Discussion Intel current and future Lakes & Rapids thread

[DrMrLordX]

38c, 270w. 10nm + meh why didn't they manufacture in 10nm ++ ?

We don't know exactly which iteration of the 10nm process is used for Ice Lake-SP. There's very little clarity as to what 10nm+ and 10nm++ really are anyway, since sometimes Intel chooses to acknowledge the old 10nm process used for Cannonlake and sometimes they don't. We've gone over that earlier in the thread.

Yeah that slide wasn’t meant for the public, nor was it directly from Intel but ASUS . I don’t doubt that it’s upto 38c but intel still hasn’t actually announced those specs

True. It's all we've got to go on thus far. If Intel can get that TDP figure down some, then great, but . . .

 

[jpiniero]

If anything, I would expect the TDP to go up, given what's been rumored with Cooper Lake.

 

[mikk]

If anything, I would expect the TDP to go up, given what's been rumored with Cooper Lake.

If you expect it goes up the chance it goes down is actually decent because you are usually wrong with your expectations.

 

[eek2121]

This.

AMD set their expectances not because they were pessimistic, but because they knew exactly, how many wafers they could reserve 6-9 months before they released the forecast for Q4 '19. I think they forecast in a way that calculates with every single piece of product they can produce as being sold, and give that number a - 0,5-1%, just to be sure.
That's also one of the reasons their stock is very high too - they are in a very good position regarding product strength, but still not setting delusional goals. I'd say AMD's stock price is more grounded and healthier than Intel's, despite the huge difference between their current and actual financial horsepower.

While this is not an AMD thread, nor is it a thread about financials, etc. (and therefore I'm not going to respond any further about this subject), it is important to realize that AMD is seeing decent growth in their GPU division thanks to Navi. Ryzen, Epyc, and Threadripper all continue to grow as well. We know Navi is doing well because there have been a few distributors as well as retailers that have published some numbers. Keep that in mind before you say AMD is overvalued. I expect AMD to perform rather well this year.

That being said, Intel really needs to get some desktop parts out the door on 10nm. If they don't do it this year, I fear they will lose the ability to compete. This will be bad for all users, regardless of if you are an Intel person or an AMD person.

 

[lobz]

That being said, Intel really needs to get some desktop parts out the door on 10nm. If they don't do it this year, I fear they will lose the ability to compete. This will be bad for all users, regardless of if you are an Intel person or an AMD person.

In that regard I have bad news for you, and this also brings us back to topic: Intel is not releasing any meaningful 10nm desktop parts in the next 5-6 quarters.

 

[DrMrLordX]

I fear they will lose the ability to compete. This will be bad for all users, regardless of if you are an Intel person or an AMD person.

Look at the bright side: Intel is still making a lot of money.

 

[beginner99]

So how long can Intel still brainwash IT managers into buying Xeons ?????? Over Rome/Milan

It's not brainwashing it's simply clueless people being clueless. The look at what they have already and simply buy the same thing again (or it's never version hence why intel named their xeons v1, v2 etc for some time to appease them). And while they are doing it because they are clueless and don't really care about performance or power use there is at least a small case to make about compatibility. But it's a small case and these people are thinking in software terms as in changing it will have huge costs so we just keep what we have.

 

[Richie Rich]

I still don't see how Intel will beat the Rome chips any time soon. Anybody with a brain would never buy any of these Intel server chips.

There are three main reasons:

1. Customer paid a lot of money for AVX512 optimizations. AMD has nothing to offer to them right now.
2. Intel Golden Cove is coming soon and they know what IPC jump it will bring. Long term partnership doesn't break just for one problematic year. Zen2 is great but AMD would need at least 3 years of continuous dominance to convince Intel customers they are better in long term (just like they did with K8). But if Intel shows Golden Cove preliminary performance numbers this might be explanation we see no massive exodus.
3. Rome/Zen2 EPYC is great due to chiplet design. Intel can do the same with even better EMIB bridges. AMD has clear at least one year advantage in this with Zen2. But this is just short term advantage. Golde Cove with EMIB or 3D stacking can turn the long term advantage back to Intel's side. It's very hard to abandon long term partnership with Intel when Intel's salesman shows you really cool stuff's coming soon. And if Golden Cove is really good you don't want to switch and get stuck with worse AMD Zen3/4. And that's why no price discount for crapy actuall CPUs (this can explain great Intel's revenue). IMHO Intel tries to sale future in these days.

