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

[jpiniero]

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

Now yields are going to make monolithic a thing of the past. Even on mobile, even if it's only something like one CPU and one GPU chiplet max.

 

[lobz]

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.

I have no idea which office workers you have ever talked with who need more frequency or IPC, but the ones I know are actually frustrated as hell when their systems get real slow when they have many apps + tabs opened. No amount of IPC helps there, just threads and memory.

 

[jpiniero]

I have no idea which office workers you have ever talked with who need more frequency or IPC, but the ones I know are actually frustrated as hell when their systems get real slow when they have many apps + tabs opened. No amount of IPC helps there, just threads and memory.

Probably more platter drives and lack of memory more than anything else. Thank goodness it seems that SSDs have become common even on office desktops.

 

[lobz]

I have no idea which office workers you have ever talked with who needs more frequency or IPC, but the ones I know are actually frustrated as hell when their systems get real slow when they have many apps + tabs opened. No amount of IPC helps there, just threads and memory.

Probably more platter drives and lack of memory more than anything else. Thank goodness it seems that SSDs have become common even on office desktops.

I still don't know how more frequency and IPC would help the office workers as opposed to have enough threads to accomodate their rather unorganized and cluttered workspace.

 

[Charlie22911]

I have no idea which office workers you have ever talked with who need more frequency or IPC, but the ones I know are actually frustrated as hell when their systems get real slow when they have many apps + tabs opened. No amount of IPC helps there, just threads and memory.

Memory and threads do almost jack for a business class machine (the majority of PCs) made in the last 5 years, especially when you have to wait for the 15-20 applications and startup scripts that are a part of a baseline to start simultaneously. An SSD matters more in an office environment that is bogged down by redundant bloatware, HBSS/HIDS, custom patching software, organizational banners, scripts, etc etc etc. Let’s not even mention the experience if background updates are installing.
Your average office person is just not going to benefit much if at all from anything over 4C8T threads for office suites and email as much as they will from an SSD. 4C8T is probably even overkill.
In fact the only time I’ve ever saturated the dual core 5th gen i5 in my ancient work laptop was in excel, thanks to a spreadsheet that triggered a bug with conditional formatting.

Consider a $10 difference between HDDs and SSDs isn’t much, unless you’re an organization like mine that has a few hundred thousand machines. So hard drives are the default choice, which impacts performance massively; I suspect it’s the same with other big organizations as well. This is why I question the decision to go “moar coars”. Something something diminishing returns.

 

[scannall]

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.

As long as security is foremost in the engineers minds then it should be OK.

 

[DrMrLordX]

Now yields are going to make monolithic a thing of the past. Even on mobile, even if it's only something like one CPU and one GPU chiplet max.

Maybe, maybe not. Monolithic mobile CPUs are rolling out on TSMC 7nm+. They haven't quite hit the wall. Should be interesting to see what Apple and Qualcomm do with TSMC 5nm though.

 

[lobz]

Memory and threads do almost jack for a business class machine (the majority of PCs) made in the last 5 years, especially when you have to wait for the 15-20 applications and startup scripts that are a part of a baseline to start simultaneously. An SSD matters more in an office environment that is bogged down by redundant bloatware, HBSS/HIDS, custom patching software, organizational banners, scripts, etc etc etc. Let’s not even mention the experience if background updates are installing.
Your average office person is just not going to benefit much if at all from anything over 4C8T threads for office suites and email as much as they will from an SSD. 4C8T is probably even overkill.
In fact the only time I’ve ever saturated the dual core 5th gen i5 in my ancient work laptop was in excel, thanks to a spreadsheet that triggered a bug with conditional formatting.

Consider a $10 difference between HDDs and SSDs isn’t much, unless you’re an organization like mine that has a few hundred thousand machines. So hard drives are the default choice, which impacts performance massively; I suspect it’s the same with other big organizations as well. This is why I question the decision to go “moar coars”. Something something diminishing returns.

Geez...... how did you guys turn this into a SSD vs threads conversation? Can you please read the original post I replied to?

 

[IntelUser2000]

Now yields are going to make monolithic a thing of the past. Even on mobile, even if it's only something like one CPU and one GPU chiplet max.

They won't. I think you and many others are extreme on yields being low. I'm even going to say this for Intel's case. People admitted that if anyone could pull off the original timeline and specs for the 10nm, it would have been Intel's fab team. But that's how unrealistically they were aiming for.

On Desktops adding a few watt or two, no one cares. Do that on mobile, you'll quickly recede marketshare, and that's for laptops nevermind phones and tablets!

I can see it splitting with mobiles and client chips going monolithic and K chips and servers going MCM. Servers are already going crazy with wafer scale chips and multiple 500mm2+ die packages.

Should be interesting to see what Apple and Qualcomm do with TSMC 5nm though.

Since their 5nm is a quasi-full node, its less of a jump than 14 to 10 or 10 to 7 for Intel. Area reduction seems good, but performance gains are low. It's similar to their 28nm to 20nm transition. This is why I think some compare their 3nm to Intel's 7nm. TSMC can be said to do continuous moderate jumps while Intel does a large jump then supplements it with modifications consisting of + designations at the end. Intel's 7nm+ might be better compared with TSMC 3nm, but each will likely have pluses and minuses that fit the market they are in.

 

[lobz]

They won't. I think you and many others are extreme on yields being low. I'm even going to say this for Intel's case. People admitted that if anyone could pull off the original timeline and specs for the 10nm, it would have been Intel's fab team. But that's how unrealistically they were aiming for.

On Desktops adding a few watt or two, no one cares. Do that on mobile, you'll quickly recede marketshare, and that's for laptops nevermind phones and tablets!

I can see it splitting with mobiles and client chips going monolithic and K chips and servers going MCM. Servers are already going crazy with wafer scale chips and multiple 500mm2+ die packages.



