Article "AMD CTO Mark Papermaster: More Cores Coming in the 'Era of a Slowed Moore's Law'" - @ Tom's

Page 2 - Seeking answers? Join the AnandTech community: where nearly half-a-million members share solutions and discuss the latest tech.

A///

Diamond Member
Feb 24, 2017
4,352
3,151
136
Intel's problems are systemic, they won't get fixed through the surgical intervention of an engineering prodigy.

People who still expect "Conroe" or "Keller" moments from Intel still do not understand what put this company in the position it finds itself today.
Conroe and the long Core uarch was a result of taking some great, great ideas from earlier Intel processors. I said this a week ago, but there are a lot more poor uarchs compared to good ones. And just because Keller is there, putting aside his specialty, it doesn't mean the idea will execute perfectly. Even the greatest engineers in history have messed up. Assuming a 3 year cadence for "Ocean Cove," we'll probably see an Intel response deep into '22. I'm not placing too much faith in Intel to do it right. Immense sums of money and a huge talent pool mean nothing and do not guarantee against a failure. "Ocean Cove" could be amazing, but it could also be trampled by whatever Zen is out then.

Tis the beauty of the hardware sector.
 

CentroX

Senior member
Apr 3, 2016
351
152
116
If AMDs IPC claim is accurate; they are roughly 30% better when zen 4 launches in 2 years. If intel hasnt responded by then, yikes.

That is massive.
 

moinmoin

Diamond Member
Jun 1, 2017
4,731
7,249
136
Intel doesn't have a uarch problem, they have a fab problem. They need to decide - and quickly - if they are capable of keeping up with TSMC or even Samsung, and if not, they need to at the very least start dual sourcing.
They'll want to postpone publishing such fab related decisions as long as possible. The worst case would be once it should become publicly clear that even 7nm won't fix anything, all the current well oiled business will break apart rather quickly. Better keep the current business and keep pretending until the latter may no longer be necessary.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Intel doesn't have a uarch problem, they have a fab problem. They need to decide - and quickly - if they are capable of keeping up with TSMC or even Samsung, and if not, they need to at the very least start dual sourcing.
If they decide for dual sourcing, that will be everything _but_ fast. While TSMC and Samsung respective nodes are fairliy interchangeable, I don't think that it is that easy to port their performance products that were designed for intel nodes, to foundry nodes. I might be wrong tho
 

jpiniero

Lifer
Oct 1, 2010
14,173
4,964
136
If they decide for dual sourcing, that will be everything _but_ fast. While TSMC and Samsung respective nodes are fairliy interchangeable, I don't think that it is that easy to port their performance products that were designed for intel nodes, to foundry nodes. I might be wrong tho

It would have to be a design that isn't tied to a node. Which if the rumors are true that Willow Cove is being backported to 14 nm with Rocket Lake they might be okay there. But we'll have to see if that happens. Just doing that must have been really hard from a internal political standpoint, fabbing a leading edge type product at either foundry would get extreme pushback. But given the lead times you can't wait that long and it does mean committing a lot of money and resources.
 

moinmoin

Diamond Member
Jun 1, 2017
4,731
7,249
136
If they decide for dual sourcing, that will be everything _but_ fast. While TSMC and Samsung respective nodes are fairliy interchangeable, I don't think that it is that easy to port their performance products that were designed for intel nodes, to foundry nodes. I might be wrong tho
Intel doesn't even manage to port their designs between their own nodes. That's on their todo for future designs. :tearsofjoy:
 

scannall

Golden Member
Jan 1, 2012
1,943
1,629
136
If they decide for dual sourcing, that will be everything _but_ fast. While TSMC and Samsung respective nodes are fairliy interchangeable, I don't think that it is that easy to port their performance products that were designed for intel nodes, to foundry nodes. I might be wrong tho
They use their foundries for a lot of other things besides CPUs. Chipsets, ethernet chips etc. That would be much easier to fab elsewhere, and relieve that capacity for their main business.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
It would have to be a design that isn't tied to a node. Which if the rumors are true that Willow Cove is being backported to 14 nm with Rocket Lake they might be okay there.
Well that's quite clear there, that they're doing that out of absolute necessity, not because it's optimal.
 

Zstream

Diamond Member
Oct 24, 2005
3,396
277
136
Can we get back on topic? Intel is just bridging the stupidity of the stock market, where crappy apps and software vendors are worth “billions”. The fact is that Intel is doing quite well, and will continue to do so. Only some short sided moron will say otherwise and try to downplay their results. Intel actually makes something, and have been making quality products for some time.
Once this silly stock market “app bubble” goes away, you’ll have Intel still standing. Let’s talk about more cores and how difficult it is to generate proper software that can span cores without relying on a true hardware scheduler.
 

