Significant leaks: Intel working on true 3D packaging and Raja wants to enter dGPU "with a bang"

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
Significant leaks: Intel working on true 3D packaging and Raja wants to enter dGPU "with a bang"

Ashraf Eassa, long-time Intel analyst at The Motley Fool (https://twitter.com/TMFChipFool), has become known for his reputation as being one of the most trustworthy and consistent leakers of insider information from Intel, such as code names.

Among other things, in the last half a year, he was the first (but sadly the tech press has not yet caught up to any of this information nor given him the credits he deserves) to tell us about Whiskey Lake (and that it's 14++), Alder Lake, Monette Hill, Lakefield being big.little Core/Atom mashup, that the first 10nm Cannon Lake part having been PRQ'd is a 2+0 SKU, Arctic Sound (Gen12) and Jupiter Sound (Gen13) dGPUs, Meteor Lake (7nm), Granite Rapids (7nm), Raja making changes to Arctic Sound to make it more competitive for gaming, Intel working on CUDA competitor for release this year, Arctic Sound using chiplets and Intel making a dedicated ASIC for inference. (If you want a source for any of these, feel free to ask.)

Now, he has leaked a few little more tidbits for you to enjoy.

The leaks

More specifically, on the CPU side, he confirms that Icelake-SP is still targeted for 2019 (last year D. Schor said May'19) and Sappire Rapids is still targeted for 2020 (a roadmap two years ago suggested mid-'20). As Intel said, there will be 3 gens of 10nm (with ICL being 10+ and SPR being 10++). Given Intel's yearly cadence, by the way, this implies 7nm (Granite Rapids) is still planned for 2021. (Intel's annual report suggested Intel has already started working on Fab 42 in 2017.)

Further, Intel has put Lakefield's power management in Tiger Lake for very low standby of 9mW, something Apple execs (who are likely familiar with iPhone's feeds and speeds) were "surprised" with.

Next, on the packaging side, Ashraf has come with the major scoop that Intel is working on a 3D packaging technology called Fevoros, which will be able to package "smaller dies together as well as 3D stacking". It will be used post-Sapphire Rapids (so starting with 7nm Granite Rapids); EMIB will be used until SPR.

For some context as to what this significant news implies. It will be the successor of EMIB. EMIB is a 2.5D packaging technology that allowed for connecting dies (that are within um next to each other) with a low-power and very high-bandwidth interconnect (potentially over 4TB/s). EMIB's claim to fame and differentiator compared to the competition is that it doesn't use a very big (expensive) interposer, nor does it use yield-destroying TSVs (through silicon vias), but just a small embedded bridge in the package (as the name implies obviously).

Now, I have to say 3D stacking is already done: it is used in HBM. But as everyone knows Intel is not a DRAM manufacturer, so this implies that Intel is going to 3D stack logic dies (or 3D XPoint?). This would be a first. Everyone in the industry is interested in how well cooling of 3D stacked logic dies will go.

I would guess that this will be an extension of EMIB: Intel is working on multiple generations of EMIB, for instance we know Falcon Mesa will use EMIB2, which will shrink the bump pitch to 35um from 55um for 2.5x the bandwidth, and Intel has previously said they were seeing as low as 10um in the labs.

Finally, on the GPU side of things, Ashraf has given a few more details about Arctic Sound (10++ Gen12 based, the GPU architecture that will be in Tiger Lake). He says Arctic Sound was initially targeted as an accelerator in the data center for "video streaming apps". (We know Intel is fully embracing the heterogeneous future of compute, with things like FPGAs, training and inference ASICs and GPUs.) With Raja on board, Arctic Sound has been "split into two", with the second side of the coin being the gaming market. He is said he wants to "enter the market with a bang". So it is obvious Intel will be giving NVIDIA a hard time (which is necessary because NVIDIA is doing very well wrt execution), bringing significant competition to the market.

Arctic Sound is targeted for PRQ in H1'20 (which in any case implies a 2020 launch). He also says it will consist of an MCP with up to 4 dies. (I assume these chiplets will be connected via EMIB, since that's the only thing that would be feasible.)

Now, the interesting bit here is that he says they're targeting Jupiter Sound to have leadership (i.e. >NVIDIA) performance/watt "under 60W". Now, 60W doesn't sound very high-end to me, but given that they're going the MCP route, I'd presume the 60W is for just one chiplet. So if one extrapolates to 4 chiplets, that makes for a high-end (leadership) product of 240W of 7nm Gen13 goodiness in 2021.

