An exciting conference this year, indeed. So many great speakers and topics; I hope coverage is as great as this synopsis, Ian! Are you planning to attend the entirety of the conference?
Having been there for the totally content-free SKL presentation in 2016, I don't really have much confidence that Intel will say anything of substance about Alder Lake.
The only important information that was missing from the last year's Intel presentations and which became visible only after launch, was that even the SuperFin 10-nm Intel process cannot reach the efficiency of the TSMC 7-nm process, in the sense that whenever the CPUs reach their power limits and are constrained by them, at equal power consumption and equal number of active cores, an Intel CPU will have a lower clock frequency than an AMD CPU.
Except for this important detail, the future performance of Tiger Lake and Ice Lake Server could be evaluated from the Hot Chips presentations.
We had a glimpse of the Xe-HP last year on 10esf ... 42 TFlops FP32 ... so looks like Intel 10esf is doing fine. Perhaps this will become clear this year with Sapphire Rapids, which has been sampling broadly since Nov 2020, according to Intel's SC20 slides.
It's true that Zen 3 is more efficient than Willow Cove, but it should be noted that different CPU architectures can have differences in efficiency not related to node. Otherwise, Ryzen 5000 would be 0% faster than Ryzen 3000, since they are built on the same node and have the same power limits. By evaluating Cypress Cove, Comet Lake, Willow Cove, and Zen 3, we can assume that 10SF is not as performant as 14+++, 10SF is less efficient than N7, and Willow Cove is less efficient than Zen 3, meaning that Intel is at a node and architecture deficit. In a way this actually makes sense because Zen was more efficient than Kaby Lake and Coffee Lake (in spite of an inferior node), N7 is roughly 3x as efficient as 14LPP, and Zen 3 is ~1.2x as efficient as Zen 2.
If they don't say anything substantial this year, they're in big trouble. Intel needs to build confidence that they got their version of bigLITTLE ready for prime time, or they will lose a lot of customers and also share value. The latter would bode really poorly for Intel's Chief Revenue Officer, which I was astonished to learn exists just recently.
I'm not so sure Intel's bigLITTLE means much ... I believe one of their leaked configurations are 8 cores of Golden Cove and no Gracemonts, which will be fine for me if it gets the performance lifts from pcie5, ddr5 and the Golden Cove cores.
Judging by Morpheus I, Morpheus II alone might just be worth the attendance fee: that approach is both so ingenious and necessary to avoid the risk of RoP hijacking. I can't wait to see that in real hardware at least in some security enclaves of high-performance SoCs. There is a lot of similar work being done in the ARM universe, but since CET full VM encryption nothing similarly visionary seems on the horizon for x86.
Apple's not a charity and it's not part of academia, so why exactly would they participate if it doesn't serve the business marketing goals. They have their own events where they have the power to completely control the marketing message or directly talk to developers. They don't sell data center products, or ML servers, or HPC servers. If you're a partner who needs to know details about their compute components you're not going to get that information from a public conference.
Do you remember the Apple Samsung lawsuits when the iPhone came out? And how the public reacted? The way the IP system is SUPPOSED to work is that companies reveal information in return for which they get protection for those revelations. Apple's experience in the Samsung case was that the system simply does not work the way it is supposed to. The cost of protecting against was should have been obvious infringement was ungodly and, even more important, regardless of whether they won in court, they were pilloried by the public.
And so what happened was what would have been obvious to anyone capable of engaging in more-than-one-step (ie NOT most of the internet) reasoning... If the system doesn't work for Apple, Apple will disengage from it.
Not completely so -- they do file a number of patents, and you can learn a fair bit (though certainly not everything) from those patents. But why should they make it easier for society to learn from them when society has indicated that it is on the side of clear IP theft rather than on the side of supporting IP?
Apple is the most obvious retreat from an IP commons. But it's not the only one. AMD, Intel, IBM -- all of them say VASTLY less than they used to. Sure, they still publish articles. But the articles contain ever less leading edge content. They tell you the obvious stuff, not the serious stuff. You disagree? Try actually researching something rather than just letting the latest article flow over you.
