Intel investing $1B to meet 14nm demand

[Kenmitch]

https://hothardware.com/news/intel-chip-supply-woes-amd-ryzen-sales-september-etailer

lol

 

[Mopetar]

They do this because of production requirements. The Demand for cheaper chips always outweighs the demand for higher end chips. So you bin down meet that demand. Script flips on these 28c dies as on the Server market that needs these types of CPU's the upper end has higher demand. I think Intel will be making these wafers just for the handful of 28c dies they can get just to meet demand. But for a 1600x and 2600x yeah 80% of them are probably working 8 core dies.

The server market is no different than any other. It’s a more expensive product, but not a luxury product.

Look at the cost difference between Intel’s top bin server chips and what’s below them. Losing 4 cores off the full 28 cuts the price in half and you get a clock bump on that cut chip that makes it about 93% of the theoretical performance.

The problem is that these 28 core dies are massive and that’s okay if you’re making your mainstream product on the new node and have a lot of spare capacity on your old, but mature node. However, Intel is caught in a difficult situation where they are having to make all of their products on the same node and they don’t want to give AMD any quarter in any category.

 

[DrMrLordX]

https://hothardware.com/news/intel-chip-supply-woes-amd-ryzen-sales-september-etailer

lol

Bob Swan sure is being optimistic eh? Hate to tell the guy, but it looks like not so many people are choosing Intel . . .

However, Intel is caught in a difficult situation where they are having to make all of their products on the same node and they don’t want to give AMD any quarter in any category.

Intel apparently doesn't have a choice! People who want to buy will buy something, and they aren't going to wait forever for Intel to make up for those production shortfalls. How long do you suppose it's going to take for that $1 billion investment to turn things around?

Interestingly enough, AMD has an opportunity here to take over the low-end which is not being well-served by Intel at the moment. That is the segment Su specifically wanted to avoid last year at the Ryzen launch. The reality is that any CPU AMD chooses to make on 12nm with 4c/4t or 4c/8t is going to work out just dandy. Zen+ has good characteristics in the 3 GHz range, the dice are relatively cheap, and there's no real need for an iGPU for laptops or cheapo desktop systems since OEMs seem hell-bent on getting a surplus dGPU from the warehouse and throwing that in anyway.

AMD may accidentally take over the low-end with chips that last year would have been seen as mid-range. Down-binned, underclocked/undervolted versions of the 2300x and 2500x could take over a ton of market share. AMD has Raven Ridge too, so they have other options, but if AMD wants a simple approach, there it is.

 

[prtskg]

Looks like the price hike is hitting all of Europe.

Mindfactory is seeing even worse price increases of Intel chips at 50% average. And consumer sales are moving over to AMD. Intel is probably trying to keep their larger OEMs satisfied, mainly at the high end, but even my new HP Zbook ordered by my company is being pushed back further and further as well due to supply constraints of the Intel chip.

HP is getting special love from Intel IMO. Some are saying they received less Xeon chips from Intel compared to other companies leading to 13% loss in their market share. Hence HP is recommending Epyc more now.

 

[wahdangun]

The server market is no different than any other. It’s a more expensive product, but not a luxury product.

Look at the cost difference between Intel’s top bin server chips and what’s below them. Losing 4 cores off the full 28 cuts the price in half and you get a clock bump on that cut chip that makes it about 93% of the theoretical performance.

The problem is that these 28 core dies are massive and that’s okay if you’re making your mainstream product on the new node and have a lot of spare capacity on your old, but mature node. However, Intel is caught in a difficult situation where they are having to make all of their products on the same node and they don’t want to give AMD any quarter in any category.

No, in this market license cost can be majority of the cost, so even if the cpu was twice as expensive but can deliver 30% of more performance for the targeted application, they will will choose the most expensive one, this is why server part can be so expensive because realistically software cost sometimes can be more insane

 

[arandomguy]

Outside of APUs, AMD only has to fab one (1) die. Their ability to adjust to varying demand is much quicker than Intel. Just fab the die and adjust the packaging step to the product you want.

Over the medium to long term they can adjust and fill the increase in demand on their own chips pretty well. The problem is the short term especially as the current quarter is the busiest shopping quarter. There is also the issue of how aggressive/risky they would want to be in any ramp up in chasing market share. Bad prediction on Intel's capacity will result in themselves running over supply and crashing their own margins.

Any ramp up to product on shelves is still going to be measured in the months range. So unless they've predicted in advance then CPU supply is going to be tighter for both sides this quarter then it otherwise could would have been. Typically in Q4 we see street prices fall below MSRP as everyone competes for volume. Intel's own 9xxx release should drive down street prices as well normally. We saw both these in play last Q4.

