Intel's Future Chips: News, Rumours & Reviews

[goldstone77]

Live Delidding of Intel i9-7900X (ft. Der8auer)
Gamers Nexus

Intel still using cheap thermal paste

https://www.youtube.com/watch?v=I1Bv8Mxnnlc

Now THAT is a shame! Should turn a number of potential buyers off... if they even bother to look into it.
But come on, going that route on your high-high end products? WTH, Intel. May as well do it right in the first place, it'll only add like $20(?) to the overall price.

TIM is fine. OC potential is HUGE

https://youtu.be/kpoies2JcmI

Hmm... so I should at least be capable of 4.8 with the 7820 on a good air cooler(nh-d15 for ex.), 5.0 if I win the lottery.
And if the TIM is 'fine', then why so many complaints about it? Did they run into issues with solder? I can only imagine so, since they haven't used it since 2nd gen sandy bridge. But AMD is currently using it with their new chips...

HEDT processors were always soldered.

Solder hasn't been used on newer mainstream chips because the die is smaller. Soldering that small die to the IHS would cause problems in the long run because of cracks in the solder.

Is it time to switch to AMD? Has Intel lost its mind?
JayzTwoCents
Published on Jun 5, 2017
https://www.youtube.com/watch?v=QJNRtGo5IMc

Der8auer's words come to mind:

From my point of view Skylake-X is a pretty impressive CPU. We now have have a 10-core that's a lot cheaper than the Broadwell-E before and it clocks a lot higher. [...] So we had Broadwell-E before who could run like 4.3-4.4 GHz, it cost 1700€. Now we have Skylake-X which is a lot cheaper and we CAN push it to 5 GHz, so what's all this negative press about? I don't really understand it. So from my point of view this is a very impressive CPU, so we have very high single-thread performance on the 10-core and also high MT performance, which we didn't have before in a Broadwell-E

It looks as the x299 launch is putting nervous certain people.

I agree that the 10 core chip is going to be excellent! Even with the add cost it's going to be cheaper, and the better CPU than last gen 10 core. Der8auer is deflecting from the real negativity of this product line. The quality of the chip isn't what everyone is complaining about except that they used cheap thermal paste again instead of solder. The problem is with the entire product line as a whole. The 10 core option is now the first option you have with 44 PCI-E lanes, which effective makes it the entry level chip to get into HEDT with all the extra costs for raid puts it at $1,398. The 6, and 8 core are now hamstrung to 28 PCI-E lanes, and if you want the above mentioned full function of RAID add $399 to the total price and all of a sudden these prices are not that far away from last year and you lost PCI-E lanes. And the i5 and i7 are limited to just 16 PCI-E lanes, dual channel RAM, and a higher 112W TDP over previous gen. That doesn't make sense to anyone when you have to put that chip on a much more expensive board with functionality that can't be used by the i5 and i7. Also, the 14,16, and 18 core parts were added after the announcement of ThreadRipper, and there are no details on them yet.

 

[goldstone77]

CORE X IS HERE! Intel's i9 7980XE 18-Core CPU and X299 Platform
Paul's Hardware
Published on May 30, 2017
https://www.youtube.com/watch?v=TOtvKQWUqpM

 

[gamerk316]

TR4 is half the cost of LGA2066

Half the cost... For who exactly?

Also, a bit on a tangent: I've read that Intel is giving very obvious-subtle hints/winks at Qualcomm they will sue if they go ahead with the X86 emulator in Win10 RT-Reborn. I bet Miss Lisa Su is giggling in her office reading about that.

Cheers!

Legally, this is a little bit messy. Emulating the base x86 instruction set (386) isn't a problem, it's all the patented extensions (mainly, SSE, especially SSE2) where things get a little bit murky.

Look it it this way: AMD, as part of it's cross-licensing deal with Intel, has the right to use Intels SSE extensions in exchange for Intel having access to x86-64. Qualcomm is trying to emulate SSE without paying Intel a dime. So if you take Qualcomms side, then what's preventing Intel from "emulating" x86-64 or AMD from "emulating" SSE and bailing from the cross-license agreement altogether?

So yeah, theres some legal issues to work out. x86 as an instruction set isn't protected, but all the extensions almost certainly are. And that's where Qualcomm could find itself in a lot of trouble really quickly. You also need to consider the fact Qualcomm is under Anti-trust investigation in Europe, and this isn't going to help it's case out any,

So yeah, break out the popcorn; this one is going to get really ugly.

