Thoughts on software-defined silicon [LWN.net]

Thoughts on software-defined silicon

Posted Feb 19, 2022 11:57 UTC (Sat) by khim (subscriber, #9252)
In reply to: Thoughts on software-defined silicon by Cyberax
Parent article: Thoughts on software-defined silicon

> But this works only if you don't have competition.

If you have competition then it becomes even more important.

> After all, it doesn't cost you anything extra to enable a feature that your competitor doesn't have.

It would cost you a lot. If you would stop selling $299 Ryzen 5 5600X (which is, essentially, $449 Ryzen 7 5800X with two cores disabled) and would start selling Ryzen 7 5800X for $399 instead of having these two… you would lose both on people who were ready to spend $449 and on ones who were ready to buy something for $299.

I took example from AMD book, not Intel book to show that as competition grows more acute the need to disable features and sell crippled product grows, not diminishes. Market segmentation is powerful tool.

When AMD had no ability to make powerful CPUs (for many years AMD's CPU before Ryzen were awful) — because it was hard to sell even CPUs with even all features enabled (and it was losing money as the result). When AMD made something for the top tier — it started marked segmentation games (and immediately become profitable).

> We're already seeing this with AMD, it's steadily eating into Intel's server marketshare that had been unassailable up until ~3 years ago

And this has nothing to do with the fact that AMD presented 7nm EPYC Rome two years ago while Intel still couldn't make 10nm Xeons (10nm of Intel is more-or-less the same as 7nm of TSMC), but everything to do with the fact that EPYCs are less segmented? Dream on.

to post comments

Thoughts on software-defined silicon

Posted Feb 19, 2022 15:11 UTC (Sat) by jhoblitt (subscriber, #77733) [Link] (3 responses)

It isn't just an issue of competition in the market. Fab capacity of each company and the overall market is a factor. Currently, the entire world is maxed out on wafer starts. So while it probably saves on masks, validation, etc. to cripple a perfectly functional 16core die to be sold as an 8core die, that means there is an 8core die area of silicon that is now lost. In a situation where every cpu produce can be sold, that lost real estate represents a potential 100% increase in gross margin.

Thoughts on software-defined silicon

Posted Feb 19, 2022 15:19 UTC (Sat) by khim (subscriber, #9252) [Link] (2 responses)

If we have lived in a world where every CPU can be sold then we would have seen similar craziness to GPU market where prices are 2x-3x recommended price.

And GPU makers don't, actually, embrace that craziness, they fear it because they know what comes next: governments would say that cryptocurrency mining is a criminal activity, GPU sales would drop through the floor and they would be selling them at loss for some time.

Fab capacity is strained not because it's impossible to build mode fabs, but because it's impossible to do that profitably: build cycles are many years, investment are measured in billions and unused fabs are not aging well.

Thus no, your reasoning doesn't make sense long-term. And CPU/GPU manufacturing is long-term process, it takes many years to develop CPU from scratch and year or two just to do minor alterations.

Thoughts on software-defined silicon

Posted Feb 19, 2022 16:10 UTC (Sat) by jhoblitt (subscriber, #77733) [Link] (1 responses)

I was just quoted a 26+ week lead time on any zen3 epyc cpu with >= 32cores from a major manufacturer. I ended up accepting zen2 cores in order to cut the estimated lead time in half. However, the price will start to float at the market rate 90days from the date the quote was issued. I won't even know the final cost until I'm invoiced for the shipment. If this isn't a CPU shortage, I don't know what one would look like.

The GPU and CPU makers are all bidding on the same fab capacity. When AMD reserves wafers at TSMC, that's capacity that is denied to Apple/Nvidia/Intel/etc. and vice versa.

The ability to build new fabs is not unlimited. Bleeding edge fabs need equipment from ASML who reports that they and their supply chain are already maxed out. The world is essentially already building new fab capacity at the maximum rate they can get lithography equipment.

The rumors are that TSMC is booked out *years* in advance at the 5, 4, and 3nm nodes. Are you arguing that AMD is going to produce 128core zen4s and then cripple perfect dies down to 16c core parts instead of taping out multiple designs that use the exact same logical blocks, when all that lost real estate could have instead been used for GPUs or ~5-7 additional CPUs?

