Intel Demonstrates 65W Broadwell-K Socketed Processors at GDC 2015

[cytg111]

I wonder what kinds of products he's secretly urging Intel to make.

- something smooth, lubricated and with an intel-sticker-sticker?

 

[Ajay]

I wonder what kinds of products he's secretly urging Intel to make.

- something smooth, lubricated and with an intel-sticker-sticker?

While witeken's post does seem out of context - I see no need to start 'ganging' up on him. This is a tech forum, not P&N - please do not drag this thread down to that level. This happens too much as it is, IMHO.

 

[witeken]

And is 100.000 of those witeken wearing different shades, mustaches and wigs?

It's even more: 250k.

Source: http://forums.anandtech.com/showpost.php?p=36922359&postcount=22.

 

[DrMrLordX]

Sorry, the mental image of 250k people wearing fake moustaches in a massive focus group convention is making me laugh a little too much right now. Kind of like Lego LotR in "disguise" mode.

. . . *cough*. Right then. So am I the only one who is still simultaneously concerned-yet-enthused by Broadwell? I mean, on the one hand you have an unlocked Iris Pro part which could be majorly cool (maybe), but on the other hand you have Xeon-D (also an interesting part in its own right) reinforcing the notion that Intel can't get Broadwell chips above a clockspeed of around 3 ghz at a reasonable voltage.

The Iris Pro demo was cool, but they wouldn't talk clockspeeds or anything.

 

[Enigmoid]

I would guess that the 14 nm low power process was obviously prioritized over high power given intel's ambitions in mobile. Get high performance low power high margin parts out (U chips and xeon-D followed by low margin cherry trail) then follow up with high performance high power silicon.

If I had to guess its more the heat density problem than voltage. With 14 nm, a BW core + L3 shrinks to ~10 mm^2. With 4 cores clumped together you would get something like 60W+ in a piddly 40 mm^2, making it very difficult to cool.

 

[Fjodor2001]

I would guess that the 14 nm low power process was obviously prioritized over high power given intel's ambitions in mobile. Get high performance low power high margin parts out (U chips and xeon-D followed by low margin cherry trail) then follow up with high performance high power silicon.

If I had to guess its more the heat density problem than voltage. With 14 nm, a BW core + L3 shrinks to ~10 mm^2. With 4 cores clumped together you would get something like 60W+ in a piddly 40 mm^2, making it very difficult to cool.

If it's a 14 nm issue, then I don't see how it can be solved with Skylake either. Will we have to wait for Cannonlake until we see desktop performance topping Haswell? Or will Cannonlake even exaggerate the problem further, being even more dense and thus having even higher heat density.

Maybe Skylake IPC improvements will be high enough to compensate for lower frequency, but I doubt it. At least not for general code.

 

[ehume]

. . . If I had to guess its more the heat density problem than voltage. With 14 nm, a BW core + L3 shrinks to ~10 mm^2. With 4 cores clumped together you would get something like 60W+ in a piddly 40 mm^2, making it very difficult to cool.

That sounds about right. Temp-limited; not Voltage-limited or clock-limited.

I wonder if SB represented the peak in clockspeeds. We get equivalent instruction speeds as we move from IB to HW, but the top clocks get slower and slower (DC being a reset). IDC warned about this in his thread where he looked at generations of cpu's and temps. He had this scissors graph that showed size of the process vs temps generated (I don't have the reference right now). And adding AVX and AVX2 certainly jacks up the temps. I am running some heatsink comparisons right now, and when you run Linpack, Linpack+AVX and Linpack+AVX2 you seen ever-mounting temps.

Even without Intel focusing on mobile, with the heat density issue the max temp/Voltage for even the best cooled chips may simply, inevitably fall. And while that may be good for laptops and computing in general, it is bad for e-peen.

 

[Enigmoid]

If it's a 14 nm issue, then I don't see how it can be solved with Skylake either. Will we have to wait for Cannonlake until we see desktop performance topping Haswell? Or will Cannonlake even exaggerate the problem further, being even more dense and thus having even higher heat density.

Maybe Skylake IPC improvements will be high enough to compensate for lower frequency, but I doubt it. At least not for general code.

They can move the cores around so you don't have 4 cores clumped together.

The core structure could be rearranged so that the very 'hot' parts of the core are isolated from each other and surrounded by 'cool/cooler' silicon.