 

[-Fran-]

TR4 is half the cost of LGA2066

Half the cost... For who exactly?

Also, a bit on a tangent: I've read that Intel is giving very obvious-subtle hints/winks at Qualcomm they will sue if they go ahead with the X86 emulator in Win10 RT-Reborn. I bet Miss Lisa Su is giggling in her office reading about that.

Cheers!

Legally, this is a little bit messy. Emulating the base x86 instruction set (386) isn't a problem, it's all the patented extensions (mainly, SSE, especially SSE2) where things get a little bit murky.

Look it it this way: AMD, as part of it's cross-licensing deal with Intel, has the right to use Intels SSE extensions in exchange for Intel having access to x86-64. Qualcomm is trying to emulate SSE without paying Intel a dime. So if you take Qualcomms side, then what's preventing Intel from "emulating" x86-64 or AMD from "emulating" SSE and bailing from the cross-license agreement altogether?

So yeah, theres some legal issues to work out. x86 as an instruction set isn't protected, but all the extensions almost certainly are. And that's where Qualcomm could find itself in a lot of trouble really quickly. You also need to consider the fact Qualcomm is under Anti-trust investigation in Europe, and this isn't going to help it's case out any,

So yeah, break out the popcorn; this one is going to get really ugly.

I already have my popcorn!

Cheers!

 

[juanrga]

I agree that the 10 core chip is going to be excellent! Even with the add cost it's going to be cheaper, and the better CPU than last gen 10 core. Der8auer is deflecting from the real negativity of this product line. The quality of the chip isn't what everyone is complaining about except that they used cheap thermal paste again instead of solder.

He has demonstrated that soldering the chip only provided about 4% more performance. Therefore the usage of thermal paste is not any problem for the 99.99% of users. It must be a problem only for extreme overclockers looking for breaking worldwide records; this people will have to delid the chip and replace paste by liquid metal.

 

[-Fran-]

He has demonstrated that soldering the chip only provided about 4% more performance. Therefore the usage of thermal paste is not any problem for the 99.99% of users. It must be a problem only for extreme overclockers looking for breaking worldwide records; this people will have to delid the chip and replace paste by liquid metal.

You forgot how Ivy Bridge was very criticized because they went TIM instead of keeping solder from Sandy? I do remember all temperatures and TDPs went up and it was annoying for all people with K CPUs. Not exactly a reduced group of people, mind you, among the enthusiast (or upper mainstream?) group.

Cheers!

 

[okcnaline]

That is comparing 16C CPU to 18C CPU and ignoring the cost of the rest of the platform.

Yep, the rest of the platform which includes Intel SSDs, Intel "dongles", 2 more cores, and 4 extra threads. I don't know how enterprises would justify paying $300 and more for Intel SSDs just for 2 extra cores which can be paid for if they just have that one extra server.

 

[goldstone77]

I agree that the 10 core chip is going to be excellent! Even with the add cost it's going to be cheaper, and the better CPU than last gen 10 core. Der8auer is deflecting from the real negativity of this product line. The quality of the chip isn't what everyone is complaining about except that they used cheap thermal paste again instead of solder.

He has demonstrated that soldering the chip only provided about 4% more performance. Therefore the usage of thermal paste is not any problem for the 99.99% of users. It must be a problem only for extreme overclockers looking for breaking worldwide records; this people will have to delid the chip and replace paste by liquid metal.

Actually, it doesn't effect extreme overlockers like Der8auer, because they delid and use a conventional thermal paste. A lot of "K" owners would have a problem with it. Der8auer tried to justify the usage of paste, because of micro fractures years ago. He might be slightly bias since he benefits directly from Intel's use of paste for his overclocking desires. Ultimately, I'm willing to bet it's just cheaper and faster for Intel to use paste.

Start watching at 2:20
XOC Voltages Explained ft. Der8auer (7GHz LN2 Overclock)
Gamers Nexus
Published on Jun 11, 2017
https://www.youtube.com/watch?v=fGyoR7BPyS8

 

[8350rocks]

Is it time to switch to AMD? Has Intel lost its mind?
JayzTwoCents
Published on Jun 5, 2017
https://www.youtube.com/watch?v=QJNRtGo5IMc

Der8auer's words come to mind:

From my point of view Skylake-X is a pretty impressive CPU. We now have have a 10-core that's a lot cheaper than the Broadwell-E before and it clocks a lot higher. [...] So we had Broadwell-E before who could run like 4.3-4.4 GHz, it cost 1700€. Now we have Skylake-X which is a lot cheaper and we CAN push it to 5 GHz, so what's all this negative press about? I don't really understand it. So from my point of view this is a very impressive CPU, so we have very high single-thread performance on the 10-core and also high MT performance, which we didn't have before in a Broadwell-E

It looks as the x299 launch is putting nervous certain people.