 

[uzzi38]

There are three main reasons:

1. Customer paid a lot of money for AVX512 optimizations. AMD has nothing to offer to them right now.
2. Intel Golden Cove is coming soon and they know what IPC jump it will bring. Long term partnership doesn't break just for one problematic year. Zen2 is great but AMD would need at least 3 years of continuous dominance to convince Intel customers they are better in long term (just like they did with K8). But if Intel shows Golden Cove preliminary performance numbers this might be explanation we see no massive exodus.
3. Rome/Zen2 EPYC is great due to chiplet design. Intel can do the same with even better EMIB bridges. AMD has clear at least one year advantage in this with Zen2. But this is just short term advantage. Golde Cove with EMIB or 3D stacking can turn the long term advantage back to Intel's side. It's very hard to abandon long term partnership with Intel when Intel's salesman shows you really cool stuff's coming soon. And if Golden Cove is really good you don't want to switch and get stuck with worse AMD Zen3/4. And that's why no price discount for crapy actuall CPUs (this can explain great Intel's revenue). IMHO Intel tries to sale future in these days.

1. AVX512 is extremely niche, and will remain that way for a long time.
2. Golden Cove is a long way out for servers. We're talking Granite Rapids, which is late 2022 earliest (as it's 7nm). For desktop and mobile, it's relatively soon, but for servers? Yeah, 3 years of remaining on top is almost guaranteed. Sapphire Rapids is almost confirmed to be Willow Cove. Well, this is about as close to a confirmation as one can get bar actual word from Intel.
3. See above, plus, what makes you think AMD doesn't have 3D stacking coming relatively soon-ish, or that Intel will be utilising Foveros at all in the near future on the server platform.

Granted, it's a TSMC tech that AMD may or may not use, the option is still there for AMD to utilise it. Whether they will or they won't is yet to be seen.

 

[Markfw]

There are three main reasons:

1. Customer paid a lot of money for AVX512 optimizations. AMD has nothing to offer to them right now.
2. Intel Golden Cove is coming soon and they know what IPC jump it will bring. Long term partnership doesn't break just for one problematic year. Zen2 is great but AMD would need at least 3 years of continuous dominance to convince Intel customers they are better in long term (just like they did with K8). But if Intel shows Golden Cove preliminary performance numbers this might be explanation we see no massive exodus.
3. Rome/Zen2 EPYC is great due to chiplet design. Intel can do the same with even better EMIB bridges. AMD has clear at least one year advantage in this with Zen2. But this is just short term advantage. Golde Cove with EMIB or 3D stacking can turn the long term advantage back to Intel's side. It's very hard to abandon long term partnership with Intel when Intel's salesman shows you really cool stuff's coming soon. And if Golden Cove is really good you don't want to switch and get stuck with worse AMD Zen3/4. And that's why no price discount for crapy actuall CPUs (this can explain great Intel's revenue). IMHO Intel tries to sale future in these days.

In the benchmark thread I saw, they equalled 512 performance, since everything else was 2x faster.
See above for number 2 and number 3. Also, MIlan will be out by the time Intel has Golden cove. is out, and its faster than Rome.

 

[Intel Inside]

1. AVX512 is extremely niche, and will remain that way for a long time.
2. Golden Cove is a long way out for servers. We're talking Granite Rapids, which is late 2022 earliest (as it's 7nm). For desktop and mobile, it's relatively soon, but for servers? Yeah, 3 years of remaining on top is almost guaranteed. Sapphire Rapids is almost confirmed to be Willow Cove. Well, this is about as close to a confirmation as one can get bar actual word from Intel.
3. See above, plus, what makes you think AMD doesn't have 3D stacking coming relatively soon-ish, or that Intel will be utilising Foveros at all in the near future on the server platform.

Granted, it's a TSMC tech that AMD may or may not use, the option is still there for AMD to utilise it. Whether they will or they won't is yet to be seen.

Reply removed

Please use English. It eliminates the problems that can arise from posts using another language.