Arkaign

Lifer
Oct 27, 2006
20,736
1,377
126
One bizarre thing about Intel is that they're packing absolutely massive iGPU into their consumer line from end to end, even when it makes no sense. 9700k/9900k/etc could be 12C with more cache if they used the transistor budget in a more intelligent way with no iGPU.

Even the F series just has the iGPU present but laser disabled. Waste of space.

It makes sense for the mass market corporate boxes though. Meh. I still think it should be chipset side, but I think the iGPU was a cynical move to market block AMD and Nvidia from selling premium mobo-resident onboard GPUs. Whether it is on the board or the CPU itself, they're all extremely mediocre anyway. Ancient 750ti is better than the lot of them, which says a lot about the practical limits of on-package/die/Mobo GPUs. Bandwidth, transistor count, and TDP dictate their mediocre spot in the real world. Good enough for MS word and YouTube, but beyond that a big fat meh.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
Can we get back on topic? Intel is just bridging the stupidity of the stock market, where crappy apps and software vendors are worth “billions”. The fact is that Intel is doing quite well, and will continue to do so. Only some short sided moron will say otherwise and try to downplay their results. Intel actually makes something, and have been making quality products for some time.
Once this silly stock market “app bubble” goes away, you’ll have Intel still standing. Let’s talk about more cores and how difficult it is to generate proper software that can span cores without relying on a true hardware scheduler.
I dare you to hold your breath tight until your prophecy comes true.
 

DrMrLordX

Lifer
Apr 27, 2000
21,310
10,495
136
Good point, right now amd with TSMC 7nm have designs that peak at 4.7Ghz.

Another thing to consider is that TSMC claims their 7nm+ node (Zen3) has an increase in performance at isopower. Zen2 does not seem to be Fmax-limited by design. While it's not at all reasonable to assume that Zen3 will be able to boost 20% higher than Zen2 just from the move to TSMC 7nm+, it is reasonable to assume that boost clocks will go up at least a little, which is the opposite direction Intel is going with their 10nm node.

Cell was more like a single full POWER ISA based core coupled with 7-8 128 bit SIMD engines called SPE's - kinda like a very early proto APU, I'm pretty sure I saw an article claiming some games used the SPE's for extra graphics oomph.

Originally, Sony had wanted to use Cell alone for CPU and GPU duties, until the SPEs proved incapable of handling graphics on their own. That's when they shoehorned in a dGPU for PS3 and called it a day.
 

Roland00Address

Platinum Member
Dec 17, 2008
2,196
260
126
People who still expect "Conroe" or "Keller" moments from Intel still do not understand what put this company in the position it finds itself today.
Care to elaborate, I have an idea of why Intel is in the place it is today, but I am not a mind reader so I can't instantly know what you are referring to for intel's current situation is not one cause, but multifactorial.
 

moinmoin

Diamond Member
Jun 1, 2017
4,731
7,249
136
One bizarre thing about Intel is that they're packing absolutely massive iGPU into their consumer line from end to end, even when it makes no sense. 9700k/9900k/etc could be 12C with more cache if they used the transistor budget in a more intelligent way with no iGPU.

Even the F series just has the iGPU present but laser disabled. Waste of space.

It makes sense for the mass market corporate boxes though. Meh. I still think it should be chipset side, but I think the iGPU was a cynical move to market block AMD and Nvidia from selling premium mobo-resident onboard GPUs. Whether it is on the board or the CPU itself, they're all extremely mediocre anyway. Ancient 750ti is better than the lot of them, which says a lot about the practical limits of on-package/die/Mobo GPUs. Bandwidth, transistor count, and TDP dictate their mediocre spot in the real world. Good enough for MS word and YouTube, but beyond that a big fat meh.
The market for the three big CPU & GPU players in the x86 space is approaching an interesting if not funny situation. Nvidia dominates the GPU space, with monster dies at the very top end (GV100 is 815 mm²). Intel dominated the CPU space for long, and they as well use big dies at the top end (the 28 cores Skylake-X is 698mm²). The latter has been attacked by AMD's MCM/chiplets approach used for CPUs, but their GPUs are still monolithic. Through Xe, Intel plans to essentially attack Nvidia by using the MCM/chiplets approach in the GPU space.
 

soresu

Platinum Member
Dec 19, 2014
2,371
1,583
136
Intel doesn't even manage to port their designs between their own nodes. That's on their todo for future designs. :tearsofjoy:
I think 28nm was AMD's wake up call there - it taught them to be careful over committing designs to nodes.

I think it happened before that too, when 32nm got cancelled forcing most of Northern Islands to be cancelled along with it, excepting Cayman which got ported back to 40nm?
 