Intel is quite nervous...certainly...

 

[juanrga]

Is it time to switch to AMD? Has Intel lost its mind?
JayzTwoCents
Published on Jun 5, 2017
https://www.youtube.com/watch?v=QJNRtGo5IMc

Der8auer's words come to mind:

From my point of view Skylake-X is a pretty impressive CPU. We now have have a 10-core that's a lot cheaper than the Broadwell-E before and it clocks a lot higher. [...] So we had Broadwell-E before who could run like 4.3-4.4 GHz, it cost 1700€. Now we have Skylake-X which is a lot cheaper and we CAN push it to 5 GHz, so what's all this negative press about? I don't really understand it. So from my point of view this is a very impressive CPU, so we have very high single-thread performance on the 10-core and also high MT performance, which we didn't have before in a Broadwell-E

It looks as the x299 launch is putting nervous certain people.

Intel is quite nervous...certainly...

It doesn't look so. They have a clear winner.

 

[YoAndy]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

 

[8350rocks]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

 

[jimmysmitty]

That is comparing 16C CPU to 18C CPU and ignoring the cost of the rest of the platform.

Yep, the rest of the platform which includes Intel SSDs, Intel "dongles", 2 more cores, and 4 extra threads. I don't know how enterprises would justify paying $300 and more for Intel SSDs just for 2 extra cores which can be paid for if they just have that one extra server.

This is not targeted towards enterprises and when they buy servers the hardware cost is not what matters or costs that much. It is the support that costs and matters.

Besides most servers don't come with every feature they support either. Bought a Dell server? Great. Want to use that iDRAC? Pay up. In the enterprise world there is a lot less complaining on costs and more focusing on what you need to work.

Yes cost does have a role but it is not like a person. People, unless insanely rich, focus on cost as their livelihoods matter based on what they have and spend. In a company if you need RAID5 or you need 2 more cores so you can make one server box do 9 different VM roles than you do it as in the end the cost will pay for itself.

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

Here is the problem. Productivity is usually, for the most part, in a business. The small amount of self employed people who buy their own systems will worry about how much to pay for each part or the overall system. Most companies do not. They will buy whatever is being sold as the best for their program and needs be it Intel, AMD or NVidia etc.

I also see people drooling over 64 PCIe lanes. Considering the piss poor or even non existent support for CFX/SLI in some recent games and the extremely poor scaling past 2 cards, why does it matter to have 64? Now if there were 64 and they could delegate them to other purposes, sure. But most CPU lanes are dedicated to the PCIe slots while the chipset PCIe lanes can be changed per the mobo vendor to make the overall platform more appealing.

But hey I am sure there are plenty of people who want quad SLI GTX 1080s.... Oh wait, only supports dual SLI. Maybe quad CFX Rx 580s? Possibly if the returns are worth it. Still have not ever seen a reason to go past 2 GPUs and even then support has been pretty shoddy, especially in DX12 where devs are more responsible for it.

In the end AMD better be careful. Everyone knows that Intel can do a price war and AMD might not survive it. They are just now getting back to a more competitive position. They should not start a war with a company that is spending $7 billion to renovate FAB42, More than AMDs 2016 revenue. Intel will win it.

I think AMD should focus on delivering a better platform for a slightly better price. Going too low is never a good idea, especially when your main competitor has been enjoying the top alone for many years.

 

[8350rocks]

That is comparing 16C CPU to 18C CPU and ignoring the cost of the rest of the platform.

Yep, the rest of the platform which includes Intel SSDs, Intel "dongles", 2 more cores, and 4 extra threads. I don't know how enterprises would justify paying $300 and more for Intel SSDs just for 2 extra cores which can be paid for if they just have that one extra server.

This is not targeted towards enterprises and when they buy servers the hardware cost is not what matters or costs that much. It is the support that costs and matters.

Besides most servers don't come with every feature they support either. Bought a Dell server? Great. Want to use that iDRAC? Pay up. In the enterprise world there is a lot less complaining on costs and more focusing on what you need to work.

Yes cost does have a role but it is not like a person. People, unless insanely rich, focus on cost as their livelihoods matter based on what they have and spend. In a company if you need RAID5 or you need 2 more cores so you can make one server box do 9 different VM roles than you do it as in the end the cost will pay for itself.