Admin allisolm

 

[DrMrLordX]

@Richie Rich

To address point #3:

Intel's fundamental problem has nothing to do with EMIB or Foveros, per se. They are simply unable to manufacture 28c+ monolithic dice from anything beyond a Skylake variant at this time. The problem has been known since at least 2017, if not earlier, so they've had plenty of time to say, "Okay, we're just going to keep manufacturing on 10nm and use EMIB to reach core counts that make our customers happy". They didn't do it. Skylake-SP still isn't going to be an EMIB chip. You have to wait for Sapphire Rapids to (maybe) get that. Intel is losing benchmarks thanks to them simply not selling anything other than Skylake-SP retreads, over and over again, with very little growth in core count per socket (and insanely high power usage/TCO on Cascade Lake-AP and Cooper Lake-SP).

Why they refused to utilize their most advanced packaging technology available to keep their latest cores in server products is beyond me. At speeds of 3 GHz and below, a bunch of EMIBed IceLake cores should have been at least okay-ish, and theoretically we could have seen that product shipping in 2018. Unlike with consumer devices, cost wouldn't have been a major factor either, since the margin on such CPUs is so high that Intel can afford to pay extra in the manufacturing/packaging process.

As a consequence, Intel is shipping Skylake-SP in unknown quantity in 2020 to face off against a competitor with a uarch that has undergone two years more advancement than anticipated by anyone. Skylake-SP was supposed to be a competitor to Zen1, not Zen3.

 

[itsmydamnation]

@Richie Rich


Why they refused to utilize their most advanced packaging technology available to keep their latest cores in server products is beyond me. At speeds of 3 GHz and below, a bunch of EMIBed IceLake cores should have been at least okay-ish, and theoretically we could have seen that product shipping in 2018. Unlike with consumer devices, cost wouldn't have been a major factor either, since the margin on such CPUs is so high that Intel can afford to pay extra in the manufacturing/packaging process.

As a consequence, Intel is shipping Skylake-SP in unknown quantity in 2020 to face off against a competitor with a uarch that has undergone two years more advancement than anticipated by anyone. Skylake-SP was supposed to be a competitor to Zen1, not Zen3.

That ones easy, cache/ i/o design. They would need a completely new one to take advantage of using many chiplets as there current one would be extremely complex to package at high core counts. That cache design probably would take several years of development. So im sure they are working on it now, but just like everything else for intel at the moment its probably a 2022 product for servers.

 

[DrMrLordX]

That ones easy, cache/ i/o design. They would need a completely new one to take advantage of using many chiplets as there current one would be extremely complex to package at high core counts. That cache design probably would take several years of development. So im sure they are working on it now, but just like everything else for intel at the moment its probably a 2022 product for servers.

I'm sure it would have taken a cache redesign. Had they started when Cannonlake first began looking funky, they probably could have reached a 2018 launch window (they likely noticed problems with Cannonlake going as far back as 2016). Maybe 2019.

 

[Richie Rich]

In the benchmark thread I saw, they equalled 512 performance, since everything else was 2x faster.
See above for number 2 and number 3. Also, MIlan will be out by the time Intel has Golden cove. is out, and its faster than Rome.

1) AVX512 is not only about vector width. It also contains 17 new instruction subsets. If some customer uses those new instructions AMD's AVX2 doesn't help at all even Zen's FPU is a bit more powerful. Those customers won't switch to AMD until they support those new instructions too. I don't say there is majority of those customers.

2) Sunny and Willow Coves are better than Zen2 in terms of IPC, so right now there is no big pressure (K8 was twice as fast than P4 and AMD gained only 25% server market share after several years). Performance competition is more tight today. Especially when Golden Cove (maybe with EMIB chiplet design) is "something significantly bigger" as Keller said. However tight competition is good for customers and thanks god AMD's back in the game.

@DrMrLordX
I totally agree. Intel has a lot of cool stuff but they are not able to put them together. IMHO it's clear Intel had bad leadership, no vision. Smart people as Keller will put Intel back on track however it will take three more years before we see some results like Core2 uarch was.

 

[Markfw]

1) AVX512 is not only about vector width. It also contains 17 new instruction subsets. If some customer uses those new instructions AMD's AVX2 doesn't help at all even Zen's FPU is a bit more powerful. Those customers won't switch to AMD until they support those new instructions too. I don't say there is majority of those customers.