 

[ClockHound]

Even without Intel focusing on mobile, with the heat density issue the max temp/Voltage for even the best cooled chips may simply, inevitably fall. And while that may be good for laptops and computing in general, it is bad for e-peen.

Are you saying our e-peen could fall off? That is indeed bad news. ;-)

 

[Fjodor2001]

They can move the cores around so you don't have 4 cores clumped together.

The core structure could be rearranged so that the very 'hot' parts of the core are isolated from each other and surrounded by 'cool/cooler' silicon.

Yeah, I guess. But why didn't they do that alread on the Broadwell design then? That issue ought to be well known by now, and I bet they have tools simulating and detecting heat hotspots on designs.

 

[witeken]

Yeah, I guess. But why didn't they do that alread on the Broadwell design then? That issue ought to be well known by now, and I bet they have tools simulating heat hotspots for a particular design.

All your negative comment about 14nm products rest on the assumption, incorrectly, that there are any problems with 14nm which make it worse than 22nm.

 

[Dave2150]

All your negative comment about 14nm products rest on the assumption, incorrectly, that there are any problems with 14nm which make it worse than 22nm.

We're just looking at confirmed facts.

It's a fact that so far the maximum clock speed of their 14nm process is 3.1Ghz (3.4Ghz turbo) for a dual core part.

We haven't seen any quad-core parts yet, and the 8 core parts have a max clock of 2.2Ghz so far (Xeon D).

 

[witeken]

We're just looking at confirmed facts.

It's a fact that so far the maximum clock speed of their 14nm process is 3.1Ghz (3.4Ghz turbo) for a dual core part.

We haven't seen any quad-core parts yet, and the 8 core parts have a max clock of 2.2Ghz so far (Xeon D).

Instead of blaming, unfoundedly so, Intel's manufacturing process, why not blame their empirically conformed bad yields? BDW-U/Y has a smaller die area. Further, that 3.1GHz you quote is a 28W SKU. Lastly, it isn't far-fetched to release Broadwell Xeon-D, which is based on Broadwell-U/Y and thus have lower clock speeds and TDPs.

 

[Fjodor2001]

All your negative comment about 14nm products rest on the assumption, incorrectly, that there are any problems with 14nm which make it worse than 22nm.

How do you know that it's incorrect? Do you have some benchmarks or data backing that statement up?

 

[Dave2150]

Instead of blaming, unfoundedly so, Intel's manufacturing process, why not blame their empirically conformed bad yields? BDW-U/Y has a smaller die area. Further, that 3.1GHz you quote is a 28W SKU. Lastly, it isn't far-fetched to release Broadwell Xeon-D, which is based on Broadwell-U/Y and thus have lower clock speeds and TDPs.

I'm not sure what to blame. I just know there aren't any high clocked Broadwell parts available, announced or demonstrated thus far. Also no desktop Broadwell part available either.

There was an Intel article that showed that their 14nm yields were 1 year behind where they would like, though that article/slide is quite old also now - so again we have no idea of their current yields.

 

[Enigmoid]

We're just looking at confirmed facts.

It's a fact that so far the maximum clock speed of their 14nm process is 3.1Ghz (3.4Ghz turbo) for a dual core part.

We haven't seen any quad-core parts yet, and the 8 core parts have a max clock of 2.2Ghz so far (Xeon D).

All current 14 nm intel chips on the market are using the low power process. With that in mind, perhaps a 3.1 ghz ceiling is not surprising.

 

[Fjodor2001]

All current 14 nm intel chips on the market are using the low power process. With that in mind, perhaps a 3.1 ghz ceiling is not surprising.

But if there were no problems reaching higher frequencies with the high performance 14 nm process, we'd already see chips on the market by now.

 

[ehume]

Are you saying our e-peen could fall off? That is indeed bad news. ;-)

:biggrin::biggrin::biggrin:

 

[frozentundra123456]

We're just looking at confirmed facts.

It's a fact that so far the maximum clock speed of their 14nm process is 3.1Ghz (3.4Ghz turbo) for a dual core part.

We haven't seen any quad-core parts yet, and the 8 core parts have a max clock of 2.2Ghz so far (Xeon D).

I in fact dont know why there are not quad core or hi frequency parts out. But your reasoning is somewhat circular. Just because you havent seen a part yet, doesnt mean it is not feasible or wont appear. The "confirmed facts" are that 14nm is late and had (or still has) poor yeilds. There are not "confirmed facts" about how the chip will clock at higher power envelopes or the performance of quad core parts. Am I expecting great performance from quad core parts? No, not if you mean a great improvement from Haswell.