For example, I recently had reason to question EXACTLY how a modern L1D cache works, in particular how multiple accesses are supported in a single cycle even though the caches are not multiported. Yes yes, multiple banks, we all know that, we're not idiots. But DETAILS... How many distinct accesses per cycle does the TLB support? How many tag lookups? Are piggyback ports used? Is any sort of D2D (storing way information in TLB entries) used? etc etc. I have (somewhat -- still many unresolved issues) reverse-engineered this for the M1. But no thanks to any article from Intel, AMD, or IBM. Intel said that with Haswell they redesigned the L1D to eliminate bank collisions. What did they do? If you know, write up an article -- because there is NOTHING on the internet about this.
Hence my point. EVERYONE is silent, in a way that was not true in the 1990s. IBM, Intel, AMD pretend to be open -- because they need to market themselves, both to new hires and to vendors/enterprise/HPC/national labs, ... Apple doesn't need to do either of these things. So it doesn't. It's not about "take ideas and don't share", it's about economic incentives and the failure of the IP system (largely at the hands of the internet and a jeering crowd who insist that anyone should be allowed to copy anything).
That's not a fact-free assessment but let's not pretend BSD has become a viable desktop platform despite Apple using it as the basis for OS X for years... What gets pushed upstream is mostly low level code Apple want more eyes and bug fixes on. Looks to me like a mostly one-way street as far as taking and giving back; I'm sure Linux was overlooked because the licence would have forced Apple to give back. As for Apple vs. Samsung, they definitely should have kept those rounded corners a trade secret...
"I'm sure Linux was overlooked because the licence would have forced Apple to give back."
Seriously, dude? You do know that there is a HISTORY here? A history that runs through CMU and Mach, and that begins before Linux even existed, that begins, hell, almost before the 80386 even existed?
The Apple Samsung cases regarding the 3gs and galaxy s had almost nothing to do with technical patents, it was almost all about design, the shape of the phone and icon images.
It's unclear how you can draw the conclusions you do based on that.
However I agree that the US patent office is a failure in regards to that it's possible to get a patent issued even though it's clear that prior art exists, which goes completely against what the patent system was set up to do. It has been hijacked by lawyers esp in the US but it's spreading unfortunately and to a large degreefrom lobbying by transnationals.
My statement was about the failure of the IP system as a whole. Samsung was the extremely visible case. But on the more technical side was Apple vs WARF, where, once again, the Internet leaped in to insist that Apple must be guilty of infringement while knowing almost nothing about the case, either the nature of WARF's claim, or why Apple felt that it did not infringe -- even though the relevant Apple patents date from 2012 and can be seen to be somewhat different because the load-store aliasing predictor also stores register information. (This register information can be used in a way completely different from anything Moshovos imagined, to accelerate traditional loads, sometimes by shaving off a cycle, sometimes down to simply a Rename operation.)
You can nitpick irrelevant details. Or you can consider the point I made about the increasing vapidity of talks and papers by every CPU company. Your choice.
While I think WARF is a patent troll (produces no products), whether Moshovos imagined the specific use of it or not is irrelevant. They patented store-sets using PC hashing as a way of disambiguating them (I believe). If you use that mechanism, whatever additional metadata you store to get goodness on top of it becomes irrelevant. You're using the mechanism, pay up.
I think the IBM paper I linked is a pretty solid argument against vapidity. You're probably looking in the wrong places. Perhaps such papers exist from apple and I'm not aware of them? (unlikely, but there's a non-zero chance of everything)
Have you ever seen a research paper authored by apple employees in any conference? Never happens. Attend ISCA etc and you'll find several apple employees, but I've never seen one presentation to date from them. PhD students once hired by apple are not permitted to present research work done during their PhD under the pretext of divulging apple's secrets. You can google Yan Le Cun's critique of apple and its lack of success in AI (at the time). His critique was that apple is great at taking other people's ideas and mature technologies and polishing/packaging it to an unprecedented level, but doesn't come up with original ideas. I think this is still true.
Patents don't count as publications, btw. They're designed to be as broad and obfuscating as possible.
