Intel's 10nm only coming to servers in 2020 with Ice Lake

[Taf the Ghost]

1 minute ago, The Benjamins said:

 

ok at stock: 2700x is 24% faster then a 8700k ( clock x cores)

2700x 8c @ 4.0 = 32

8700k 6c @ 4.3 = 25.8

 

OC: 2700x is 7.7% faster then a 8700k ( clock x cores)

2700x 8c @ 4.2 = 33.6

8700k 6c @ 5.2 = 31.2

 

Rumor: 9900k is 17.5% faster then a 2700x ( clock x cores)

9900k 8c @ 4.7 = 37.6

9700k 8c @ 4.6 = 36.8

If things scale appropriately, though the 9700k is going to be 8c/8t, which, for anything that can leverage beyond 8 threads, is going to add up pretty fast. In Cinebench R15, we can generally expect the 9700k to land around 1500 points. About 100 points ahead of the 8700k at stock. Given the likely soldered reality, there's actually not going to be a huge amount of headroom for OC'ing.

 

That should put the 9700k behind the 1700 at 4.0 Ghz and all of the 2700 SKUs. By a bit. 9900k will be faster by dent of ~15% higher clocks, though HT isn't as important on Intel as it is on AMD. (AMD's SMT can get up to 25% improvement depending on the application.)

[The Benjamins]

1 minute ago, Taf the Ghost said:

If things scale appropriately, though the 9700k is going to be 8c/8t, which, for anything that can leverage beyond 8 threads, is going to add up pretty fast. In Cinebench R15, we can generally expect the 9700k to land around 1500 points. About 100 points ahead of the 8700k at stock. Given the likely soldered reality, there's actually not going to be a huge amount of headroom for OC'ing.

 

That should put the 9700k behind the 1700 at 4.0 Ghz and all of the 2700 SKUs. By a bit. 9900k will be faster by dent of ~15% higher clocks, though HT isn't as important on Intel as it is on AMD. (AMD's SMT can get up to 25% improvement depending on the application.)

I was more point out the flawed math then my personal in depth estimates, I would add estimated IPC difference and HT vs SMT scaling.

[Taf the Ghost]

1 minute ago, djdwosk97 said:

Your numbers are off. An 8700k needs a 33% single core performance lead to equal a 2700x in multithreaded, perfectly scaling, loads. It might be easier to see if you use Ryzen as the baseline.

 

E.g.

 

let 1 Ryzen core at 4.2ghz be 1 unit of performance. Then a 2700x would be 8 units powerful.

 

An 8700k at 5.2 has a 24% clock speed advantage and let's just say a 5% IPC lead at the same clocks (I don't know what the IPC difference is with Ryzen 2). So then 1 8700k core would have 1*1.24*1.05=1.30 units of power. Multipled by 6 gives you 7.8 units -- just shy of a 2700x.

5% is about right. It's AVX loads and quick cache stuff that Intel gets the advantage. In the "saturated core" discussions, AMD's Zen core is actually more powerful than Intel's current Skylake cores, outside of AVX2 & above. Issue is Intel has better I/O & Cache systems, which is part of why AMD gets more out of SMT than Intel does.

[Taf the Ghost]

1 minute ago, The Benjamins said:

I was more point out the flawed math then my personal in depth estimates, I would add estimated IPC difference and HT vs SMT scaling.

Cinebench doesn't scale well on Intel with HT, but it does a lot more with AMD and SMT. A few programs get more out of the HT (Corona is one, I believe), but AMD's design is "wider" at the front, so there's reasons for that. It's part of why more Skylake cores really matters for Intel CPUs.

 

Unless Intel is rolling out some new Cache or other tech with Coffee Lake-R, the 8700k is actually the far better buy for the consumer. HT doesn't scale too great under most render loads (~35% in Cinebench going from 6c/6t to 6c/12t; roughly the same in other tasks), but the value of HT shows up in a lot of gaming situations because the GPU spams a massive amount of very small calculations rapidly. This makes putting 2 threads on the same core extremely valuable, given the much shorter latency between them.

 

The Windows Scheduler, when working properly, can provide a much better user experience when keeping small tasks grouped on the same core. That's part of the original "Ryzen just seems smoother" issue. Tasks kept within the CCX just respond much faster. Same thing. 

