Intel is a 'Rhodium' sponsor of Hot Chips this year, which is one above the standard top Platinum sponsorship level. Last year they sponsored all the lunches, break sessions, and the post-event breakouts.
I always found these levels a joke, especially on people who don;t understand that they don't mean anything. I've seen so many people flaunting "platinum" cards at different events without realizing that the order was Platinum<Titanium<Carbon or something along those lines. Or that one guy in a phone store shouting that he's a Gold customer not knowing that for that company Gold was s*literally* the lowest level. Everyone has to feel good so they all get a crown, they're all belle of the ball. Gold, Diamond, Graphite, Virtuoso, Opal, Carbon, Platinum, Ultimate, Wolfram, Unobtainium, Master Mason. In any order.
What the... that TPU v3 board looks like it has swollen and possibly leaking aluminum electrolytic capacitors! The only time I've seen caps that had bulging rounded tops like that was on hardware made between 1999 and 2007.
I'm hoping it's just a trick of the light (though the one really looks like it's leaking). Still... Why doesn't it use solid caps?
Money. It is hard to tell from that photo, but it looks like there are 15 capacitors at 100 uF, 100 V, probably these: EKZE101ELL101MK16S which run $0.30 each. So, that would be $4.50 of capacitors. After a quick search the cheapest non-electrolitic capacitors that I could find at 100 uF and 100 V are these: R60EW61005000K which run over $9 each in bulk. So, that would be $135 of capacitors.
I went with 100 V because the first capacitor on the bottom right of the column looks like it says 100 V upside down. The rest look like model KZE and 100 uF. Of course, a better photo would make this a lot easier to tell.
If those film caps are the only others that are available (I checked too and that seems to be the case), then the problem is most likely that there are no suitable caps available with those specifications other than aluminum electrolytics, so they are using what is available. Perhaps they are using massively overrated capacitors (higher voltage) to try to compensate for a lack of available higher-end alternatives. I don't know a lot about this stuff but 100v seems incredibly high for any motherboard.
Still... they must be horribly mistreating these things for them to look like bloated, leaking Fujiyuu caps pulled from a generic PSU made in 2005. The ones in the other picture don't look like that but they're definitely the same caps.
Several look like they are leaking and almost all look bulged. Maybe the reason that board was used for display is because its a failed board with bad caps?
Cheers for the update mate always wanted to fly out for it but always clashed w the summer break. This year with no summer break seems like a no brainer just signed up! J
Ian, you keep mentioning that the Infinity Fabric has been decoupled from main memory as a new thing on Zen2 APU's -- when in fact regular old Matisse has fully decoupled infinity fabric as well. You can set your IF speed to anything you want -- it doesnt need to be coupled to the memory speed, or memory controller speed.
It seems that the new feature is that on the Zen2 APU's the IF can scale dynamically with load -- whereas on Matisse the IF clock speed is static. If you set it to 1800Mhz, it is locked at 1800Mhz all the time, where it seems like on the APU's it can scale up and down while running. Can you verify this please?
Thanks also for flagging the Jim Keller profile - thankfully Fortune are dumb enough to embed the plaintext of the article in the page source, even if you can't bust the paywall! :-x
AMD has been repeatedly clear that Zen 3 is, architecturally speaking, a minor change over Zen 2 so I’m not surprised that they’re not discussing it. HotChips is where people go to present or learn about major design changes and challenges.
You have it backwards. Zen2 was more of "just an upgrade" of Zen and Zen+, while Zen3 will be a major architecture change. The wording was along the lines that it was unusual for Zen2 to be as big of an improvement as it was because it was NOT a new architecture. Zen3 is a new architecture.
All from the same designs that Keller left then with - minor iterative upgrades - not a new arch. Zen was a new arch ... Zen 3 is still Zen. Much like Intel made changes to Skylake - AMD is making the same sort of changes to Zen.