There's levels of openness. There's apple, which likely implements TAGE, and tons of other algorithms invented in academia but gives nothing in return. Intel research does publish good ideas, many of which become products years down the line. IBM papers have far more detail than what you might expect out of a company. Take this as an example: https://conferences.computer.org/isca/pdfs/ISCA202... but there are several. I recall an AMD paper detailing their uop-cache organization. Apple won't even tell you their ROB size.
Intel, AMD, and IBM reveal at least enough information that someone writing hand-tuned software knows about bottlenecks and how to get around them. Intel's software optimization guide is full of such info. Haven't looked at AMD's much. Your particular interest might not be covered, but most reasonable cases (scheduler organization, branches per cacheline/16b, loop streaming buffer sizes etc) are publicly stated. Do we know apple's rob size from Apple itself?
Samsung vs. apple is a red herring argument. That was about a UI "look and feel" not hardware. Apple and Intel had to pay WARF (Univ of Wisconsin)for their memory disambiguation patent infringement, a sign that hard IP is protected. Even if an array of rounded-cornered squares is not patentable.
(As you would have learned if you had read the Apple patents. But of course you don't do that, do you? Complain about how Apple never tells the world anything -- but don't bother to consult one fairly rich source of ground truth.)
This is my point. The internet has created a pool of very loud people Dunning-Kruger's who don't even know how little they know -- but are very certain about how and why they are right and everyone else is wrong.
And this supports the claim that apple does not add knowledge/value to the larger cpu design community? Exactly how does that logic follow? But sure, classic internet troll behavior of fixating on one inconsequential point to conveniently shift the conversation to how right you were. Show me papers published by apple in hardware. Start with One. I linked an ibm paper, told you where to find others. But no, your response is that I haven’t kept up with some courtroom relitigation. Congrats! And kudos on knowing about denning kruger.
As to patents, a) patents make broad claims for winning lawsuits. Implementation may or may not be exactly as specified in them and b) your company lawyer will advise against reading patents as it may be used against you
Are these PIM solutions (and the flash equivalent, where you do the processing in the flash drive) actually being used? Obviously they're not being used by Apple or in your basic Wintel laptop, but are they being used by cloud providers, or HPC or universities or anywhere? For 20 years this ideas has been pushed academically, and for maybe ten years I've been seeing "product-like" announcements. But never any reference to these things in the wild.
I am well aware that I don't know the full range of computing. Hence this question. Not snark, just a genuine curiosity as to whether/when/if these things will exit the lab.
Should be an exciting conference! One comment about the sponsorship levels, which currently top out at "Rhodium"; they should have added an "Unobtainium" level above that, would be very a propos the current shortages and prices.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
32 Comments
Back to Article
LordSojar - Tuesday, May 18, 2021 - link
An exciting conference this year, indeed. So many great speakers and topics; I hope coverage is as great as this synopsis, Ian! Are you planning to attend the entirety of the conference?Andrei Frumusanu - Tuesday, May 18, 2021 - link
We'll be covering the whole show, yes.LordSojar - Tuesday, May 18, 2021 - link
Excellent! Thanks Andrei, Ian, the rest of and team Anandtech for doing so. Looking forward to seeing more about rDNA2 and Bluefield-2, especially!SarahKerrigan - Tuesday, May 18, 2021 - link
Having been there for the totally content-free SKL presentation in 2016, I don't really have much confidence that Intel will say anything of substance about Alder Lake.AdrianBc - Tuesday, May 18, 2021 - link
Last year's presentations about Tiger Lake and Ice Lake Server were informative enough.Obviously they were excessively optimistic, but if you looked at them critically it was possible to assess how the real products will be.
AdrianBc - Tuesday, May 18, 2021 - link
The only important information that was missing from the last year's Intel presentations and which became visible only after launch, was that even the SuperFin 10-nm Intel process cannot reach the efficiency of the TSMC 7-nm process, in the sense that whenever the CPUs reach their power limits and are constrained by them, at equal power consumption and equal number of active cores, an Intel CPU will have a lower clock frequency than an AMD CPU.Except for this important detail, the future performance of Tiger Lake and Ice Lake Server could be evaluated from the Hot Chips presentations.