Last edited:

jamescox

Senior member
Nov 11, 2009
635
1,103
136
Intel doesn't have a uarch problem, they have a fab problem. They need to decide - and quickly - if they are capable of keeping up with TSMC or even Samsung, and if not, they need to at the very least start dual sourcing.

It is kind of both. My assumption is that all of Intel’s designs were monolithic until AMD shook up the market. The lead time for a completely new design is very long, so I expect that their cpus, regardless of process tech, will be monolithic up until 2022. They are leaking all kinds of stuff to try to keep people interested, but they just don’t have competition for quite a few AMD parts. They have talked about their chip stacking tech, which was probably started quite a while ago, but that was probably for very low power mobile parts. Chip stacking is the norm in cell phone processors and such.

The shortage of 14 nm part indicates a lot of accelerated, cancelled, and changed plans inside intel. I kind of wonder if it is partially due to suddenly needing to make a huge number of 8 core parts on 14 nm when they never intended to do so. We should have had 8 core parts for the high end on 20 nm. AMD (technically) did it on 28 nm with excavator based parts, but they had shared FPUs. Integer ALUs are tiny, but FPUs are comparatively huge. Intel just kept on selling 4 core parts because they could, even though 8 core should have been mainstream at 14 nm. Now they have to make all of these 8 core parts that would have been high priced Xeons or HEDT parts before.

When they finally get a 10 nm or below part out, it is probably going to be a monolithic 10 core part. That isn’t going to compete well with AMD’s chiplet based parts that will be much cheaper to produce. Also, I think intel will have a very tough time competing with their own 14 nm parts. They will probably have lower performance for gaming compared to high clocked 14 nm parts. I find the gaming performance thing kind of ridiculous anyway, since it is mostly at low quality and at ridiculously high FPS for all processors. It would be bad PR to lose to their own older cpus and possibly to AMD’s parts , even if it isn’t a particularly valid use case. The rumored back port of a 10 nm design to 14 nm ++(+?) May allow them to compete in the desktop market with Zen 2, although Zen 3 may change that. They will take a lot more power even if they win some performance benchmarks. They may also have trouble meeting demand due to the large die size on 14 nm. If it was planned for 10 nm, then it will be huge and power hungry on 14 nm variant. They could get closer feature parity by supporting pci-e 4.0 though.

The server landscape is even worse, although intel has the advantage that a lot of systems are planned significantly in advance, so even if AMD is better in almost every way, they will continue to sell server processors for years. The server market is very conservative. They will not switch easily. Intel isn’t going to be competitive with their monolithic designs and a design actually optimized for a chiplet architecture will take years of work. They tried with the Xeon platinum 9200, but that is a marketing stunt really. No one will buy that part. It is 2 almost 700 square mm die, so it will be ridiculously expensive. Even if it wasn’t ridiculously expensive, it is still more than 2x the power of a competing AMD part which doesn’t work for big server installations.

One thing to keep in mind is that TSMC has good reason to support AMD strongly. Intel has their own fab so they are technically competition to TSMC. TSMC can’t get a part of the massive AMD64 server market except by supporting AMD or through ARM processor overtaking the AMD64 market. They probably get a small part for the revenue through accessory chips, but they would get that anyway. Intel had, I believe, close to 99% of the server market share. With AMD now taking a chunk of that, TSMC also gets a part of that. Without AMD they were almost completely locked out. There has been some move by big vendors to just make their own chips, so ARM could be a big factor also going forward. I don’t know how the current ARM server chips compare to Epyc though.

So I think intel does have a uarch problem, depending on what you consider part of the uarch. While their mesh network L3 cache does provide a monolithic last level cache, it seems quite expensive to do so. The power consumption on chips with the mesh network were higher than expected and higher than Intel’s ratings would indicate. The mesh network may not be that worthwhile for 8 cores or less and it can’t work across multiple chiplets. The power consumption would go through the roof unless they are split off into separate NUMa nodes, which means it isn’t a monolithic last level cache any more. While intel will have a response eventually, it is actually unclear whether they can pull ahead. They aren’t going to have the process tech advantage anymore like they did compared to AMD’s fab. TSMC represents a massive portion of the entire market, so they have the resources to push the process tech. As far as design, it is also unclear whether they can pull ahead. We now have many companies with the expertise to design high performance cpus.

A lot of customers also realize how bad it is to have a single supplier. The cloud computing companies took a big hit with spectre, meltdown, etc. They will want to have some diversity such that their entire business isn’t shutdown by some vulnerability from one supplier. That is good for AMD, but may also be good for ARM based solutions. The next few years will be interesting. As a customer, I hope intel and nvidia get some good competition. It isn’t in the consumers best interest to have a single dominant company.
 

jpiniero

Lifer
Oct 1, 2010
14,173
4,964
136
Intel just kept on selling 4 core parts because they could, even though 8 core should have been mainstream at 14 nm

Where would they have gotten the fab capacity? If anything, the shortages caused by increasing the core count has caused more damage to Intel than AMD directly has in the client market. And now OEMs are mad because Intel looks like it has cut Comet Lake production in favor of Whiskey but the OEMs want Comet because it's newer.
 