Thoughts on software-defined silicon

Posted Feb 19, 2022 17:12 UTC (Sat) by khim (subscriber, #9252) [Link]

> I was just quoted a 26+ week lead time on any zen3 epyc cpu with >= 32cores from a major manufacturer. I ended up accepting zen2 cores in order to cut the estimated lead time in half.

That's nice piece of data. Let's try to decipher it, shell we?

You asked for latest-and-greatest cores and got promise to receive them when they would be made (you may not know that, but half-year or more is not atypical if we are talking about manufacturing of complex 7nm chip: they require almost hundred masks and application of one mask takes full day or more, depending on how much load fab experiences right now). The exact same thing happened with surprisingly good R300 back when it was hot — yet somehow back then noone was bawling their eyes out and complained about shortage of fabs. Everyone accepted that it was simple management miscalculation.
When you asked for less “hot” chips made by the exact same manufacturer in the exact same lab by the exact same producer you got shorter times because, apparently because there are surplus of these in production.
Yet it's very easy to buy consumer-grade chips with Zen3 core. The full nomenclature from lowly Ryzen 5 5600X (which my friend bought in India two days ago) to Ryzen 9 5950X (which I bought at the same time). Prices are less than what recommended prices are. Many of these are artificially crippled.

Doesn't look like shortage of CPUs to me, sorry.

> If this isn't a CPU shortage, I don't know what one would look like.

Indeed, you don't know that. When you need to pay 10x or 100x price to receive 250nm chip (like some automotive chips are selling for right now) and situation stays that way for years — then you can say there are shortage of chips.

Till then it's normal reaction of market on changes in demand and supply for something that takes years to produce coupled with customers who, naïvely, expect to buy the same thing with lead times measured in weeks.

JIT-manufacturing, both good and bad sides of it.

> The rumors are that TSMC is booked out *years* in advance at the 5, 4, and 3nm nodes.

Why do you say these are rumors? That's the reality. Latest nodes are always booked years in advance. Because fabs are extremely expensive yet they can only ask for extra-premium prices for latest nodes for a few years… nobody builds spares. You don't even need rumors to confirm something that was always true.

> The GPU and CPU makers are all bidding on the same fab capacity.

Nope. That's not true because of, as you have said yourself: fab capacities are booked years in advance. Essentially they are booked when labs are built or maybe a bit later. If they are already booked then there are no competition between customers.

> When AMD reserves wafers at TSMC, that's capacity that is denied to Apple/Nvidia/Intel/etc. and vice versa.

No, they just build more fabs.

> Are you arguing that AMD is going to produce 128core zen4s and then cripple perfect dies down to 16c core parts instead of taping out multiple designs that use the exact same logical blocks, when all that lost real estate could have instead been used for GPUs or ~5-7 additional CPUs?

No, I'm saying that if you don't want shortages you don't disrupt markets by printing trillions of unbacked money and using all tricks you can imagine to avoid 40-50% inflation… mechanism where you have to calculate number of chips you need to order five years in advance but where buyers expect lead times measured in weeks works if you can predict number of buyers, but if you break that mechanism… it stops working. This is completely unrelated to selling crippling CPUs. Ryzen 5 5600X and Ryzen 9 5900 are still being produced and sold despite the fact that you can sell the exact same chips as Ryzen 7 5800X and Ryzen 9 5990X.

Turning 128 core chip into 16 core chip wouldn't make any sense because AMD embraces chiplets architecture thus you can just use more or less chiplets. But if you want to shave off 2 cores or 4 cores... AMD does that and is happy to sell these at discount prices.

Thoughts on software-defined silicon

Posted Feb 19, 2022 19:07 UTC (Sat) by Cyberax (✭ supporter ✭, #52523) [Link] (4 responses)

> It would cost you a lot. If you would stop selling $299 Ryzen 5 5600X (which is, essentially, $449 Ryzen 7 5800X with two cores disabled) and would start selling Ryzen 7 5800X for $399 instead of having these two… you would lose both on people who were ready to spend $449 and on ones who were ready to buy something for $299.

Except that you can undercut Intel and sell Ryzen 7 5800X for $299 and immediately crush Intel. Otherwise Intel will just push you out with cheaper and faster CPUs.