Targon - Tuesday, May 18, 2021 - link
Intel: We have the fastest processors, but they need 350W for the CPU alone to even have a chance to compete with AMD.JayNor - Thursday, May 20, 2021 - link
We had a glimpse of the Xe-HP last year on 10esf ... 42 TFlops FP32 ... so looks like Intel 10esf is doing fine. Perhaps this will become clear this year with Sapphire Rapids, which has been sampling broadly since Nov 2020, according to Intel's SC20 slides.Otritus - Saturday, May 22, 2021 - link
It's true that Zen 3 is more efficient than Willow Cove, but it should be noted that different CPU architectures can have differences in efficiency not related to node. Otherwise, Ryzen 5000 would be 0% faster than Ryzen 3000, since they are built on the same node and have the same power limits. By evaluating Cypress Cove, Comet Lake, Willow Cove, and Zen 3, we can assume that 10SF is not as performant as 14+++, 10SF is less efficient than N7, and Willow Cove is less efficient than Zen 3, meaning that Intel is at a node and architecture deficit. In a way this actually makes sense because Zen was more efficient than Kaby Lake and Coffee Lake (in spite of an inferior node), N7 is roughly 3x as efficient as 14LPP, and Zen 3 is ~1.2x as efficient as Zen 2.eastcoast_pete - Tuesday, May 18, 2021 - link
If they don't say anything substantial this year, they're in big trouble. Intel needs to build confidence that they got their version of bigLITTLE ready for prime time, or they will lose a lot of customers and also share value. The latter would bode really poorly for Intel's Chief Revenue Officer, which I was astonished to learn exists just recently.JayNor - Thursday, May 20, 2021 - link
I'm not so sure Intel's bigLITTLE means much ... I believe one of their leaked configurations are 8 cores of Golden Cove and no Gracemonts, which will be fine for me if it gets the performance lifts from pcie5, ddr5 and the Golden Cove cores.abufrejoval - Tuesday, May 18, 2021 - link
Judging by Morpheus I, Morpheus II alone might just be worth the attendance fee: that approach is both so ingenious and necessary to avoid the risk of RoP hijacking. I can't wait to see that in real hardware at least in some security enclaves of high-performance SoCs. There is a lot of similar work being done in the ARM universe, but since CET full VM encryption nothing similarly visionary seems on the horizon for x86.abufrejoval - Tuesday, May 18, 2021 - link
Correction: Aquabolt seems just as interesting, it's been a long PIM winter over the last couple of years.artk2219 - Tuesday, May 18, 2021 - link
Ah typical Apple, always happy take everyone else's ideas, and share nothing in return.flgt - Tuesday, May 18, 2021 - link
Apple's not a charity and it's not part of academia, so why exactly would they participate if it doesn't serve the business marketing goals. They have their own events where they have the power to completely control the marketing message or directly talk to developers. They don't sell data center products, or ML servers, or HPC servers. If you're a partner who needs to know details about their compute components you're not going to get that information from a public conference.name99 - Tuesday, May 18, 2021 - link
That's a misunderstanding of Apple's behavior.Do you remember the Apple Samsung lawsuits when the iPhone came out? And how the public reacted?
The way the IP system is SUPPOSED to work is that companies reveal information in return for which they get protection for those revelations. Apple's experience in the Samsung case was that the system simply does not work the way it is supposed to. The cost of protecting against was should have been obvious infringement was ungodly and, even more important, regardless of whether they won in court, they were pilloried by the public.
And so what happened was what would have been obvious to anyone capable of engaging in more-than-one-step (ie NOT most of the internet) reasoning... If the system doesn't work for Apple, Apple will disengage from it.
Not completely so -- they do file a number of patents, and you can learn a fair bit (though certainly not everything) from those patents. But why should they make it easier for society to learn from them when society has indicated that it is on the side of clear IP theft rather than on the side of supporting IP?
Apple is the most obvious retreat from an IP commons. But it's not the only one. AMD, Intel, IBM -- all of them say VASTLY less than they used to. Sure, they still publish articles. But the articles contain ever less leading edge content. They tell you the obvious stuff, not the serious stuff.
You disagree? Try actually researching something rather than just letting the latest article flow over you.