But I am keeping an open mind until the products actually appear.

 

[Dave2150]

I in fact dont know why there are not quad core or hi frequency parts out. But your reasoning is somewhat circular. Just because you havent seen a part yet, doesnt mean it is not feasible or wont appear. The "confirmed facts" are that 14nm is late and had (or still has) poor yeilds. There are not "confirmed facts" about how the chip will clock at higher power envelopes or the performance of quad core parts. Am I expecting great performance from quad core parts? No, not if you mean a great improvement from Haswell.

But I am keeping an open mind until the products actually appear.

It's fact that the CPU's mentioned don't exist yet. Intel haven't released them. I guess you're implying that Intel could release them, if they wanted to, but are choosing not to at the moment? Perhaps to clear more Haswell stock, or to make higher profits from selling higher yielding 22 parts?

 

[Enigmoid]

But if there were no problems reaching higher frequencies with the high performance 14 nm process, we'd already see chips on the market by now.

You still don't get it.

There are two 14 nm processes. low power and high power/performance.

Given intel's ambitions, it is extremely likely that low power was prioritized and pushed upwards on the roadmap. Hence all 14nm products intel is currently shipping are made on the 14nm low power process. 14nm high performance hasn't been launched yet. There are no current chips made on the high performance process so how can you say that the 14nm high performance process cannot reach higher frequencies?

Very likely 14nm high performance is suffering from heat density problems (which is not a process problem) and intel needs some more refinement. That doesn't mean that its impossible or a failure.

 

[Dave2150]

You still don't get it.

There are two 14 nm processes. low power and high power/performance.

Given intel's ambitions, it is extremely likely that low power was prioritized and pushed upwards on the roadmap. Hence all 14nm products intel is currently shipping are made on the 14nm low power process. 14nm high performance hasn't been launched yet. There are no current chips made on the high performance process so how can you say that the 14nm high performance process cannot reach higher frequencies?

Very likely 14nm high performance is suffering from heat density problems (which is not a process problem) and intel needs some more refinement. That doesn't mean that its impossible or a failure.

That's personal opinion. Personally, I think it's a complete failure that it's almost been 2 years since mainstream desktop haswell launched, with no product to replace it, apart from the enthusiast grade x99.

 

[Dave2150]

You still don't get it.

There are two 14 nm processes. low power and high power/performance.

Given intel's ambitions, it is extremely likely that low power was prioritized and pushed upwards on the roadmap. Hence all 14nm products intel is currently shipping are made on the 14nm low power process. 14nm high performance hasn't been launched yet. There are no current chips made on the high performance process so how can you say that the 14nm high performance process cannot reach higher frequencies?

Very likely 14nm high performance is suffering from heat density problems (which is not a process problem) and intel needs some more refinement. That doesn't mean that its impossible or a failure.

If what you say is true and Intel do have two different 14nm processes, then it would also have been true for previous generations.

Instead, we saw the desktop and mobile parts launched together, or within the same quarter at least. I dont think Intel would suddenly stop doing this, after all the previous generations of the Core architecture have had similar release cadences.

Far more likely that 14nm just didn't let them high the same clocks speeds as 22nm did.

 

[NTMBK]

Intel had several different variations of 22nm, too. Haswell was on a different process variant from Bay Trail. Density vs. performance.

 

[DrMrLordX]

Far more likely that 14nm just didn't let them high the same clocks speeds as 22nm did.

There's another, potentially scarier possibility: what if Intel realizes that delaying 14nm products and sticking with 22nm Haswell/Haswell-E for a bit longer is their best financial bet in every arena outside of mobile/low-power laptop?

Are they going to sell any/many more desktops with Broadwell than they will with Haswell? Think about what Idontcare was saying, especially in the Q12015 financial thread: Haswell Refresh is only just now catching up with them.

Sure the 14nm they're using for Broadwell-U/Broadwell-Y may not be adequate for high-performance desktop and server processors. They can always respin 14nm if they have to, and given what NTMBK just said, they've done it before on other processes. Trying to make guesses about the limitations of ALL their available process tech based on a selection of mobile CPUs and Xeon D only takes us so far. The subject of this thread - Broadwell Iris Pro - ought to tell us more in three months, especially since we can (presumably) get our hands on retail examples and experiment with clockspeeds and voltage to learn the chip, so to speak.

So be patient, we'll find out more very soon.