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

Here is the problem. Productivity is usually, for the most part, in a business. The small amount of self employed people who buy their own systems will worry about how much to pay for each part or the overall system. Most companies do not. They will buy whatever is being sold as the best for their program and needs be it Intel, AMD or NVidia etc.

I also see people drooling over 64 PCIe lanes. Considering the piss poor or even non existent support for CFX/SLI in some recent games and the extremely poor scaling past 2 cards, why does it matter to have 64? Now if there were 64 and they could delegate them to other purposes, sure. But most CPU lanes are dedicated to the PCIe slots while the chipset PCIe lanes can be changed per the mobo vendor to make the overall platform more appealing.

But hey I am sure there are plenty of people who want quad SLI GTX 1080s.... Oh wait, only supports dual SLI. Maybe quad CFX Rx 580s? Possibly if the returns are worth it. Still have not ever seen a reason to go past 2 GPUs and even then support has been pretty shoddy, especially in DX12 where devs are more responsible for it.

In the end AMD better be careful. Everyone knows that Intel can do a price war and AMD might not survive it. They are just now getting back to a more competitive position. They should not start a war with a company that is spending $7 billion to renovate FAB42, More than AMDs 2016 revenue. Intel will win it.

I think AMD should focus on delivering a better platform for a slightly better price. Going too low is never a good idea, especially when your main competitor has been enjoying the top alone for many years.

How about 8x NVME m.2 x4 PCIe 3.0 SSDs and 2x GPUs @ 16X/16X?

 

[jimmysmitty]

That is comparing 16C CPU to 18C CPU and ignoring the cost of the rest of the platform.

Yep, the rest of the platform which includes Intel SSDs, Intel "dongles", 2 more cores, and 4 extra threads. I don't know how enterprises would justify paying $300 and more for Intel SSDs just for 2 extra cores which can be paid for if they just have that one extra server.

This is not targeted towards enterprises and when they buy servers the hardware cost is not what matters or costs that much. It is the support that costs and matters.

Besides most servers don't come with every feature they support either. Bought a Dell server? Great. Want to use that iDRAC? Pay up. In the enterprise world there is a lot less complaining on costs and more focusing on what you need to work.

Yes cost does have a role but it is not like a person. People, unless insanely rich, focus on cost as their livelihoods matter based on what they have and spend. In a company if you need RAID5 or you need 2 more cores so you can make one server box do 9 different VM roles than you do it as in the end the cost will pay for itself.

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

Here is the problem. Productivity is usually, for the most part, in a business. The small amount of self employed people who buy their own systems will worry about how much to pay for each part or the overall system. Most companies do not. They will buy whatever is being sold as the best for their program and needs be it Intel, AMD or NVidia etc.

I also see people drooling over 64 PCIe lanes. Considering the piss poor or even non existent support for CFX/SLI in some recent games and the extremely poor scaling past 2 cards, why does it matter to have 64? Now if there were 64 and they could delegate them to other purposes, sure. But most CPU lanes are dedicated to the PCIe slots while the chipset PCIe lanes can be changed per the mobo vendor to make the overall platform more appealing.

But hey I am sure there are plenty of people who want quad SLI GTX 1080s.... Oh wait, only supports dual SLI. Maybe quad CFX Rx 580s? Possibly if the returns are worth it. Still have not ever seen a reason to go past 2 GPUs and even then support has been pretty shoddy, especially in DX12 where devs are more responsible for it.

In the end AMD better be careful. Everyone knows that Intel can do a price war and AMD might not survive it. They are just now getting back to a more competitive position. They should not start a war with a company that is spending $7 billion to renovate FAB42, More than AMDs 2016 revenue. Intel will win it.

I think AMD should focus on delivering a better platform for a slightly better price. Going too low is never a good idea, especially when your main competitor has been enjoying the top alone for many years.

How about 8x NVME m.2 x4 PCIe 3.0 SSDs and 2x GPUs @ 16X/16X?

So we are just spouting off ridiculous specs now? Why not get a setup with quad GV100s?

And what do you do with 8 NVMe M.2 drives? Does AMD even support RAID or would they all have to be individual drives?

For the masses 64PCIe lanes is pointless. For gamers, 64 is still pointless, especially when larger single drives makes more sense and multiple companies are pushing towards NVDIMMs.

The only purpose I see for 64 PCIe lanes is in the HPC market where they already have specialized boards to run as many GPGPU accelerators as possible.