 

Clock for clock, unless Intel is up to something, these parts should be slower in Gaming. An 8c ring is slower. Same reason the 7700k at 5 Ghz was either even or better than the 8700k, with the exception of the few games that could leverage the extra cores. (Coffee Lake does have some improved Cache stuff, so it's not exactly the same Skylake core, but some of that is just down to node optimization.)

[AngryBeaver]

29 minutes ago, Brooksie359 said:

200% performance is 2 times the performance. You said 15% clock increase mean 84% more compute power which it doesn't. 15% increase in clocks means 15% increase in compute power. You can't just use percentages how you like. If someone says this cpu has 150% higher compute power that means you would simply multiply the other cpus compute power  by 1.5 and it should give you the compute power of the higher performing chip. So based on what you said you are simply wrong. If you had worded it differently maybe what you said would make more sense but in its current form it is incorrect. 

It is 2 times the performance... you are missing That we were comparing the performance of. I was not referring to the power of the ENTIRE cpu, but the power of the cores. So each CORE was X amount faster. So if you have a value of 150% for example that means you gained 1.5x more cores essentially. 200% CORE performance would mean you gained 2 more cores in performance.

 

[Taf the Ghost]

There's a reason the speculation about solder on the 8c part has been around for a while. It was really Intel's only "silver bullet" to make eight Skylake cores behave under load. These are going to draw around 250w at stock, under all-core load. Probably around 350w at 5.1 Ghz. For context, that's going to be more power than the upcoming 32c Threadripper, at stock, under load.

 

Solder is at least going to allow these not to hit thermal limits so quickly. Is a 240mm Rad a requirement to even think about OC'ing these?

[Brooksie359]

9 minutes ago, AngryBeaver said:

It is 2 times the performance... you are missing That we were comparing the performance of. I was not referring to the power of the ENTIRE cpu, but the power of the cores. So each CORE was X amount faster. So if you have a value of 150% for example that means you gained 1.5x more cores essentially. 200% CORE performance would mean you gained 2 more cores in performance.

 

You can't talk about the performance of a single core then take the fact that there are 6 cores into account. Each core was 15% faster if the frequencies are 15% higher. If you want to talk about the total compute power then talk about the total compute power. If you want to talk about the total compute power. Don't start talking about one metric and then use a completely different one. 

[AngryBeaver]

1 minute ago, Brooksie359 said:

You can't talk about the performance of a single core then take the fact that there are 6 cores into account. Each core was 15% faster if the frequencies are 15% higher. If you want to talk about the total compute power then talk about the total compute power. If you want to talk about the total compute power. Don't start talking about one metric and then use a completely different one. 

Or maybe you just follow the conversation instead of confusing the hell out of everything. Apparently you are the only one so far to miss the meaning of the prior conversations.

[Brooksie359]

2 minutes ago, AngryBeaver said:

Or maybe you just follow the conversation instead of confusing the hell out of everything. Apparently you are the only one so far to miss the meaning of the prior conversations.

Have you even read the last like 10 posts? The majority of them were about your flawed math. The only one confusing people is you so I am unsure why you would think it's me. 

[leadeater]

5 hours ago, yian88 said:

Im not buying that for a second unless you have some well documented/tests source for that. Im not saying dual channel was ever enough for certain tasks.

Just like first 8 core intel was an i9 HEDT and now they move it to mainstream dual channel RAM 8 core i9 9900k, i have no doubts 8-12 cores can run without issues with dual 3000+ mhz RAM for most  workstation tasks, sorry i have not documented on this issue so i dont know specifically for intel 14 nm kaby lake cores, how much bandwidth they use in various tasks, i would assume if you throw work 100% on 12 cores at the same time with intensive memory read/write dual channel is not enough but otherwise for workstation/multitasking it should be enough. By that i mean general software like CAD or photoshop or streamers/gamers dont need more than 2 channel for now.

There are members on this forum that can attest to this and have commented about it before.

 

 

http://www.int.washington.edu/PROGRAMS/12-2c/week3/joo_03.pdf

 

An applicable example for people here would be video encoding and I'll post below some gaming benchmarks that test this below, even though gaming is the least likely to be memory bound typically.

 

https://www.hardwaresecrets.com/does-dual-channel-memory-make-difference-in-gaming-performance/4/

 

10 or 12 cores with Dual Channel DDR4 is asking to be memory bound, it'll happen like it already can even on 4 cores. It would be a very unbalanced product so not likely to happen, either DDR5 will allow it or triple channel main stream.