For example, I recently had reason to question EXACTLY how a modern L1D cache works, in particular how multiple accesses are supported in a single cycle even though the caches are not multiported. Yes yes, multiple banks, we all know that, we're not idiots. But DETAILS...
How many distinct accesses per cycle does the TLB support? How many tag lookups? Are piggyback ports used? Is any sort of D2D (storing way information in TLB entries) used? etc etc.
I have (somewhat -- still many unresolved issues) reverse-engineered this for the M1. But no thanks to any article from Intel, AMD, or IBM.
Intel said that with Haswell they redesigned the L1D to eliminate bank collisions. What did they do? If you know, write up an article -- because there is NOTHING on the internet about this.
Hence my point.
EVERYONE is silent, in a way that was not true in the 1990s.
IBM, Intel, AMD pretend to be open -- because they need to market themselves, both to new hires and to vendors/enterprise/HPC/national labs, ...
Apple doesn't need to do either of these things. So it doesn't.
It's not about "take ideas and don't share", it's about economic incentives and the failure of the IP system (largely at the hands of the internet and a jeering crowd who insist that anyone should be allowed to copy anything).
BushLin - Tuesday, May 18, 2021 - link
That's not a fact-free assessment but let's not pretend BSD has become a viable desktop platform despite Apple using it as the basis for OS X for years... What gets pushed upstream is mostly low level code Apple want more eyes and bug fixes on. Looks to me like a mostly one-way street as far as taking and giving back; I'm sure Linux was overlooked because the licence would have forced Apple to give back.As for Apple vs. Samsung, they definitely should have kept those rounded corners a trade secret...
name99 - Tuesday, May 18, 2021 - link
"I'm sure Linux was overlooked because the licence would have forced Apple to give back."Seriously, dude? You do know that there is a HISTORY here? A history that runs through CMU and Mach, and that begins before Linux even existed, that begins, hell, almost before the 80386 even existed?
https://en.wikipedia.org/wiki/Mach_(kernel)
Zoolook - Tuesday, May 18, 2021 - link
The Apple Samsung cases regarding the 3gs and galaxy s had almost nothing to do with technical patents, it was almost all about design, the shape of the phone and icon images.It's unclear how you can draw the conclusions you do based on that.
However I agree that the US patent office is a failure in regards to that it's possible to get a patent issued even though it's clear that prior art exists, which goes completely against what the patent system was set up to do. It has been hijacked by lawyers esp in the US but it's spreading unfortunately and to a large degreefrom lobbying by transnationals.
name99 - Tuesday, May 18, 2021 - link
My statement was about the failure of the IP system as a whole.Samsung was the extremely visible case. But on the more technical side was Apple vs WARF, where, once again, the Internet leaped in to insist that Apple must be guilty of infringement while knowing almost nothing about the case, either the nature of WARF's claim, or why Apple felt that it did not infringe -- even though the relevant Apple patents date from 2012 and can be seen to be somewhat different because the load-store aliasing predictor also stores register information. (This register information can be used in a way completely different from anything Moshovos imagined, to accelerate traditional loads, sometimes by shaving off a cycle, sometimes down to simply a Rename operation.)
You can nitpick irrelevant details.
Or you can consider the point I made about the increasing vapidity of talks and papers by every CPU company.
Your choice.
deltaFx2 - Wednesday, May 19, 2021 - link
While I think WARF is a patent troll (produces no products), whether Moshovos imagined the specific use of it or not is irrelevant. They patented store-sets using PC hashing as a way of disambiguating them (I believe). If you use that mechanism, whatever additional metadata you store to get goodness on top of it becomes irrelevant. You're using the mechanism, pay up.I think the IBM paper I linked is a pretty solid argument against vapidity. You're probably looking in the wrong places. Perhaps such papers exist from apple and I'm not aware of them? (unlikely, but there's a non-zero chance of everything)
SIDtech - Wednesday, May 19, 2021 - link
I remember you from Realworldtech forums. Haven't seen you post there in a whiledeltaFx2 - Wednesday, May 19, 2021 - link
Have you ever seen a research paper authored by apple employees in any conference? Never happens. Attend ISCA etc and you'll find several apple employees, but I've never seen one presentation to date from them. PhD students once hired by apple are not permitted to present research work done during their PhD under the pretext of divulging apple's secrets. You can google Yan Le Cun's critique of apple and its lack of success in AI (at the time). His critique was that apple is great at taking other people's ideas and mature technologies and polishing/packaging it to an unprecedented level, but doesn't come up with original ideas. I think this is still true.Patents don't count as publications, btw. They're designed to be as broad and obfuscating as possible.