 

[8350rocks]

That is comparing 16C CPU to 18C CPU and ignoring the cost of the rest of the platform.

Yep, the rest of the platform which includes Intel SSDs, Intel "dongles", 2 more cores, and 4 extra threads. I don't know how enterprises would justify paying $300 and more for Intel SSDs just for 2 extra cores which can be paid for if they just have that one extra server.

This is not targeted towards enterprises and when they buy servers the hardware cost is not what matters or costs that much. It is the support that costs and matters.

Besides most servers don't come with every feature they support either. Bought a Dell server? Great. Want to use that iDRAC? Pay up. In the enterprise world there is a lot less complaining on costs and more focusing on what you need to work.

Yes cost does have a role but it is not like a person. People, unless insanely rich, focus on cost as their livelihoods matter based on what they have and spend. In a company if you need RAID5 or you need 2 more cores so you can make one server box do 9 different VM roles than you do it as in the end the cost will pay for itself.

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

Here is the problem. Productivity is usually, for the most part, in a business. The small amount of self employed people who buy their own systems will worry about how much to pay for each part or the overall system. Most companies do not. They will buy whatever is being sold as the best for their program and needs be it Intel, AMD or NVidia etc.

I also see people drooling over 64 PCIe lanes. Considering the piss poor or even non existent support for CFX/SLI in some recent games and the extremely poor scaling past 2 cards, why does it matter to have 64? Now if there were 64 and they could delegate them to other purposes, sure. But most CPU lanes are dedicated to the PCIe slots while the chipset PCIe lanes can be changed per the mobo vendor to make the overall platform more appealing.

But hey I am sure there are plenty of people who want quad SLI GTX 1080s.... Oh wait, only supports dual SLI. Maybe quad CFX Rx 580s? Possibly if the returns are worth it. Still have not ever seen a reason to go past 2 GPUs and even then support has been pretty shoddy, especially in DX12 where devs are more responsible for it.

In the end AMD better be careful. Everyone knows that Intel can do a price war and AMD might not survive it. They are just now getting back to a more competitive position. They should not start a war with a company that is spending $7 billion to renovate FAB42, More than AMDs 2016 revenue. Intel will win it.

I think AMD should focus on delivering a better platform for a slightly better price. Going too low is never a good idea, especially when your main competitor has been enjoying the top alone for many years.

How about 8x NVME m.2 x4 PCIe 3.0 SSDs and 2x GPUs @ 16X/16X?

So we are just spouting off ridiculous specs now? Why not get a setup with quad GV100s?

And what do you do with 8 NVMe M.2 drives? Does AMD even support RAID or would they all have to be individual drives?

For the masses 64PCIe lanes is pointless. For gamers, 64 is still pointless, especially when larger single drives makes more sense and multiple companies are pushing towards NVDIMMs.

The only purpose I see for 64 PCIe lanes is in the HPC market where they already have specialized boards to run as many GPGPU accelerators as possible.

There *are* PCIe RAID controllers out there...broadcom has one for sure, for example.

I know they support RAID natively, just not sure if that applies to NVMe directly.

 

[juanrga]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

 

[juanrga]

 

[gamerk316]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

Adding cores is just a performance multiplayer. Doubling single core performance offers the same performance benefit of doubling the amount of cores. However, with less then perfect scaling, the first solution can offer more performance.

Taking this case, AMD may have 60% cores, but Intel has 40% per core performance, so the theoretical maximum performance advantage for AMD, assuming perfect scaling and disregarding AVX512, is 20%. And if you get less then perfect scaling, the performance advantage is less.

Those 10% IPC improvements per generation add up after a few generations.

 

[goldstone77]

PathForward: US Dept. of Energy Awards $258M in Research Contracts To Develop Exascale Supercomputer Technology
by Ryan Smith on June 15, 2017 2:06 PM EST

http://www.anandtech.com/show/11547/us-dept-of-energy-awards-258m-to-develop-exascale-supercomputer-technology

 

[juanrga]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

Adding cores is just a performance multiplayer. Doubling single core performance offers the same performance benefit of doubling the amount of cores. However, with less then perfect scaling, the first solution can offer more performance.

Taking this case, AMD may have 60% cores, but Intel has 40% per core performance, so the theoretical maximum performance advantage for AMD, assuming perfect scaling and disregarding AVX512, is 20%. And if you get less then perfect scaling, the performance advantage is less.

Right. Those 6 extra cores will surely produce only a performance gain of 20--40% due to non-linear scaling of performance with number of cores (Amdahl law).

