, for other applications, that might not always be the case. Claims of “always connected” notwithstanding, many systems today do not connect. Creating a payment system that requires a connection would have to change that.

It’s already the case that more and more chips are looking for a connection out. The proliferation of on-chip monitors that can respond to requests for chip data is a separate development. Those require at least a periodic connection. A fully tracked payment system would likewise require at least an occasional connection.

Applications and economics
This may not be a useful model for all chips going forward. Not every application space will have the variety of possible implementations that 5G presents. “The business model seems suitable mostly for industries where feature adoption is not easily predictable — and often underestimated — and where the cost of replacement is high,” observed Boillet.

But it does give system and chip designers an opportunity to rethink how they go to market. Done properly, this can be a win for both producer and consumer. The theory is that the sum of base and subscription prices should end up lower than what would be charged if a separate design were done for each variant.

The industry might be tempted to simply to convert one-time chip payments entirely into longer-term annuities, with the net result being higher cost for chips in applications where this is a less obvious fit. One could imagine a down-payment for a chip and a monthly subscription just to keep the chip working — no features involved. That would likely spur some industry pushback, unless it was implemented by a behemoth company in a take-it-or-leave-it scenario.

This presumably will take some time to sort out, but EdgeQ is careful to point out that this isn’t a trial balloon idea. “We have customers for this chipset, and they have adopted this model,” Ravuri pointed out. “So it’s not an experiment. It’s not theory. It’s real.”

EdgeQ appears to view this as the future, seeing economic benefits both for themselves and for its customers. “As the industry evolves, I strongly believe this is where the chipset world is going,” said Ravuri.

But as Boillet pointed out, “The pay-as-you-go model comes with an implied promise of reduced cost of ownership, which can be satisfied only if clear win-win dynamics are present.”