[ravenshrike]

On 7/26/2018 at 3:00 AM, Master Disaster said:

Anybody ever thought Intel are playing the smart game?

No. That would be plausible if this were the first or second delay but 10nm in 2020 will put it 5 years behind schedule. Allowing AMD to hit Zen 2+ or start sampling Zen 3 is not a particularly well thought out maneuver and I doubt even Intel is that conceited.

[mr moose]

I think Intel genuinely hit a stumbling block they weren't expecting with 10nm.  Had it worked as it does on paper it would be almost in full production now and showing some impressive results.  However given the lack of information on the subject and the failure of many insiders to come good with "watch this space" tweets, I think it's fair to say they just went too dense for the node and are struggling to find a work around.   Maybe they'll jump to 7nm as some have predicted or maybe they'll find that workaround.

[Mira Yurizaki]

2 hours ago, mr moose said:

I think Intel genuinely hit a stumbling block they weren't expecting with 10nm.  Had it worked as it does on paper it would be almost in full production now and showing some impressive results.  However given the lack of information on the subject and the failure of many insiders to come good with "watch this space" tweets, I think it's fair to say they just went too dense for the node and are struggling to find a work around.   Maybe they'll jump to 7nm as some have predicted or maybe they'll find that workaround.

And I'm also under the impression that they've found it better to refine existing processes to get the most out of them rather than try to race to get to the smallest node. As an example, NVIDIA managed to attain the same performance while reducing the number of transistors by about 40%, boost the clock speed by about 28% (though this might be related to reducing the transistor count), all while doing this on the same process node as the previous generation.

 

Intel has a lot more to worry about than manufacturing. TSMC and GloFo can focus 100% on manufacturing techniques, more or less.

Edited July 28, 2018 by M.Yurizaki

[cj09beira]

15 hours ago, leadeater said:

10 or 12 cores with Dual Channel DDR4 is asking to be memory bound, it'll happen like it already can even on 4 cores. It would be a very unbalanced product so not likely to happen, either DDR5 will allow it or triple channel main stream.

even still amd already announced that the next epyc cpu will be 48 cores which means a 12 core mainstream cpu, which would help them compete against the 9900k

[leadeater]

47 minutes ago, cj09beira said:

even still amd already announced that the next epyc cpu will be 48 cores which means a 12 core mainstream cpu, which would help them compete against the 9900k

There has been no announcement from AMD about how many cores EPYC 2 will have, only the usual speculation/rumor mil. The way in which Zen uses memory is a bit different due to the CCXs and IF though. We're 2 weeks away from getting an indication of how memory bandwidth limitations show up with AMD Zen and TR2, double the cores without increasing memory bandwidth. TR2 will exacerbate the issue a bit more than increasing cores in the CCX due to remote memory for 2 of the dies (unless AMD has managed to work some pin out magic) but there will be plenty of core scaling benchmarks to look through and we should see tests where going above a certain number of threads gives no performance increase.

[Stefan Payne]

1 hour ago, leadeater said:

TR2 will exacerbate the issue a bit more than increasing cores in the CCX due to remote memory for 2 of the dies (unless AMD has managed to work some pin out magic) but there will be plenty of core scaling benchmarks to look through and we should see tests where going above a certain number of threads gives no performance increase.

They could redesign the Package so that every die gets a single Channel.

Shouldn't be that big of a deal.

 

 

As for the Remote Cores:
Remember MSI K8T Master?

That should give you an idea of what we could have...

[S w a t s o n]

19 hours ago, Taf the Ghost said:

These are going to draw around 250w at stock, under all-core load. Probably around 350w at 5.1 Ghz. For context, that's going to be more power than the upcoming 32c Threadripper, at stock, under load.

 

How did you come to 250W? As demonstrated by your comparison that is A LOT of power for 8 cores, even highly clocked ones. 350W seems almost impossible

[Taf the Ghost]

28 minutes ago, Swatson said:

How did you come to 250W? As demonstrated by your comparison that is A LOT of power for 8 cores, even highly clocked ones. 350W seems almost impossible

https://www.techspot.com/review/1497-intel-core-i7-8700k/page4.html

 

 

270 * (8/6) = 360. 

 

Now it's not going to be that much down the rails, that's probably going to max around 200w, but Skylake cores are thirsty. 14nm++ moved the frequency max efficiency point up, but pushing the silicon to 5.1 Ghz on 8 cores is going to use a lot of power.