Charlie22911

Senior member
Mar 19, 2005
613
228
116
More cores doesn’t do much for you if the software can’t be parallelized to use them, so I don’t understand the push. Sure for us enthusiasts who think this stuff is cool, we may be able to use it. We are the minority though.
If the goal is to push performance forward, then there are other areas we can focus on, storage for example; every PC within my organization uses spinning rust. It’s horrid, and these types of machines represent the majority of the PCs that are manufactured and sold.

Now if the mythical beast known as “reverse hyper threading” could be done, things would be very interesting.
 

soresu

Platinum Member
Dec 19, 2014
2,371
1,583
136
Now if the mythical beast known as “reverse hyper threading” could be done, things would be very interesting.
Speculative Multi Threading is its proper name.

I have wondered more recently if AI/ML research might prove to forge a new path for Spec MT, but nothing concrete has come about yet - though I don't read every academic paper, it could have been done already.
 

Charlie22911

Senior member
Mar 19, 2005
613
228
116
Speculative Multi Threading is its proper name.

I have wondered more recently if AI/ML research might prove to forge a new path for Spec MT, but nothing concrete has come about yet - though I don't read every academic paper, it could have been done already.

I wasn’t aware, thanks. I’ll have some reading material to look at now.
It’s not uncommon to see blocks for AI type workloads in arm SOCs. It’d be interesting to See it more tightly integrated into the execution cores to extract more parallelism, though I wonder if latency would negate many of the benefits.

Or I may just be thinking out of my bum, I’m an enthusiast; I don’t know squat about such low level things.
 

Cardyak

Member
Sep 12, 2018
72
159
106
Speculative Multi Threading is its proper name.

I have wondered more recently if AI/ML research might prove to forge a new path for Spec MT, but nothing concrete has come about yet - though I don't read every academic paper, it could have been done already.

I do think speculative multithreading will happen one day, and when it does it will offer a new way to extract more parallelism and increase performance for single threaded linear code.

Akin to how OoO execution offered a long runaway for future performance gains over the last 20 or so years, speculative execution will hopefully offer the same for the next 20 years.

The pertinent questions are: How do you identify the parallelism to be extracted? Along with power and latency.
 

lobz

Platinum Member
Feb 10, 2017
2,057
2,856
136
More cores doesn’t do much for you if the software can’t be parallelized to use them, so I don’t understand the push. Sure for us enthusiasts who think this stuff is cool, we may be able to use it. We are the minority though.
If the goal is to push performance forward, then there are other areas we can focus on, storage for example; every PC within my organization uses spinning rust. It’s horrid, and these types of machines represent the majority of the PCs that are manufactured and sold.

Now if the mythical beast known as “reverse hyper threading” could be done, things would be very interesting.
This kind of thinking was exactly what intel was hoping to achieve among consumers during the quad-core era. Don't forget, originally they wanted to launch even the SKL-X HEDT with another 10 cores at the top-end.

You also forget the cheekiest catch: innovating and developing to further increase IPC is not going against offering more cores and overall compute resources at the same time. This is another mindst intel achieved very successfully: avg users didn't long for more cores / threads because they were made to believe they didn't need them, and then everyone was amazed when the HEDT went from 6 to 8 and then 8 to 10 cores. Such engineering marvel...

Most people started thinking only when AMD started to offer mainstream 8 cores on an inferior and less dense node then intel's current node.

100% agree on the storage front with you though. I'd rather hang myself than having to work on a PC with a HDD in it ever again.

But that has nothing to do with not holding back the otherwise inevitable HW and SW evolution artifically by IHVs, that's simply a rather unintelligent way of trying to cut costs by SIs. They always underestimate the value of customer satisfaction.
 
Last edited:

jpiniero

Lifer
Oct 1, 2010
14,173
4,964
136
You also forget the cheekiest catch: innovating and developing to further increase IPC is not going against offering more cores and overall compute resources at the same time.

It is a tradeoff, since you have die space, power consumption, costs among other considerations.

And I'm telling you, the majority of the desktop market is office workers who don't need and won't use extra cores. They can use more frequency and/or IPC however.
 

IntelUser2000

Elite Member
Oct 14, 2003
8,686
3,785
136
Chip stacking is the norm in cell phone processors and such.

That's very different. It's package on package with memory. Older Atom chips that went into phones and tablets used the same thing(because it had to in order to be competitive).

Mobile will continue to be monolithic or close to it because it has to.

MCM-approach is for desktops and servers.