Coffeelake and Donutlake

[batboy]

I'm in the process of getting an i9 7920X, y'all can have Coffeelake. That's what I decided to do after my sensitive ears (not big ears) heard all the rumors that aren't even worth repeating.

 

[wingman99]

I'm in the process of getting an i9 7920X, y'all can have Coffeelake. That's what I decided to do after my sensitive ears (not big ears) heard all the rumors that aren't even worth repeating.

What do you expect comparing a $359.00 i7 8700k to $1199.00 i9 7920X.

 

[batboy]

True, you make a valid point.

 

[Super Nade]

Is it just me or is this launch a tad underwhelming? No IPC improvements and the platform is not much of an upgrade over Z170.

 

[batboy]

Basically, Kaby Lake v2. Not that it's bad, it'll be a good overclocker. Just not anything to write home about. Especially when you have to buy a new chipset. Did I mention Intel make a lot on chipsets?

 

[wingman99]

Is it just me or is this launch a tad underwhelming? No IPC improvements and the platform is not much of an upgrade over Z170.

I think when it comes to silicon technology we are at a point where there is not much room for innovation left.

 

[mackerel]

I see CL as Skylake v3. At consumer level, what more do you want? The main selling point is 2 more cores.

As for IPC, you can look at it at an instruction level as well as overall. Instructions that have been around a long time most likely are about as optimised as they're going to get. The only way to significantly add IPC is to add new instructions to do specific things faster than multiple existing instructions. I love the AVX series for example, but new instructions take time to be used, and aren't universally applicable.

 

[Woomack]

Most improvements in last ~10 years were AVX instructions and faster cache. Not much else is better. There are of course power saving features or smaller process but that isn't affecting performance. This is why most users can't see big difference in last 5 Intel generations. Most changes are in chipsets and additional devices. We are also getting more cores just because single core performance won't go up much. I only wonder why software is not fully using more cores while marketing is pushing 6+ cores for some time already.

 

[mackerel]

It isn't easy to use multiple cores, and not all software tasks benefit from multiple-cores even if implemented. I think a lot of software that can use make use of multiple cores already do. The only difference I think we might see going forwards is how good is that multi-core scaling? For example, on Prime95 in multi-thread mode, I suspect there is some limitation in it that means once you go above about 8 cores per system, the scaling falls off. It can't efficiently spread the work out. Light workloads like Cinebench R15 are pretty much best case, in that each thread can work independently from others. With some other tasks, where there is inter-thread synchronisation required, things get difficult fast.

I'm waiting for the day when people wake up and realise core count isn't everything even if the software scales. It is only one contribution to total performance. Especially when we get to higher (double digit) core counts as mainstream, this will be more important to consider. As a generalisation, within a given architecture/process technology, more cores enables more throughput by offering more performance in a given power budget. But you need to scale well to achieve that, and it isn't easy.

 

[SPL Tech]

Most improvements in last ~10 years were AVX instructions and faster cache. Not much else is better. There are of course power saving features or smaller process but that isn't affecting performance. This is why most users can't see big difference in last 5 Intel generations. Most changes are in chipsets and additional devices. We are also getting more cores just because single core performance won't go up much. I only wonder why software is not fully using more cores while marketing is pushing 6+ cores for some time already.

Well Intel has to do something. They cant just release a new chip and say the only improvement is that it uses 5 watts less and supports 4 more PCI-e bands. Consumers expect performance updates with every chip so if they cant get higher clockrates they need more cores. I am a bit interested why they havent tried upping the cache more. Like 20 MB L3 and 1 MB L1 for a quadcore. Wonder how much it would help.

 

[SPL Tech]

Well everyone talked a big game about how Z370 mobos are going to be over $300. More than 50% of the boards on Newegg are under that price and there are several at or under $150.

 

[mackerel]

I am a bit interested why they havent tried upping the cache more. Like 20 MB L3 and 1 MB L1 for a quadcore. Wonder how much it would help.

It will cost to implement due to extra die area required, plus there are performance tradeoffs. Cache is generally faster when smaller. As it gets bigger, keeping track of where data is slows it down. Compare the L2 cache of Skylake-S and Skylake-X. X has the bigger cache, but the bandwidth dropped as a result. It is a performance tradeoff between cache hit rate and speed/latency. Outside of exotic Xeons and Skylake-X, you get typically 1.5 to 2.0MB L3 per core on desktop CPUs, going up to 2.5MB/core for historic midrange Xeons. In a way, more cores will mean more total cache, and adequacy of cache quantity is not a linear measure. You tend to either have enough to maintain high performance or you don't.

Personally I'd like to see Crystal Well make a return to desktop, as it was only ever implemented in Broadwell mainstream CPUs (and some mobile ones). That was a separate 128MB eDRAM on the same CPU package acting as L4 victim cache. It made a fair ram substitute, allowing a system with low performance ram to keep up with those with faster ram. Bandwidth was good compared to DDR3 system ram, and was rated at 50GB/s from memory which is about double that of DDR3 1600 dual channel. I think if it were to be revisited it would need to be higher than that now. DDR4 3200 would roughly equal it in bandwidth, although the eDRAM would be lower latency.