2) Sunny and Willow Coves are better than Zen2 in terms of IPC, so right now there is no big pressure (K8 was twice as fast than P4 and AMD gained only 25% server market share after several years). Performance competition is more tight today. Especially when Golden Cove (maybe with EMIB chiplet design) is "something significantly bigger" as Keller said. However tight competition is good for customers and thanks god AMD's back in the game.

@DrMrLordX
I totally agree. Intel has a lot of cool stuff but they are not able to put them together. IMHO it's clear Intel had bad leadership, no vision. Smart people as Keller will put Intel back on track however it will take three more years before we see some results like Core2 uarch was.

Its a very small number of customers that use AVX512, and AMD is not a little ahead, they almost tied in avx512 because all the other benchmarks they were ahead at twice the performance. On number 2, did you not read that its Zen3 that will come up against those ? And Zens is supposed to be better than Zen2, the same as Sunny and Willow cove are.

 

[lobz]

There are three main reasons:

1. Customer paid a lot of money for AVX512 optimizations. AMD has nothing to offer to them right now.
2. Intel Golden Cove is coming soon and they know what IPC jump it will bring. Long term partnership doesn't break just for one problematic year. Zen2 is great but AMD would need at least 3 years of continuous dominance to convince Intel customers they are better in long term (just like they did with K8). But if Intel shows Golden Cove preliminary performance numbers this might be explanation we see no massive exodus.
3. Rome/Zen2 EPYC is great due to chiplet design. Intel can do the same with even better EMIB bridges. AMD has clear at least one year advantage in this with Zen2. But this is just short term advantage. Golde Cove with EMIB or 3D stacking can turn the long term advantage back to Intel's side. It's very hard to abandon long term partnership with Intel when Intel's salesman shows you really cool stuff's coming soon. And if Golden Cove is really good you don't want to switch and get stuck with worse AMD Zen3/4. And that's why no price discount for crapy actuall CPUs (this can explain great Intel's revenue). IMHO Intel tries to sale future in these days.

Yeah, no.

Intel Golden Cove is not coming soon. Ice Lake should finally happen first. And then still a few other steps along the way. I'm sure Golden Cove is coming, it's just anything but soon.
Your 3. point is the funniest. If it was that simple as you say it, they would have done that like 2 years ago, when EMIB first ready to use. Long term partnership is not hard to abandon in business - personal greed in exec position, now that is really hard to abandon. All this has literally nothing to do with either Golden Cove or EMIB.

 

[lobz]

1) AVX512 is not only about vector width. It also contains 17 new instruction subsets. If some customer uses those new instructions AMD's AVX2 doesn't help at all even Zen's FPU is a bit more powerful. Those customers won't switch to AMD until they support those new instructions too. I don't say there is majority of those customers.

You can bet that AMD is not after grabbing sales for those 6 customers' 8 AVX 512 projects.

 

[lobz]

2) Sunny and Willow Coves are better than Zen2 in terms of IPC, so right now there is no big pressure (K8 was twice as fast than P4 and AMD gained only 25% server market share after several years).

Do you realize, that 1: no matter how good of an IPC Sunny Cove cores have, when they can't do any meaningful core count increase without burning down entire towns when a server is turned on, and 2: Rome is already twice as fast as Cascade Lake?

I'm just asking.

 

[reb0rn]

Intel for sure is with pants down, but if they have brain left they for sure can respond with atom many cores servers and chiplets design with mixed chip (high low IPC mix etc) but they need to slash and offer a lot more with a lot less margins
Then AMD would be in trouble, will that happen i have no idea... atm AMD is just smart with not so gr8 arhitecture they beat them with brute force

 

[NostaSeronx]

So... the dood says =>
7nm i7/i9/Xeon w/ DDR5-5600
(It will be using the 7nm Multi-die Fabric Interface(Ring for cores(not MDFI), Mesh for chiplets(MDFI)))
late 2020 sampling and mid-2021 shipping

With a "most x86 cores in a single package" ever made in history. Every single die is sub-cm2 including I/O dies(DDR5, CXL, Ethernet, etc), Xeon/server-class McIVR, AI/FPGA accelerators, etc.

 

[uzzi38]

So... the dood says =>
7nm i7/i9/Xeon w/ DDR5-5600
(It will be using the 7nm Multi-die Fabric Interface(Ring for cores(not MDFI), Mesh for chiplets(MDFI)))
late 2020 sampling and mid-2021 shipping

With a "most x86 cores in a single package" ever made in history. Every single die is sub-cm2 including I/O dies(DDR5, CXL, Ethernet, etc), Xeon/server-class McIVR, AI/FPGA accelerators, etc.