To end, he also notes Knights Crest will be fabbed in-house, but everyone had thought that already. But Intel is also working on a dedicated DL chip for inference called Spring Hill.

Sources
He has not disclosed his sources, but he often has multiple independent ones and he has a good track record.

1. Arctic Sound PRQ=H1'20, targeting perf/watt leadership under 60W
2. Knights Crest will be in-house fabbed (that Spring Crest is TSMC was already known)
3. Intel's DL inference ASIC (was already known Intel was working on this based on job postings) called Spring Hill
4. Intel working on EMIB successor for true 3D packaging, called Fevoros
5. Fevoros will succeed EMIB post-Sapphire Rapids, ICL-SP targeted for 2019, SPR for 2020
6. There will be "only" 3 gens of 10nm (10, 10+, 10++)
7. Raja has redefined Arctic Sound, split in two: orignally for video streaming in data center, now also specifically for gaming, want to enter market with a bang
8. Intel has pulled the the power management from Lakefield into Tiger Lake, has very low standby power of 9mW
9. Arctic Sound will consist of 2-4 chiplets
10. Apple execs were said to be "surprised" by (35W) Tiger Lake's low standby power of 9mW

 
Last edited:

VirtualLarry

No Lifer
Aug 25, 2001
56,229
9,990
126
Finally, on the GPU side of things, Ashraf has given a few more details about Arctic Sound (10++ Gen12 based, the GPU architecture that will be in Tiger Lake). He says Arctic Sound was initially targeted as an accelerator in the data center for "video streaming apps". (We know Intel is fully embracing the heterogeneous future of compute, with things like FPGAs, training and inference ASICs and GPUs.) With Raja on board, Arctic Sound has been "split into two", with the second side of the coin being the gaming market. He is said he wants to "enter the market with a bang". So it is obvious Intel will be giving NVIDIA a hard time (which is necessary because NVIDIA is doing very well wrt execution), bringing significant competition to the market.
LOL. I'll believe Intel taking the "Gamer market by storm", when I actually see it. I'm not giving them any benefit of the doubt in this area, since they've tried MULTIPLE times, and FAILED.
 

NTMBK

Lifer
Nov 14, 2011
10,208
4,940
136
It's going to be interesting to see this all play out. If Arctic Sound really was aimed at the video streaming market, it was probably meant as a replacement for the Intel Visual Compute Accelerators: https://www.intel.com/content/www/us/en/products/servers/accelerators.html Which literally strapped three CPUs with integrated Iris Pro graphics onto a PCIe card. The question is whether Arctic Sound is a single big GPU which was meant to replace those multiple iGPUs, or whether it is a small GPU and the PCIe card would mount multiple GPUs. I suspect it is a small GPU, as a big GPU solution would require very robust virtualization support for their datacenter customers' needs- not something I've seen any evidence of Intel developing. Whereas a small GPU can also be reused as a replacement for the AMD GPU in Kaby Lake G, and aimed at the laptop market.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
(1) Please someone update Wikipedia ;-)
(2) Spread this info to the tech press please

I like Ashraf for his leaks, and they seem quite accurate.

But I'd be very weary of a big press like Anandtech using it as a pipeline story. They don't do it because once started, the line between rumor and leaks become very blurry and might quickly end up as a site viewed as untrustworthy as WCCFTech. Because our ability to discern things as individuals vary, without having something solid it becomes nothing but a rumor. It needs to have something, before they can even think of putting it as a news story. Trust can be lost far quicker than the time it took to gain it.

There also have been sources that were correct a few times, but fizzed out after that. Ultimately the rumors stay as rumors until evidence starts to surface.
 

imported_bman

Senior member
Jul 29, 2007
262
54
101
Good to see Intel's pipeline having interesting products for the next few years.

https://twitter.com/TMFChipFool/status/982083072823840768

If I am reading this right he has heard that Icelake-Y is scheduled for Q4 2018. That would be pretty nice to have quad core fanless on the market by Christmas. Does this mean that Cannon-Y has been abandoned outside of industrial/embedded applications?
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
If I am reading this right he has heard that Icelake-Y is scheduled for Q4 2018. That would be pretty nice to have quad core fanless on the market by Christmas. Does this mean that Cannon-Y has been abandoned outside of industrial/embedded applications?