Google tends to use 48V DC power from the rack with local voltage regulators bringing it down to the right voltage on the motherboards, these capacitors are likely the input and possibly output filtering for this.
These input capacitors thus needs to be at least 63V rated and using 100V capacitors isn't uncommon for 48V DC systems. Using through-hole capacitors for input filtering is perfectly sensible and besides high-voltage SMD capacitors are really expensive.
The output capacitors could easily be SMD aluminium-poly or tantalum-poly that is less visible (extremely low-profile and COULD be on the back!).
If they did use some of the through-hole capacitors for output, yes, they probably could use 2.5V! or 5V rated capacitors for this but for through-holes the price difference likely was so small that they used the 100V models to save space in the "picker" (the machine that places all components on the motherboard - if you have too many different components you need extra passes!) and/or their parts storage.
Thanks for the info. Sounds reasonable to me. I was suggesting that they use through-hole solid caps but I see the selection of those is fairly limited too.
Anyway, I'm not even remotely qualified to critique the engineers who actually design these things, but I'm still amazed by the visibly bad capacitors. That just isn't a common way for caps to fail these days unless they were made during the "plague" era of the early to mid 2000s. Looks really bad on display for such a cutting edge device.
They will have to come up with new material ... and of course they will never retire the "Yeah, this AMD generation is good, but wait until the NEXT one"
The Xe talk is by David Blythe, the chief architect on Xe. That talk could go anywhere he wants, but I wouldn't guess that he'd be spending a lot of time talking about the oneAPI software.
He did blurt out something about Xe ray tracing acceleration previously, so maybe someone should follow up on that.
Sorry Ian, it's actually the absolute worst conference, for the very reasons you describe.
Because people like me have to live through half a year's worth of suspense and anticipation being teased and outright tortured with hints and half-facts by people like you, until they finally lower the paywall for plebs like me.
Oops, without the flights, hotels and this entry fee: It really has become a bargain!
Pretty sure the ETH design is based on their PULP and HERO research, targeted at very low power edge inferencing, even if the core count is high.
It's an 'architecture' because all ETH PULP chips are open source with HDL currently for STM FD-SOI. It comes validated with test chips in that process, but they don't sell chips or OEM support services.
So you can then fab your own variant directly with STM I guess, or you can go with one of the existing partner companies, who tend to package it with additional IP, produce smaller production chip runs, perhaps PCBs and provide OEM support.
And you would be free to transform the design to another process, if that made any sense commercially.
If you couldn't guess, abufrejoval was being a little sarcastic - the fact he has to wait for the content of the conference to be public, suggesting that they say a lot but he has to wait. Hot Chips is great for content.
Intel continues to rehash the same stuff with very minor improvements. Until they actually prove they can get back on track and deliver meaningful upgrades, I don't know that you can trust anything they have to say about roadmaps or future plans.
And if you're a big OEM, such as Dell or HP, maybe you should be rethinking your vendor lock-in, if you haven't already.
Are the bigger OEMs really locked into Intel for processor technologies through some contractual agreement? I vaguely recall court cases in which Intel was involved with creating supply contracts that contained prohibitive or punitive clauses should a company decide to purchase processors from other companies (AMD pretty much in specific since it was the only credible alternative) but those sorts of contracts landed Intel in lawsuits, resulted in bad press, and I don't think supply contracts still contain those clauses these days.
It seems more like the larger OEMs purchase mainly Intel processors for the sake of platform integration and other value-adds along with consideration for buyer demand and product fit for the intended purpose. Not to say that AMD isn't offering credible competition these days across all performance metrics we commonly measure. Beyond that, the big OEMs do sell modern AMD systems, but since laptops represent the vast majority of computer sales and AMD's presence there has not been competitive across the board until VERY recently (namely with the release of 4000-series APUs) Intel has a lot of existing momentum and we won't see AMD-based designs in large numbers right away. Pandemic economic decline is also probably influencing conservative, cost-cutting decisions at corporate offices across the board so expect that process to ramp up even more slowly than it would otherwise. There is just going to be a lot less demand-side pull on the supply chain for the foreseeable future.