In reality, cores are usually locked because they fail internal QA - this is indeed a perfectly valid strategy.

> I took example from AMD book, not Intel book to show that as competition grows more acute the need to disable features and sell crippled product grows, not diminishes. Market segmentation is powerful tool.

Not following. Once the competition starts biting, there's a pressure to unlock more and more features on the low-end of the spectrum.

Thoughts on software-defined silicon

Posted Feb 20, 2022 0:40 UTC (Sun) by khim (subscriber, #9252) [Link]

> Except that you can undercut Intel and sell Ryzen 7 5800X for $299 and immediately crush Intel.

Seriously? Immediately crush Intel? Weren't you celebrating 25% market share which AMD achieved in four years after introducing CPU which totally kicked Intel's ass?

If you would start selling Ryzen 7 5800X for $299 the only thing you would achieve would be stiffing your own R&D. Which would mean that your next CPU would be worse that CPU of competitors and you would “crush” yourself instead.

> In reality, cores are usually locked because they fail internal QA - this is indeed a perfectly valid strategy.

How do you know? We know for a fact from the times when it was possible to enable them that they were perfectly functional back then. We know from sales figures today that it's, most likely, still true today (Ryzen 5 5600X sales are much higher than Ryzen 7 5800X). Nope, these cores are not locked because they fail QA. They are locked to be able to sell CPUs at different price points.

> Once the competition starts biting, there's a pressure to unlock more and more features on the low-end of the spectrum.

Where would that pressure come from? It's zero-sum game. We only have two competitors (in mobile space there are more, but still not that many). If you stop offering CPUs at different price points then you wouldn't, suddenly, get more money from selling more CPUs because you couldn't produce more CPUs at the snap of fingers. You pay a lot to ramp up capacity and it happens slowly. If you would deprive yourself from receiving more money they you would just fail to produce enough CPUs, leave money on the table and would lose in the next round of competition.

That's how we ended up with just two manufacturers of x86 CPUs BTW.

Thoughts on software-defined silicon

Posted Feb 20, 2022 0:44 UTC (Sun) by excors (subscriber, #95769) [Link]

> Except that you can undercut Intel and sell Ryzen 7 5800X for $299 and immediately crush Intel.

Why would you want to crush Intel in the lower-end market segment where (per the economy seat analogy) their profit margin is approximately zero? The production cost of your 8-core chip will be similar to their 8-core-with-2-perfectly-good-ones-disabled chip, so you'll make no profit either. And now you can't do any market segmentation yourself, because you're already selling your most powerful chip for no profit. That seems much worse than copying Intel's segmentation strategy and getting a small share of high-margin segments.

Thoughts on software-defined silicon

Posted Feb 21, 2022 4:35 UTC (Mon) by timrichardson (subscriber, #72836) [Link] (1 responses)

IN pricing theory, you want to extract the maximum value. If you sell a product at only one price, you are forced to compromise at both ends: there are potential customers who would pay above your marginal price, so are potentially profitable but don't buy because your offered price doesn't meet the value they see in the product, and you leave money on the table from customers who have more value in your product than what you charge; they would have paid more if you asked, but you didn't.

The conventional answer to is create differentiated products at different price points. Intel does this, nothing new. It is commonly accepted that this is something like an happy accident of the variation in how CPUs are made. A comment above says that this is greatly exaggerated but even if not, the distribution of different working cores is not a random accident: it would be a deliberately chosen manufacturing strategy affected by how the production process is configured. I doubt that Intel or AMD is very surprised by the output they get, and I expect they could tweak their production process to avoid nearly all locked cores, although at the cost of lower total output ( I have some manufacturing experience behind that comment, but it think it is not a controversial statement). The difference between accepting binned manufacturing output or achieving the same thing with software seems really invisible to me. I find it ironic that a computer science community is having trouble with the concept of abstracting hardware into software.

Thoughts on software-defined silicon

Posted Feb 21, 2022 13:40 UTC (Mon) by gnb (subscriber, #5132) [Link]

Whether the difference is really invisible depends a lot on the implementation: are the feature enablements being sold liable to expiry or revocation by the vendor? If so the difference between that and actually owning the feature seems pretty clear-cut. I suspect what is making a lot of commenters on this article uneasy is a suspicion that this is part of a move to a rental model.