 

[8350rocks]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

Adding cores is just a performance multiplayer. Doubling single core performance offers the same performance benefit of doubling the amount of cores. However, with less then perfect scaling, the first solution can offer more performance.

Taking this case, AMD may have 60% cores, but Intel has 40% per core performance, so the theoretical maximum performance advantage for AMD, assuming perfect scaling and disregarding AVX512, is 20%. And if you get less then perfect scaling, the performance advantage is less.

Right. Those 6 extra cores will surely produce only a performance gain of 20--40% due to non-linear scaling of performance with number of cores (Amdahl law).

Considering it is effectively 12 cores with SMT...you are probably going to see much closer to 60-70% gain.

Besides...IPC gap is negligible, especially in the high end parts, and I very seriously doubt you see 5 GHz on any 10+ core part reliably. Not to mention AVX512 is not in any programs that I am aware of at the moment, especially professional workstation tasks. Maybe some scientific HPC stuff that is custom coded, but you will not see that in Joe Q Public's normal workflow.

 

[8350rocks]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

Adding cores is just a performance multiplayer. Doubling single core performance offers the same performance benefit of doubling the amount of cores. However, with less then perfect scaling, the first solution can offer more performance.

Taking this case, AMD may have 60% cores, but Intel has 40% per core performance, so the theoretical maximum performance advantage for AMD, assuming perfect scaling and disregarding AVX512, is 20%. And if you get less then perfect scaling, the performance advantage is less.

Those 10% IPC improvements per generation add up after a few generations.

The gap between haswell and skylake per clock is 10%...the gap between haswell and kabylake per clock is also 10%.

My, 4 generations later and we got 10%? I will take those "generational improvements" adding up odds...

 

[-Fran-]

Amdahl's law doesn't apply to server workloads. Stop quoting that out of context. That law states explicitly that is for single algorithms and not wide-range applications that are composed of several *million* algorithms. Simple counter-example: Virtualization.

And yes, AMD does have a good CPU with EPYC; I can say confidently it is the case, because here there's talk about getting a couple racks with them since they'll be great for multi-purpose servers.

Cheers!

 

[juanrga]

That 10 core i9 7900K clocked @5ghz is going to be a monster.

20% more clockspeed will not offset 60% more cores in workloads that a productivity class platform will often utilize though.

20% more clockspeed plus 20--25% more IPC plus AVX512 support will do, even in workloads that scale up to 32-threads.

Adding cores is just a performance multiplayer. Doubling single core performance offers the same performance benefit of doubling the amount of cores. However, with less then perfect scaling, the first solution can offer more performance.

Taking this case, AMD may have 60% cores, but Intel has 40% per core performance, so the theoretical maximum performance advantage for AMD, assuming perfect scaling and disregarding AVX512, is 20%. And if you get less then perfect scaling, the performance advantage is less.

Right. Those 6 extra cores will surely produce only a performance gain of 20--40% due to non-linear scaling of performance with number of cores (Amdahl law).

Considering it is effectively 12 cores with SMT...you are probably going to see much closer to 60-70% gain.

Besides...IPC gap is negligible, especially in the high end parts, and I very seriously doubt you see 5 GHz on any 10+ core part reliably. Not to mention AVX512 is not in any programs that I am aware of at the moment, especially professional workstation tasks. Maybe some scientific HPC stuff that is custom coded, but you will not see that in Joe Q Public's normal workflow.

16 cores cannot provide a 20--40% gain over 10-core whereas 12 cores provide a 60--70% gain over 10-core. It is simply impossible.

The current IPC gap is not negligible and that is why increasing the IPC is on engineers' TO-DO list for Zen2.

The mentioned 20% clock advantage was already considering that 5GHz is not the norm.

AVX512 is an standard on HPC. Any time you heard some computer is using "Xeon Phi", it is running AVX512 code.

https://www.top500.org/lists/2016/11/highlights/

The big change is that now Intel also adds AVX512 to server and desktop CPUs. Therefore we will see server/workstation software using AVX512 soon. Google is already using AVX512 in its servers

https://www.overclock3d.net/news/cpu_mainboard/google_confirms_that_they_are_using_skylake_xeon_cpus_which_support_avx-512/1

And Sandra is already preparing for AVX512

http://www.sisoftware.eu/2016/02/24/future-performance-with-avx512-in-sandra-2016-sp1/

Amdahl's law doesn't apply to server workloads.

It is a general law of computing and yes it applies to server workloads as well.