Ian, Please don't eat the samples! We've talked about this before--it tends to make you a walking conductor--stay away from wall sockets for at least an hour after eating! I know it can be very hard--what with your friends egging you on and yelling, "Bet you can't eat just one!" and so on. Hang in there!...;)
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
64 Comments
Back to Article
Xex360 - Thursday, May 21, 2020 - link
Can you make a piece regarding quantum computing. And whether it's the future of computing, or part of it stuck as security??MrSpadge - Wednesday, July 8, 2020 - link
In short: sure it's part of the future. When and in which form it may also become part of the present is still up to R&D to decide.wr3zzz - Thursday, May 21, 2020 - link
Interesting that Alibaba, Baidu and Google are present but not Amazon.mmrezaie - Wednesday, July 8, 2020 - link
Amazon unfortunately doesn't represent much in conferences.ikjadoon - Wednesday, July 8, 2020 - link
Following Apple's lead.Two of the fastest ARM systems available today, in their respective segments, and next to nothing but virtualized SPEC benchmarks.
Anandtech has done great work, but these two are—to the dismay of the technology industry, but perhaps not consumers—zipped lips 24/7.
Teckk - Thursday, May 21, 2020 - link
Lot of sessions by Intel. Hopefully we'll get some good details on 10nm and 7nm too.Ian Cutress - Thursday, May 21, 2020 - link
Intel is a 'Rhodium' sponsor of Hot Chips this year, which is one above the standard top Platinum sponsorship level. Last year they sponsored all the lunches, break sessions, and the post-event breakouts.ingwe - Thursday, May 21, 2020 - link
They should have gone with 'Iridium'. Not that it at all matters.eastcoast_pete - Friday, May 22, 2020 - link
Or "Plutonium"; 10 nm and below is probably a radioactive subject at Intel right now.Arnulf - Sunday, May 24, 2020 - link
At least they have quite some hot chips to talk about.twtech - Wednesday, July 8, 2020 - link
Hot because they're still 14nm.close - Thursday, July 9, 2020 - link
I always found these levels a joke, especially on people who don;t understand that they don't mean anything. I've seen so many people flaunting "platinum" cards at different events without realizing that the order was Platinum<Titanium<Carbon or something along those lines. Or that one guy in a phone store shouting that he's a Gold customer not knowing that for that company Gold was s*literally* the lowest level. Everyone has to feel good so they all get a crown, they're all belle of the ball. Gold, Diamond, Graphite, Virtuoso, Opal, Carbon, Platinum, Ultimate, Wolfram, Unobtainium, Master Mason. In any order.CrystalCowboy - Wednesday, July 8, 2020 - link
This would be the best year to sponsor lunches. Shouldn't cost much at all.twtech - Wednesday, July 8, 2020 - link
Doordash for 1700 online attendees could actually be quite expensive.Strunf - Thursday, July 9, 2020 - link
Seeing that a sandwich menu is going for $65 I'm not sure about that.FreckledTrout - Wednesday, August 5, 2020 - link
LOL This is funny now that 7m is delayed months later :)ozzuneoj86 - Thursday, May 21, 2020 - link
What the... that TPU v3 board looks like it has swollen and possibly leaking aluminum electrolytic capacitors! The only time I've seen caps that had bulging rounded tops like that was on hardware made between 1999 and 2007.I'm hoping it's just a trick of the light (though the one really looks like it's leaking). Still... Why doesn't it use solid caps?