So consumers will be worse off regardless of brand this quarter if Intel's lack of supply is an issue. Even if Ryzen 2xxx prices hold at MSRP that is still a net loss compared to what would have been. The various computer component shortages the last few years have really been bad for consumers.

 

[Topweasel]

The server market is no different than any other. It’s a more expensive product, but not a luxury product.

Look at the cost difference between Intel’s top bin server chips and what’s below them. Losing 4 cores off the full 28 cuts the price in half and you get a clock bump on that cut chip that makes it about 93% of the theoretical performance.

The problem is that these 28 core dies are massive and that’s okay if you’re making your mainstream product on the new node and have a lot of spare capacity on your old, but mature node. However, Intel is caught in a difficult situation where they are having to make all of their products on the same node and they don’t want to give AMD any quarter in any category.

Nope. Wrong. A couple of grand makes little difference in the long run. Performance/Power/Density matters more so once you get up to this point demand is going higher on a 28c product than the 26, 24, 22 core versions.

Edit: I should note. The reason Intel will sell a bunch of the other time is if the shortage is bad enough the lead times might be to long for certain projects. But again even if I am wrong. Which I am not. But lets say I am. The demand for the full speed full 28c dies will exceed the demand presented by the wafers in comparison to the demand of other chips. This gets worse when you now need 2 of those per CPU.

 

[Mopetar]

No, in this market license cost can be majority of the cost, so even if the cpu was twice as expensive but can deliver 30% of more performance for the targeted application, they will will choose the most expensive one, this is why server part can be so expensive because realistically software cost sometimes can be more insane

That really depends on how the licensing is done, assuming it exists at all as not every server runs software that requires license fees. Some software uses a per core model for licensing, in which case the widest chip possible isn't necessarily the best option if you can get a chip with fewer cores that are individually more powerful or have higher clock speeds. Other software uses a per CPU license model in which case chips with more cores are going to be better. Some software just charges based on users and leaves it up to you to pick whatever hardware you want to run it on.

Just as with the consumer market, there are different use cases. For some people it won't make sense to get the CPU with the most cores, especially if it costs that much more and the other cost factors aren't enough to outweigh the additional up-front costs.

Nope. Wrong. A couple of grand makes little difference in the long run. Performance/Power/Density matters more so once you get up to this point demand is going higher on a 28c product than the 26, 24, 22 core versions.

All of that comes down to the amount of computational power needed, space availability, cooling costs, etc. At a certain point, the chip with the most cores is no longer worth it. Otherwise Intel would simply increase the price further since the demand is so much greater. Intel would have no incentive not to raise prices until they balance demand against the supply for their inventory. They maximize profits and they don't end up getting stuck with unwanted parts.

Also, if you're on a per core licensing cost model, more cores isn't necessarily better if the fewer cores can be clocked much higher, especially if licensing costs are high relative to costs for rack space, cooling, etc. Obviously you still always want as few defects as possible on a chip. You can always artificially bin a good chip, but you can't move the other direction.

 

[Topweasel]

All of that comes down to the amount of computational power needed, space availability, cooling costs, etc. At a certain point, the chip with the most cores is no longer worth it. Otherwise Intel would simply increase the price further since the demand is so much greater. Intel would have no incentive not to raise prices until they balance demand against the supply for their inventory. They maximize profits and they don't end up getting stuck with unwanted parts.

Also, if you're on a per core licensing cost model, more cores isn't necessarily better if the fewer cores can be clocked much higher, especially if licensing costs are high relative to costs for rack space, cooling, etc. Obviously you still always want as few defects as possible on a chip. You can always artificially bin a good chip, but you can't move the other direction.

Your points only work in a vacuum where these are the only dies Intel makes. They have LCC HCC and XCC dies. As we saw with the SL-X LCC and HCC dies, there is almost no power or heatsavings, and clock speed between lowest and highest core counts was at most like 100MHz. Some one looking for a balance of clock speeds and cores isn't going to be looking at a low end XCC chip. They will be looking at a high end HCC chip that can run significantly faster than a XCC chip with a bunch of cores disabled.

The market for XCC dies. Is about system density. Fitting as much workload into as small a package as possible. These are for incredibly large businesses to fill up large data centers where it's about how many they can fit per square foot and with available cooling. The reason Intel is getting all super worried is probably specifically because the 28c parts are demands way exceed production and if their next step is to glue two of them together that is just making a bad problem worse and if I am right I am guessing orders for those exceed that of other XCC configurations maybe even combined.

 

[Mopetar]

Your points only work in a vacuum where these are the only dies Intel makes. They have LCC HCC and XCC dies.