 

https://www.overclock3d.net/reviews/cpu_mainboard/intel_8th_generation_core_i3_8350k_and_core_i5_8600k_review/14

 

If you go with OC3D's numbers, 33% higher is going to be pushing 390-400w. (This is going to massively depend on motherboard, but we're talking over 5 Ghz OC's on all-core.)

[S w a t s o n]

4 minutes ago, Taf the Ghost said:

snip

That's system consumption though. Both the graph and the oc3d link

https://www.anandtech.com/show/11859/the-anandtech-coffee-lake-review-8700k-and-8400-initial-numbers/5

This indicates the 8700k for example pulls package power of ~90W

[Taf the Ghost]

3 hours ago, cj09beira said:

even still amd already announced that the next epyc cpu will be 48 cores which means a 12 core mainstream cpu, which would help them compete against the 9900k

 

2 hours ago, leadeater said:

There has been no announcement from AMD about how many cores EPYC 2 will have, only the usual speculation/rumor mil. The way in which Zen uses memory is a bit different due to the CCXs and IF though. We're 2 weeks away from getting an indication of how memory bandwidth limitations show up with AMD Zen and TR2, double the cores without increasing memory bandwidth. TR2 will exacerbate the issue a bit more than increasing cores in the CCX due to remote memory for 2 of the dies (unless AMD has managed to work some pin out magic) but there will be plenty of core scaling benchmarks to look through and we should see tests where going above a certain number of threads gives no performance increase.

Epyc 2 will be find due to be Octa-channel. Generally, I don't think we've really found a way to truly memory-bound these CPUs yet. I think the I/O & Latency will be the bigger issues for the foregoing future. It looks like Zen3 on 7nm+ (or ++) should be the DDR5 shift in 2021.

 

The few places you can invent to memory-bound, for bandwidth, a task on Desktop, we can probably just say, "well, buy a Threadripper". We'll have to see, but I don't think it'll be an issue, unless they really do bring all 16c to AM4 before moving to AM5. 

[Taf the Ghost]

6 minutes ago, Swatson said:

That's system consumption though. Both the graph and the oc3d link

https://www.anandtech.com/show/11859/the-anandtech-coffee-lake-review-8700k-and-8400-initial-numbers/5

This indicates the 8700k for example pulls package power of ~90W

Power Consumption gets messy when you're talking about different parts, but we're talking about situations underload. Even with just a CPU load, it's going to be intense. We know what 8c Zen parts do, and Skylake cores are a good bit more power hungry.

[Zodiark1593]

5 minutes ago, Taf the Ghost said:

Power Consumption gets messy when you're talking about different parts, but we're talking about situations underload. Even with just a CPU load, it's going to be intense. We know what 8c Zen parts do, and Skylake cores are a good bit more power hungry.

Are Skylake cores hungrier under AVX load, or is this just some general integer + SSE load?

 

I know AVX tends to be a lot faster on Intel chips than on Ryzen, so it would make sense for Skylake cores to be hungrier when subjected to AVX loads.

[S w a t s o n]

3 minutes ago, Zodiark1593 said:

Are Skylake cores hungrier under AVX load, or is this just some general integer + SSE load?

 

I know AVX tends to be a lot faster on Intel chips than on Ryzen, so it would make sense for Skylake cores to be hungrier when subjected to AVX loads.

Yes the AVX2/AVX512 units just suck power but you can't use that as a baseline because that's not a very typical use case generally speaking

[Drak3]

1 minute ago, Zodiark1593 said:

Are Skylake cores hungrier under AVX load, or is this just some general integer + SSE load?

 

I know AVX tends to be a lot faster on Intel chips than on Ryzen, so it would make sense for Skylake cores to be hungrier when subjected to AVX loads.

Everything from Haswell up roughly matches Zen in power consumption at similar core count and clocks in non AVX loads, when looking at desktop chips.

[Zodiark1593]

7 minutes ago, Drak3 said:

Everything from Haswell up roughly matches Zen in power consumption at similar core count and clocks in non AVX loads, when looking at desktop chips.

Sort of about what I expected. Ryzen closely matches Haswell in everything but AVX anyway, as far as performance is concerned.

 

8 minutes ago, Swatson said:

Yes the AVX2/AVX512 units just suck power but you can't use that as a baseline because that's not a very typical use case generally speaking

Understandable. Wasn't sure how the comparison tests were done is all.