dullard - Thursday, May 21, 2020 - link
Money. It is hard to tell from that photo, but it looks like there are 15 capacitors at 100 uF, 100 V, probably these: EKZE101ELL101MK16S which run $0.30 each. So, that would be $4.50 of capacitors. After a quick search the cheapest non-electrolitic capacitors that I could find at 100 uF and 100 V are these: R60EW61005000K which run over $9 each in bulk. So, that would be $135 of capacitors.PixyMisa - Thursday, May 21, 2020 - link
These are expensive, low-volume boards, but still, $135 extra to the BOM is not insignificant.extide - Thursday, May 21, 2020 - link
Yeah they are probably 10v caps, nmo need for 100v caps on a board like that. That would make the cost a lot easier to swallow.dullard - Thursday, May 21, 2020 - link
I went with 100 V because the first capacitor on the bottom right of the column looks like it says 100 V upside down. The rest look like model KZE and 100 uF. Of course, a better photo would make this a lot easier to tell.ozzuneoj86 - Thursday, May 21, 2020 - link
There's another angle here. Definitely looks like 100uf 100v:https://3s81si1s5ygj3mzby34dq6qf-wpengine.netdna-s...
If those film caps are the only others that are available (I checked too and that seems to be the case), then the problem is most likely that there are no suitable caps available with those specifications other than aluminum electrolytics, so they are using what is available. Perhaps they are using massively overrated capacitors (higher voltage) to try to compensate for a lack of available higher-end alternatives. I don't know a lot about this stuff but 100v seems incredibly high for any motherboard.
Still... they must be horribly mistreating these things for them to look like bloated, leaking Fujiyuu caps pulled from a generic PSU made in 2005. The ones in the other picture don't look like that but they're definitely the same caps.
extide - Thursday, May 21, 2020 - link
Hrmm, interesting.flgt - Thursday, May 21, 2020 - link
Would custom data center hardware use 48VDC power distribution now?PixyMisa - Thursday, May 21, 2020 - link
Facebook does 48V power distribution. Wouldn't be surprised if Google does too but I haven't seen anything specific.jeremyshaw - Wednesday, July 8, 2020 - link
Even if the external power isn't 48V, the internal power may be. Nvidia's DGX pods have been using 48V internally.zmatt - Thursday, May 21, 2020 - link
Several look like they are leaking and almost all look bulged. Maybe the reason that board was used for display is because its a failed board with bad caps?Jon Tseng - Thursday, May 21, 2020 - link
Cheers for the update mate always wanted to fly out for it but always clashed w the summer break. This year with no summer break seems like a no brainer just signed up! Jextide - Thursday, May 21, 2020 - link
Ian, you keep mentioning that the Infinity Fabric has been decoupled from main memory as a new thing on Zen2 APU's -- when in fact regular old Matisse has fully decoupled infinity fabric as well. You can set your IF speed to anything you want -- it doesnt need to be coupled to the memory speed, or memory controller speed.It seems that the new feature is that on the Zen2 APU's the IF can scale dynamically with load -- whereas on Matisse the IF clock speed is static. If you set it to 1800Mhz, it is locked at 1800Mhz all the time, where it seems like on the APU's it can scale up and down while running. Can you verify this please?
Jon Tseng - Thursday, May 21, 2020 - link
Thanks also for flagging the Jim Keller profile - thankfully Fortune are dumb enough to embed the plaintext of the article in the page source, even if you can't bust the paywall! :-xsmilingcrow - Thursday, May 21, 2020 - link
Fortunate for us but not so fortunate for Fortune. ;)Thanks.
FunBunny2 - Thursday, May 21, 2020 - link
why nothing about storage class memory? would seem to be a big answer to a big question.brucethemoose - Thursday, May 21, 2020 - link
Intel will probably talk up Optane for Ice Lake. MS will talk up the Xbox's flash.Other than that, its pretty far away from mass commercialisation. Then again, so is quantum computing....