There's supposed to be what looks like Cannonlake-based Intel NUCs coming. Other than that, it seems pretty quiet. If that's all we get for Cannonlake-Y it does open up the possibilities for late 2018 ICL-Y.
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
More specifically, on the CPU side, he confirms that Icelake-SP is still targeted for 2019 (last year D. Schor said May'19)

Yeah, that aint happening. BK said that the Optane DIMMs aren't even going to be material in 2018 so you have an idea of how little Cascade Lake is going to be available this year outside of maybe Google and Amazon. Which obviously impacts when they would realistically ship Icelake-SP even with Matisse looming.

If I am reading this right he has heard that Icelake-Y is scheduled for Q4 2018. That would be pretty nice to have quad core fanless on the market by Christmas. Does this mean that Cannon-Y has been abandoned outside of industrial/embedded applications?

Same sort of deal, Icelake-Y might get announced in Q4 but you won't be able to buy it until next year. Cannonlake-Y is completely dead, the only thing coming is a U part.
 

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
I like Ashraf for his leaks, and they seem quite accurate.

But I'd be very weary of a big press like Anandtech using it as a pipeline story. They don't do it because once started, the line between rumor and leaks become very blurry and might quickly end up as a site viewed as untrustworthy as WCCFTech. Because our ability to discern things as individuals vary, without having something solid it becomes nothing but a rumor. It needs to have something, before they can even think of putting it as a news story. Trust can be lost far quicker than the time it took to gain it.

There also have been sources that were correct a few times, but fizzed out after that. Ultimately the rumors stay as rumors until evidence starts to surface.
Obviously not all tech press post rumors/leaks.

But for instance press like Next Platform and other big guys covered the big Skylake-SP leak in 2015, although it was very clear that it was a leak because it came from an Intel power point.

Of course here there are no power points So that's why you have take someone's track record into account. For instance when Even Blass posts something everyone takes it seriously.

In any case, the media exists to inform the public, so when a reputable source gives a lot of information, then sure some media should be interested.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Yeah, that aint happening. BK said that the Optane DIMMs aren't even going to be material in 2018 so you have an idea of how little Cascade Lake is going to be available this year outside of maybe Google and Amazon.

It doesn't have to be this way. Cascade Lake SP is Q2/Q3. Optane DIMMs aren't going to usable in all Cascade Lake SP chips. You could have the Optane DIMM supporting ones in Q4, but the platform itself much earlier. The Skylake SP chips were in the leaks to have on-package FPGA variants, but we'll only see them later this year.

Same sort of deal, Icelake-Y might get announced in Q4 but you won't be able to buy it until next year. Cannonlake-Y is completely dead, the only thing coming is a U part.

You are right on the Cannonlake-Y part. I think even the NUC is -U.
 
  • Like
Reactions: witeken
Mar 10, 2006
11,715
2,012
126
It doesn't have to be this way. Cascade Lake SP is Q2/Q3. Optane DIMMs aren't going to usable in all Cascade Lake SP chips. You could have the Optane DIMM supporting ones in Q4, but the platform itself much earlier. The Skylake SP chips were in the leaks to have on-package FPGA variants, but we'll only see them later this year.



You are right on the Cannonlake-Y part. I think even the NUC is -U.

Optane DIMM supporting SKUs could come later, they are distinct from the standard Cascade Lake parts. So you are correct.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Witeken: I think the type of information Ashraf leaks are great for forum discussions. But I can totally understand why you wouldn't want them on your site. Because things will happen with time anyway, its best to take a conservative approach to things to not ruin your reputation, which is more important than getting off the record to the front page as soon as possible.

Optane DIMM supporting SKUs could come later, they are distinct from the standard Cascade Lake parts. So you are correct.

We can discern quite a lot about what Intel was planning just using the Purley leaks.

https://3s81si1s5ygj3mzby34dq6qf-wp...ntent/uploads/2015/05/intel-kdm-roadmap-2.jpg

Skylake+FPGA: Will be here this year
Cannonlake graphics & media transcode: So the plans for expanding graphics existed for quite some time. Since they are talking about not just media, but graphics, and its on a server platform, they want to expand the graphics part. It's probably delayed, but by ICL-SP time I think we'll get a much clearer picture.
 

witeken

Diamond Member
Dec 25, 2013
3,899
193
106
Witeken: I think the type of information Ashraf leaks are great for forum discussions. But I can totally understand why you wouldn't want them on your site. Because things will happen with time anyway, its best to take a conservative approach to things to not ruin your reputation, which is more important than getting off the record to the front page as soon as possible.
Fair enough, but just Google Intel roadmap. Many sites are eager to post these kind of things ;).
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
My opinion on the leaks.