You should already know this is bogus as Intel's 7nm isn't ready for production of any Xeon class products in 2021, but instead in 2022.

 

[uzzi38]

1) AVX512 is not only about vector width. It also contains 17 new instruction subsets. If some customer uses those new instructions AMD's AVX2 doesn't help at all even Zen's FPU is a bit more powerful. Those customers won't switch to AMD until they support those new instructions too. I don't say there is majority of those customers.

2) Sunny and Willow Coves are better than Zen2 in terms of IPC, so right now there is no big pressure (K8 was twice as fast than P4 and AMD gained only 25% server market share after several years). Performance competition is more tight today. Especially when Golden Cove (maybe with EMIB chiplet design) is "something significantly bigger" as Keller said. However tight competition is good for customers and thanks god AMD's back in the game.

I totally agree. Intel has a lot of cool stuff but they are not able to put them together. IMHO it's clear Intel had bad leadership, no vision. Smart people as Keller will put Intel back on track however it will take three more years before we see some results like Core2 uarch was.

1. I know that. Doesn't make it any less niche.
2. Neither Sunny nor Willow is relevant in servers, so why does this matter? By the time they are, Zen 3/Zen4 will be in place (respectively), and AMD's server lineup will still have cores that are on-par at worst, as well as nearly twice as many in a lower TDP package than Ice Lake-SP. And if Ice Lake-U has shown us anything, it's that Sunny Cove on 10nm is a very, very minor improvement in power efficiency over heavily refined Skylake on 14nm. What Intel gain in IPC is mostly dropped in clocks whilst power draw remains almost identical.

Granite Rapids is where Intel has a chance again, so you're not wrong with Golden Cove. Just... Golden Cove in servers is 2022, it's competing with Zen 5. We're yet to see how that goes.

 

[lobz]

So... the dood says =>
7nm i7/i9/Xeon w/ DDR5-5600
(It will be using the 7nm Multi-die Fabric Interface(Ring for cores(not MDFI), Mesh for chiplets(MDFI)))
late 2020 sampling and mid-2021 shipping

 

[Richie Rich]

2. Neither Sunny nor Willow is relevant in servers, so why does this matter? By the time they are, Zen 3/Zen4 will be in place (respectively), and AMD's server lineup will still have cores that are on-par at worst, as well as nearly twice as many in a lower TDP package than Ice Lake-SP. And if Ice Lake-U has shown us anything, it's that Sunny Cove on 10nm is a very, very minor improvement in power efficiency over heavily refined Skylake on 14nm. What Intel gain in IPC is mostly dropped in clocks whilst power draw remains almost identical.

Granite Rapids is where Intel has a chance again, so you're not wrong with Golden Cove. Just... Golden Cove in servers is 2022, it's competing with Zen 5. We're yet to see how that goes.

I agree, 2022 is the year where Intel could put everything together and strike back.

AMD has clear advantage in high core count server CPU, no doubt about that. However Intel has huge monolithic 28c cores with shared L3 cache (AMD is behind in development L3 ring bus cache, Zen3 should have that finaly) and this type of CPU is better in certain HPC load (CFD doesn't scale well). There are many applications where 28core at 5GHz is twice as fast than 64c Rome at 2.6 GHz. Nobody cares about 1000W consumption (SW license and engineer hour cost is 1000x higher).

Also 10nm production can be solved faster than everybody think. There is only one suplier for EUV machines: ASML. When Intel buys those machines then problem is solved because Intel struggles with last DUV. And they can move to 7nm EUV (equal to 5nm TSMC) almost in the same time as TSMC to 5nm. And this could be 10nm EUV (equal to 7nm+ TSMC) in 2021 Alder Lake (Golden Cove) and 7nm EUV in 2022. IMHO EUV will solve all Intel's production problems. This means AMD's advantage will be gone.

Anyway, Apple has +83% higher IPC using 6xALU wide core, so who cares about single digits advantage in some slow prehistoric 4xALU core. AMD and Intel should race for first 6xALU core in x86 world and bring those +50% IPC. They should challenge these ultimate Apple cores not each other.