ksec - Thursday, May 21, 2020 - link
Looks like still too early for Zen 3Elstar - Friday, May 22, 2020 - link
AMD has been repeatedly clear that Zen 3 is, architecturally speaking, a minor change over Zen 2 so I’m not surprised that they’re not discussing it. HotChips is where people go to present or learn about major design changes and challenges.Targon - Friday, May 22, 2020 - link
You have it backwards. Zen2 was more of "just an upgrade" of Zen and Zen+, while Zen3 will be a major architecture change. The wording was along the lines that it was unusual for Zen2 to be as big of an improvement as it was because it was NOT a new architecture. Zen3 is a new architecture.Deicidium369 - Sunday, May 24, 2020 - link
All from the same designs that Keller left then with - minor iterative upgrades - not a new arch. Zen was a new arch ... Zen 3 is still Zen. Much like Intel made changes to Skylake - AMD is making the same sort of changes to Zen.arashi - Monday, May 25, 2020 - link
Gimme whatever you're smoking, coz that's some good shit.TLindgren - Thursday, May 21, 2020 - link
Google tends to use 48V DC power from the rack with local voltage regulators bringing it down to the right voltage on the motherboards, these capacitors are likely the input and possibly output filtering for this.These input capacitors thus needs to be at least 63V rated and using 100V capacitors isn't uncommon for 48V DC systems. Using through-hole capacitors for input filtering is perfectly sensible and besides high-voltage SMD capacitors are really expensive.
The output capacitors could easily be SMD aluminium-poly or tantalum-poly that is less visible (extremely low-profile and COULD be on the back!).
If they did use some of the through-hole capacitors for output, yes, they probably could use 2.5V! or 5V rated capacitors for this but for through-holes the price difference likely was so small that they used the 100V models to save space in the "picker" (the machine that places all components on the motherboard - if you have too many different components you need extra passes!) and/or their parts storage.
ozzuneoj86 - Thursday, May 21, 2020 - link
Thanks for the info. Sounds reasonable to me. I was suggesting that they use through-hole solid caps but I see the selection of those is fairly limited too.Anyway, I'm not even remotely qualified to critique the engineers who actually design these things, but I'm still amazed by the visibly bad capacitors. That just isn't a common way for caps to fail these days unless they were made during the "plague" era of the early to mid 2000s. Looks really bad on display for such a cutting edge device.
Jumangi - Friday, May 22, 2020 - link
I didn't know Ian was such a "biter".Arbie - Wednesday, July 8, 2020 - link
You must have missed the first dozen times those pictures were posted.Beaver M. - Friday, May 22, 2020 - link
TL;DRSo is Intels next Desktop generation getting 14 nm again?
eastcoast_pete - Friday, May 22, 2020 - link
Well, the conference is called "Hot Chips", so their 14 nm++++(!) would fit right in. With those thermals, hot chips indeed.Eulytaur - Sunday, May 24, 2020 - link
I would usually like to clown with you but this brings the introduction of 10nm Xeon, the era is ending.Deicidium369 - Sunday, May 24, 2020 - link
They will have to come up with new material ... and of course they will never retire the "Yeah, this AMD generation is good, but wait until the NEXT one"JayNor - Thursday, May 28, 2020 - link
The Xe talk is by David Blythe, the chief architect on Xe. That talk could go anywhere he wants, but I wouldn't guess that he'd be spending a lot of time talking about the oneAPI software.He did blurt out something about Xe ray tracing acceleration previously, so maybe someone should follow up on that.
JayNor - Thursday, May 28, 2020 - link
This article covered David Blythe's presentation on gen11 graphics. Perhaps something similar for Xe...https://techreport.com/review/34319/intel-talks-ab...
abufrejoval - Tuesday, June 2, 2020 - link
Sorry Ian, it's actually the absolute worst conference, for the very reasons you describe.Because people like me have to live through half a year's worth of suspense and anticipation being teased and outright tortured with hints and half-facts by people like you, until they finally lower the paywall for plebs like me.
abufrejoval - Tuesday, June 2, 2020 - link
Oops, without the flights, hotels and this entry fee: It really has become a bargain!Pretty sure the ETH design is based on their PULP and HERO research, targeted at very low power edge inferencing, even if the core count is high.