 

[Woomack]

Well Intel has to do something. They cant just release a new chip and say the only improvement is that it uses 5 watts less and supports 4 more PCI-e bands. Consumers expect performance updates with every chip so if they cant get higher clockrates they need more cores. I am a bit interested why they havent tried upping the cache more. Like 20 MB L3 and 1 MB L1 for a quadcore. Wonder how much it would help.

Point is that they did barely anything. Most performance gains are from faster internal connections and cache. Their marketing is working great and it's not hard to convince most users that new is better. If you were on any Intel event then you know how much BS they can say in one hour I will only say one thing what I heard ... Intel IGP is good graphics for gamers ... list is long.

As it was already said, larger cache = slower cache. They have to improve it in other way to make it at the same time faster and larger but it's not so easy. It's also one of the reasons why AMD FX was so slow ... while Intel had 2-4MB L2 then AMD was pushing 8MB. Also difference in popular software like games is visible because of cache. Many people are complaining because of Skylake-X slow L3 cache.

Most changes are in chipsets, pcie lanes, additional features/controllers, lower power usage, more cores etc. Performance on single cores is not much higher than couple of years ago but there is higher turbo clock, faster memory controllers and some other things.

Finally 8th gen of mobile Intel CPUs will have more cores. Since first Core gen almost all i3 and i5 and most i7 were 2 cores. Now i5 are 2/4 or 4+HT so something is moving.

In desktops there are less changes than in other series. Now you can get 16 core Atom CPU which has 32W TDP but has full functionality of higher series. Couple of years ago at this TDP were 2-4 cores. For home/office users it doesn't matter but in business in global scale it's huge difference. In laptops in the same TDP you can get 2x more cores what makes possible to build ultrabooks with more than 2 cores. Even though performance per core isn't much better then overall it's a big difference.

 

[EarthDog]

Well everyone talked a big game about how Z370 mobos are going to be over $300. More than 50% of the boards on Newegg are under that price and there are several at or under $150.

lol, who said that? Its the mainstream platform...

 

[Woomack]

lol, who said that? Its the mainstream platform...

I think he saw comments that higher series of motherboards on Z chipsets cost $300 as we were talking about it in couple of threads. It's because Z motherboards used to be much cheaper and nowadays highest versions cost as much as X series mobos. Price on Z mobos is going up each generation.

 

[Tír na nÓg]

I checke a few reviews, and there is something I can hardly understand: 8700K@5GHz Superpi scores (1M and 32M), are equal/lower than my [email protected].

CB R15
[email protected]: 1484
8700k@5GHz: +/- 1600 (depending on reviews)

FS Physics
[email protected] : 64FPS
8700K@5GHz : 67 FPS

Which means there is little to no improvement in IPC between HAswell-E and Coffee Lake...

Could it be because of dual vs quad channel memory?

I was really considering getting one, but even overclocking it to 5.2GHz would net a 10% only IPC increase compared to what I run now...

 

[EarthDog]

32M is more memory dependent.

IPC increases really werent tauted here. It was a core bump and move to 14++ for higher clocks, etc...

 

[Tír na nÓg]

32M is more memory dependent.

IPC increases really werent tauted here. It was a core bump and move to 14++ for higher clocks, etc...

OK, but we are talking 2 generations gaps here (HAswell->Broadwell->SL/KL/CL)...

SP32M 5GHz: 385s (5820k) vs 425s (8700k)
SP1M 5GHz: 7.293s(5820k) vs 7.37s (8700k)

 

[petteyg359]

It isn't easy to use multiple cores, and not all software tasks benefit from multiple-cores even if implemented.

The second part is partly true. Many games are dependent on user input to control flow, and you can't multithread user input. You CAN multithread the effects of user input, though. If you've designed your AI to actually play well rather than just cheating to implement difficulty levels, in a simultaneous turn-based game where you're not supposed to know what actions everybody else took until the next turn, every AI player should be running in its own thread taking their turns at the same time. If you're writing a GUI app in any language, you're probably already using multithreading even if you don't know it; there's an "Event Dispatch Thread" that deals with diplaying the GUI and user input, and then there's the main thread that you're actually doing whatever work you have in.

The first part is not so true. If all schools from primary up did a better job of teaching logic (obviously not possible, because that would deplete various parties' gullible voter pools), and CS/IT schools did a better job of pointing out how various workloads can be separated into steps... Almost anything can be split at least to two threads if only people were taught what to look for.

 

[EarthDog]

OK, but we are talking 2 generations gaps here (HAswell->Broadwell->SL/KL/CL)...

SP32M 5GHz: 385s (5820k) vs 425s (8700k)
SP1M 5GHz: 7.293s(5820k) vs 7.37s (8700k)

32M is memory dependent... big time.. trash it.

Not sure on 1M..Cache amount and speed is different, etc. but, unless you have the systems next to each other and the settings, comparing like this is anecdotal at best.

Again, IPC wasn't tauted here a couple % at best between them.