Knights Crest on Intel process. I think this could be easily deduced because the "Knights" naming is used on their Xeon Phi chips. In fact, I think it wouldn't be too surprising if Knights Crest is Goldmont Plus, even Tremont cores but with Nervana IP and Deep Learning focused. Intel did say it was a "bootable Intel Xeon processor", same as for Knights Landing Xeon Phi.

60W seems ok for a product, something that will succeed Kabylake-G. With 200-240W, they'd have to nail everything, and that is so much harder to do, so I think its less likely.

Lakefield is the most exciting part for me. Not for the claimed power management improvements, but hetereogenous core combo of big and little cores.

The standby power consumption claims for Tigerlake needs more information. I can tell you easily based on what the number represents, it can either be "meh", or "impressive".
 

jpiniero

Lifer
Oct 1, 2010
14,510
5,159
136
I had thought maybe Cascade was delayed because of Meltdown, and with it Optane. Optane is unoptaneable.

That and Spectre too. Have to think they did a couple more respins, and that takes time. It kind of works out since it's obvious that even now 10 nm yields are still in the toilet so it's not like they could ship Icelake-SP early anyway.

Fevoros sounds like the chip stacking mentioned in the Sea of Cores patent. Cooling the bottom dies is an obvious issue.
 

LTC8K6

Lifer
Mar 10, 2004
28,520
1,575
126
LOL. I'll believe Intel taking the "Gamer market by storm", when I actually see it. I'm not giving them any benefit of the doubt in this area, since they've tried MULTIPLE times, and FAILED.
We need the competition very badly, so I hope they do enter the ring with something interesting.
 

nathanddrews

Graphics Cards, CPU Moderator
Aug 9, 2016
965
534
136
www.youtube.com
LOL. I'll believe Intel taking the "Gamer market by storm", when I actually see it. I'm not giving them any benefit of the doubt in this area, since they've tried MULTIPLE times, and FAILED.
Not to sound overly critical, but I don't remember them trying. Ever. They always talk a big game, but have never actually delivered on a single GPU design worth having. I think that if they were serious and actually trying, they would have had at least one competitive product in the past 40 years.

To be fair, they absolutely dominate the entry-level market with IGP.
 
  • Like
Reactions: Burpo

wahdangun

Golden Member
Feb 3, 2011
1,007
148
106
Not to sound overly critical, but I don't remember them trying. Ever. They always talk a big game, but have never actually delivered on a single GPU design worth having. I think that if they were serious and actually trying, they would have had at least one competitive product in the past 40 years.

To be fair, they absolutely dominate the entry-level market with IGP.


I think the most fundamental problem was in driver and making big gpu, because big gpu is hard, expensive and have so many IP minefield.
 
Last edited:

realibrad

Lifer
Oct 18, 2013
12,337
898
126
Not to sound overly critical, but I don't remember them trying. Ever. They always talk a big game, but have never actually delivered on a single GPU design worth having. I think that if they were serious and actually trying, they would have had at least one competitive product in the past 40 years.

To be fair, they absolutely dominate the entry-level market with IGP.

In volume sure, but there is a reason they went with Vega for some of their higher end chips. If you want IGP AMD is far better.
 

Kenmitch

Diamond Member
Oct 10, 1999
8,505
2,248
136
....Currently the Intel bar is pretty low when it comes to a GPU at least so a big bang could be not so big in the end.
 

dahorns

Senior member
Sep 13, 2013
550
83
91
In volume sure, but there is a reason they went with Vega for some of their higher end chips. If you want IGP AMD is far better.

Actually, we see from kabylake g that you arent entirely correct. For gaming, AMD's solution is better. But for a variety of general use tasks, Intel's product is actually superior.
 

realibrad

Lifer
Oct 18, 2013
12,337
898
126
Actually, we see from kabylake g that you arent entirely correct. For gaming, AMD's solution is better. But for a variety of general use tasks, Intel's product is actually superior.

superior in more than power draw?