It's an 'architecture' because all ETH PULP chips are open source with HDL currently for STM FD-SOI. It comes validated with test chips in that process, but they don't sell chips or OEM support services.
So you can then fab your own variant directly with STM I guess, or you can go with one of the existing partner companies, who tend to package it with additional IP, produce smaller production chip runs, perhaps PCBs and provide OEM support.
And you would be free to transform the design to another process, if that made any sense commercially.
Beaver M. - Wednesday, July 8, 2020 - link
Never was a fan of conferences like these.They talk a lot but dont say much.
Ian Cutress - Wednesday, July 8, 2020 - link
If you couldn't guess, abufrejoval was being a little sarcastic - the fact he has to wait for the content of the conference to be public, suggesting that they say a lot but he has to wait. Hot Chips is great for content.Brane2 - Wednesday, July 8, 2020 - link
As photos show, you often chew much more than you can swallow.iranterres - Wednesday, July 8, 2020 - link
Explains why Anandtech is orbiting heavily around Intel on this year's event.twtech - Wednesday, July 8, 2020 - link
Intel continues to rehash the same stuff with very minor improvements. Until they actually prove they can get back on track and deliver meaningful upgrades, I don't know that you can trust anything they have to say about roadmaps or future plans.And if you're a big OEM, such as Dell or HP, maybe you should be rethinking your vendor lock-in, if you haven't already.
Deicidium369 - Thursday, July 9, 2020 - link
OEMs will build what their customers want, and it seems what the don't want is AMD."AMD continues to rehash the same stuff with very minor improvements" - I fixed that for you.
PeachNCream - Friday, July 10, 2020 - link
Are the bigger OEMs really locked into Intel for processor technologies through some contractual agreement? I vaguely recall court cases in which Intel was involved with creating supply contracts that contained prohibitive or punitive clauses should a company decide to purchase processors from other companies (AMD pretty much in specific since it was the only credible alternative) but those sorts of contracts landed Intel in lawsuits, resulted in bad press, and I don't think supply contracts still contain those clauses these days.It seems more like the larger OEMs purchase mainly Intel processors for the sake of platform integration and other value-adds along with consideration for buyer demand and product fit for the intended purpose. Not to say that AMD isn't offering credible competition these days across all performance metrics we commonly measure. Beyond that, the big OEMs do sell modern AMD systems, but since laptops represent the vast majority of computer sales and AMD's presence there has not been competitive across the board until VERY recently (namely with the release of 4000-series APUs) Intel has a lot of existing momentum and we won't see AMD-based designs in large numbers right away. Pandemic economic decline is also probably influencing conservative, cost-cutting decisions at corporate offices across the board so expect that process to ramp up even more slowly than it would otherwise. There is just going to be a lot less demand-side pull on the supply chain for the foreseeable future.
JorgeE1 - Thursday, July 9, 2020 - link
HOT CHIPSState of Technology: CPUs, Hardware, Integrated Platforms and SoCs
WaltC - Thursday, July 9, 2020 - link
Ian, Please don't eat the samples! We've talked about this before--it tends to make you a walking conductor--stay away from wall sockets for at least an hour after eating! I know it can be very hard--what with your friends egging you on and yelling, "Bet you can't eat just one!" and so on. Hang in there!...;)Deicidium369 - Thursday, July 9, 2020 - link
He's a professional wafer biter - pretty much sure he knows all the common sense advice you gave him.Deicidium369 - Friday, July 10, 2020 - link
Why does this story keep popping up, when the event is in August?Ian Cutress - Thursday, July 30, 2020 - link
Updates. Changing the keynote speaker.hacktolife - Tuesday, September 29, 2020 - link
I read many articles But this article to good for all of them. I am telling you a nice article<a href="https://hacktolife.com/how-to-make-homemade-orange... How to make orange juice</a>