And yet Intel still has a product like the 8156 (https://en.wikichip.org/wiki/intel/xeon_platinum/8156) which has 4C/8T at 3.7 GHz. Something has to be massively wrong with a die for a bin like that to exist for non-artificial reasons. The funniest bit about that is that chip has a list price of ~$7,000. That's about $400 less than an 8170 that has 26C/52T at 2.1 GHz. There are plenty of other configurations that sell for well under the $7,000 that a 8156 goes for. If no one wanted to buy that chip, it wouldn't exist. Yes, Intel probably sells more 8170 chips than it does 8156, but the market isn't as simple as you want to paint it to be. They'll certainly artificially bin chips if people want to pay for them, and obviously someone needs a Xeon Platinum that can hit 3.7 GHz even if it only has 4 cores more than they need one that has 24 cores, but much lower clocks.

The reason Intel is getting all super worried is probably specifically because the 28c parts are demands way exceed production

That's what I've been saying all along, but in a much broader sense. They wouldn't have this problem at all if 10 nm was ready to go. Instead they have to try to fit all of their different product lines onto the same node and a ~700 mm^2 die eats up loads of wafers. The yields on 14+++ are probably really good, but it still doesn't matter because even if you can get most of your chips into bin that offers that best margins, you can't possibly make as many as you want when your other product lines also need to share that production capacity. Even if they shift production to meet demand in the server market, they'll just end up with shortages somewhere else.

 

[LTC8K6]

I think the 8156 has to be the result of bad dies. It makes absolutely no sense at all for anyone to buy such a chip.
The 8158 is exactly the same price and you get 12 cores at 3 ghz.

 

[Mopetar]

I think the 8156 has to be the result of bad dies. It makes absolutely no sense at all for anyone to buy such a chip.
The 8158 is exactly the same price and you get 12 cores at 3 ghz.

Obviously Intel wouldn’t cut a full die to make such a part unless they have more than they can sell, but such a die that would make a 8156 or 8158 is unlikely to occur naturally. They almost assuredly hobble dies that could be sold in other bins.

Both of those parts make perfect sense in a world where there’s some software you need with a hideously expensive per core licensing cost. If you need that software then your best value is few cores with high clock speeds as opposed to more cores with lower clock speeds.

I’m not sure what what needs a Xeon Platinum as there are other low core Xeon chips with high clocks and much lower clock speeds. There’s a 4C/8T Xeon Gold for about one-fifth the price, so I imagine it’s a niche chip. But Intel knows more than we do about all of their customers and their needs.

 

[jpiniero]

I’m not sure what what needs a Xeon Platinum as there are other low core Xeon chips with high clocks and much lower clock speeds. There’s a 4C/8T Xeon Gold for about one-fifth the price, so I imagine it’s a niche chip. But Intel knows more than we do about all of their customers and their needs.

The Gold one is basically the same chip, but it only has 4S and 2 UPI links.

 

[LTC8K6]

The Gold one is basically the same chip, but it only has 4S and 2 UPI links.

Yeah, if you are buying a 4 core Xeon because of per core costs, then it seems unlikely that you need or want the ability to add 3 or 7 more CPUs.

There are too many better choices for the 8156 to make any sense, imo.

 

[Mopetar]

Yeah, if you are buying a 4 core Xeon because of per core costs, then it seems unlikely that you need or want the ability to add 3 or 7 more CPUs.

There are too many better choices for the 8156 to make any sense, imo.

Even if your per core costs are high, you may still need or want more cores to run the software on if the work is valuable enough. Then it's a matter of whether its more cost effective to have to split that over twice as many servers or just eat the extra cost for more sockets.

If the 8156 didn't make any sense, Intel wouldn't sell it. It's quite unlikely that they manage to end up with a piece of silicon that's still functional, but only 4 of the 28 cores work, so they're certainly binning it. If they can make more money selling one of those at $7,000 than as a 24-core 8160 at $4,700, why would they not do so?

But we're all getting pretty off topic at this point.

 

[jpiniero]

If they can make more money selling one of those at $7,000 than as a 24-core 8160 at $4,700, why would they not do so?

I guess it's worth re-iterating that Intel's RCPs are extremely fake. The 8156 is probably way cheaper than something than a 8160 for an OEM to buy.

 

[ehume]

. . . Intel's RCPs . . .

RCP? I'm afraid my google-fu is not up to the task.

 

[jpiniero]

RCP? I'm afraid my google-fu is not up to the task.

Recommended Customer Price I guess.

 

[IntelUser2000]

I guess it's worth re-iterating that Intel's RCPs are extremely fake. The 8156 is probably way cheaper than something than a 8160 for an OEM to buy.

Pentium Silver J5005 has RCP of $161, but I can buy a motherboard that integrates the chip for $120. So yea.

 

[Dayman1225]

There are rumors of Intel's Fab in Ireland being expanded, It currently fabs 14nm chips, though I don't know this rumored expansion is necessarily for 14nm


and Intel's Management is holding a 'strategic' meeting with Israel's Prime Minister later this week, Intel's interim CEO is one of the executives going