Hacker News Reader

Comment by noobermin | Hacker News Reader

>Along with the cancellation of the 7LP, GlobalFoundries essentially canned all pathfinding and research operations for 5 nm and 3 nm nodes.

Is it just me or does this seem like a drastically bad move for the long term?

View on HN

Each generation of fab is significantly more expensive than the last. Pushing the edge of physics is expensive. Transistors are so small these days they can't use the normal light frequencies any more. Things like precisely focusing light gets trickier when extreme frequencies are used.

The result is you have to almost double your volume with each generation, as a result there are less and less fabs running the current process. Makes AMD decision to split off Global Foundries look pretty good in hindsight.

Even those companies with the leading process make a substantial number of chips on older process. So the bleeding edge CPU gets the latest greatest, but the chipset, flash chips, and memory chips are often a generation or more behind.

Seems realistic that if you are behind and don't have a huge customer (like apple or nvidia) lined up that you just save a few $billion and let TMSC have it. TMSC will of course charge more without competition, and make chips using TMSC less competitive. If Samsung can't compete with TMSC (which remains to be seen) TMSC might well delay future shrinks.

The market loves Moore's law, but the stress is really starting to show. Physics is starting to interfere with what the market wants. Things like CPU clock speeds stagnating, power per chip doubling for the first time in the newest generation, and of course the ever lengthening product cycles.

It does make you wonder when AMD and Intel double the normal CPU socket from 95 watts to 180 watts or so. What are they going to do for the next generation?

> It does make you wonder when AMD and Intel double the normal CPU socket from 95 watts to 180 watts or so.

could they, and still expect to sell? it seems hard to tell the data center and supercomputing customers that they're going to have install massively more cooling capacity.

ThreadRipper 2 is 250W and I don't see people complaining, so I wouldn't be at all surprised if K SKUs move up to higher TDP.

In servers it's already rumored that Intel will increase TDP to ~300W per socket. This doesn't require more cooling per se, you just fit fewer sockets into the data center.

These days TDP is a lie and everyone knows it. Both Intel and AMD are measuring TDP at base-clocks and their processors will happily pull another 75-100% over TDP while boost clocks are active, even without overclocking.

The idea that you could run an octocore in the same TDP as a quad-core is obviously not correct, and it's largely because Intel and AMD keep pushing the baseclocks down farther and farther.

That 95 watt number they provide is a minimum to provide the advertised speeds more than anything else. Current stock chips from both AMD and Intel already use much more than their rated amount if there is headroom thermally and there is power availability. The 8700k for example has no problem drawing >150 watts when turbo'ing all cores and being well cooled. AMD does this as well.

Advertised heat output should also be taken with a grain of salt. Both die size and material between the die and heatspreader can make two 95 watt chips have very different cooling requirements.

Not necessarily, I actually think this is a smart move on their part, TSMC is ahead on 7nm and FD-SOI, RF, SiGe, etc. are big, profitable and rapidly growing markets in which GF has a differential advantage.

Agree this a smart move. They should be targeting lower volume markets where the 30% decrease in per-die cost at 7nm isn't offset by the increased upfront costs. If they focus instead on decreasing fixed costs of lower volume ASICs, they will have competitive advantage in long-tail markets.

And maybe, if they will push their resources really hard on decreasing costs of lower volume ASICs, that would be a big win over GF in long term. It depends on progress in hardware synthesis tools and software mainly. If the tools like auto-generating optimized RTL from Haskell code would became mainstream, and cost for production of custom, reasonably small ASICs on modern fab processes drop by an orders of magnitude - we would get something similar to what happened to PCB recently, when there is no point to made it by yourself even if you need a dozen, just send gerber files to OSH Park, and get your PCBs for $5 in two days.

GloFo was never long-term sustainable, so they basically had a choice between going bankrupt in 1-2 years trying to do 7 nm or coasting on 14 nm and slowly withering away over 5-10 years.

GloFo is long-term sustainable, just not as a leading edge foundry. Old process tech doesn't go away, they just move to producing lower-margin products. 65nm is still in mass production. 14nm and, more importantly the fdx processes (as they have lower chip development costs and therefore are more attractive for the low-end market) will still be in full production 20 years from now.

Just not making fast cpus.

Exactly. GF is making lots of chips on 130 and 180 nm process. There’s a lot out there that isn’t bleeding edge CPUs. We’ve got power chips, port controllers, high speed linear re-drivers. Not every ASIC needs tiny transistors, in fact many use comparatively huge ones and it makes no sense to use an expensive process.

I don’t see GF going away any time soon. They’re a huge player with a lot of customers.

More info here: https://news.ycombinator.com/item?id=17851278

I think it is a great short and long term move. The new smaller processes aren't getting us to faster clock rates, they are making things cheaper. GF can get to the same point by optimizing their current processes.

Reply on news.ycombinator.com