There's levels of openness. There's apple, which likely implements TAGE, and tons of other algorithms invented in academia but gives nothing in return. Intel research does publish good ideas, many of which become products years down the line. IBM papers have far more detail than what you might expect out of a company. Take this as an example: https://conferences.computer.org/isca/pdfs/ISCA202... but there are several. I recall an AMD paper detailing their uop-cache organization. Apple won't even tell you their ROB size.
Intel, AMD, and IBM reveal at least enough information that someone writing hand-tuned software knows about bottlenecks and how to get around them. Intel's software optimization guide is full of such info. Haven't looked at AMD's much. Your particular interest might not be covered, but most reasonable cases (scheduler organization, branches per cacheline/16b, loop streaming buffer sizes etc) are publicly stated. Do we know apple's rob size from Apple itself?
Samsung vs. apple is a red herring argument. That was about a UI "look and feel" not hardware. Apple and Intel had to pay WARF (Univ of Wisconsin)for their memory disambiguation patent infringement, a sign that hard IP is protected. Even if an array of rounded-cornered squares is not patentable.
name99 - Wednesday, May 19, 2021 - link
You do know that Apple DIDN'T pay WARF, right? They were found non-infringing.https://www.reuters.com/article/us-usa-court-apple...
(As you would have learned if you had read the Apple patents. But of course you don't do that, do you? Complain about how Apple never tells the world anything -- but don't bother to consult one fairly rich source of ground truth.)
This is my point. The internet has created a pool of very loud people Dunning-Kruger's who don't even know how little they know -- but are very certain about how and why they are right and everyone else is wrong.
deltaFx2 - Wednesday, May 19, 2021 - link
And this supports the claim that apple does not add knowledge/value to the larger cpu design community? Exactly how does that logic follow?But sure, classic internet troll behavior of fixating on one inconsequential point to conveniently shift the conversation to how right you were. Show me papers published by apple in hardware. Start with One. I linked an ibm paper, told you where to find others. But no, your response is that I haven’t kept up with some courtroom relitigation. Congrats! And kudos on knowing about denning kruger.
deltaFx2 - Wednesday, May 19, 2021 - link
As to patents, a) patents make broad claims for winning lawsuits. Implementation may or may not be exactly as specified in them and b) your company lawyer will advise against reading patents as it may be used against youJon Tseng - Tuesday, May 18, 2021 - link
D'ya reckon they'll be tortilla chip care packs again this year?? :-pname99 - Tuesday, May 18, 2021 - link
Are these PIM solutions (and the flash equivalent, where you do the processing in the flash drive) actually being used?Obviously they're not being used by Apple or in your basic Wintel laptop, but are they being used by cloud providers, or HPC or universities or anywhere?
For 20 years this ideas has been pushed academically, and for maybe ten years I've been seeing "product-like" announcements. But never any reference to these things in the wild.
I am well aware that I don't know the full range of computing. Hence this question. Not snark, just a genuine curiosity as to whether/when/if these things will exit the lab.
eastcoast_pete - Tuesday, May 18, 2021 - link
Should be an exciting conference! One comment about the sponsorship levels, which currently top out at "Rhodium"; they should have added an "Unobtainium" level above that, would be very a propos the current shortages and prices.jospoortvliet - Wednesday, May 19, 2021 - link
Ha and then not actually offer that level to anyone ;-)yojanas001 - Thursday, June 17, 2021 - link
When we find issues related to this type of subject that's not easy to find but some people like you make it easy for us. Thanks for giving us precious time.https://governmentyojanas.com/sssm-id-samagra-sama...
rossmandor - Thursday, September 9, 2021 - link
Nice!