Since their 5nm is a quasi-full node, its less of a jump than 14 to 10 or 10 to 7 for Intel. Area reduction seems good, but performance gains are low. It's similar to their 28nm to 20nm transition. This is why I think some compare their 3nm to Intel's 7nm. TSMC can be said to do continuous moderate jumps while Intel does a large jump then supplements it with modifications consisting of + designations at the end. Intel's 7nm+ might be better compared with TSMC 3nm, but each will likely have pluses and minuses that fit the market they are in.

3nm will come earlier anyway

 

[JasonLD]

3nm will come earlier anyway

Hmm, doubtful. I don't expect AMD cpus will use 3nm until 2023 at the earliest. Samsung's "3nm"(Density wise, it Is comparable with TSMC 5nm and Intel 7nm) could be interesting alternative since it uses GAAFET (Potential performance advantage).

 

[lobz]

Hmm, doubtful. I don't expect AMD cpus will use 3nm until 2023 at the earliest. Samsung's "3nm"(Density wise, it Is comparable with TSMC 5nm and Intel 7nm) could be interesting alternative since it uses GAAFET (Potential performance advantage).

I meant the node itself, TSCM starts volume production in 2022. Consumer products may come at the same time as intel 7nm+ consumer products

 

[RetroZombie]

Now yields are going to make monolithic a thing of the past. Even on mobile, even if it's only something like one CPU and one GPU chiplet max.

They can do even better, how about something like the current Raven Ridge but with one chiplet port for the ability to add more cpu cores for ultra high end mobile and desktop APUs.

 

[Charlie22911]

Geez...... how did you guys turn this into a SSD vs threads conversation? Can you please read the original post I replied to?

That’s wasn’t my intent, what I’m trying to figure out for my own edification is if this is progress for progress’s sake or if there is genuine utility in having so many more cores in the mainstream.

 

[jpiniero]

They won't. I think you and many others are extreme on yields being low. I'm even going to say this for Intel's case. People admitted that if anyone could pull off the original timeline and specs for the 10nm, it would have been Intel's fab team. But that's how unrealistically they were aiming for.

Intel is going to have to ship products at bad yields past 14 nm. And they can't keep increasing the die size at 14nm obviously, they need the power savings of lower nodes.

Lakefield is more like the future. Whether you will see more than say one CPU/GPU/etc chiplet per die remains to be seen.

 

[lobz]

That’s wasn’t my intent, what I’m trying to figure out for my own edification is if this is progress for progress’s sake or if there is genuine utility in having so many more cores in the mainstream.

Ah, sorry for lashing out. We were just talking about different things

 

[Magic Carpet]

Your average office person is just not going to benefit much if at all from anything over 4C8T threads for office suites and email as much as they will from an SSD. 4C8T is probably even overkill.

Maybe 5 years ago it was, not anymore as the apps constantly get re-written. Windows 10 and Google Chrome alone feel much more responsive / start-up faster on higher-thread cpus during your usual multitasking/multitabbing. So yeah, 8T is the norm today, no wonder Intel's next years i3 will have 8T. Good times ahead, time to beat the phones!

 

[Kenmitch]

Consider a $10 difference between HDDs and SSDs isn’t much, unless you’re an organization like mine that has a few hundred thousand machines. So hard drives are the default choice, which impacts performance massively; I suspect it’s the same with other big organizations as well. This is why I question the decision to go “moar coars”. Something something diminishing returns.

Just the thought of using a HDD makes me cringe!

 

[Magic Carpet]

Just the thought of using a HDD makes me cringe!

Combined with SSDs, hdds are great when you need lots of cheap storage fast. Newer ones aren't that noisy or power hungry as your old 10/15K cheetah

 

[lobz]

Combined with SSDs, hdds are great when you need lots of cheap storage fast. Newer ones aren't that noisy or power hungry as your old 10/15K cheetah

Of course they are! Heck, I bet he uses HDDs He must have meant using a HDD as system and main workflow drive

 

[soresu]

Combined with SSDs, hdds are great when you need lots of cheap storage fast. Newer ones aren't that noisy or power hungry as your old 10/15K cheetah

That was absolutely true once - but the long delay in maturity of HAMR/MAMR technology, coupled with the tsunami that impacted manufacturing capacity has stifled price drops and capacity improvement in the HDD market.

I remember getting my first 1 TB HDD in early 2008, thinking they would have 100x the capacity by now - sadly this isn't the case.

I think we will get 8 terabit/1 terabyte per die of NAND flash before we get 100 TB HDD's - which could give 1 petabyte 2.5 inch SSD's in Samsung's enterprise 32x32 hi stack configuration (for reference sake, the entirety of the digital resources/assets for the Jungle Book remake took up 2 petabytes).

10/15K HDD's were also made redundant by the arrival of NVMe SSD's (or even SATA SSD's) - their only point was speed which is far eclipsed by SSD tech now in the multi GB/sec range.

 

[moinmoin]

If anything, the shortages caused by increasing the core count has caused more damage to Intel than AMD directly has in the client market.

Plain mismanagement on Intel's part. They managed to add cores. They added F SKUs with their iGPUs fused off. They somehow don't manage to make smaller dies without iGPUs to begin with which would recover much of the die area spent on additional cores.

 

[uzzi38]

Now yields are going to make monolithic a thing of the past. Even on mobile, even if it's only something like one CPU and one GPU chiplet max.

TSMC 7nm yields are absolutely fantastic - better than 16nm/12nm.

Maybe later on in the future, but for now it's good.

 

[Ajay]

TSMC 7nm yields are absolutely fantastic - better than 16nm/12nm.

Maybe later on in the future, but for now it's good.

What are the yields? What are they for AMD chiplets and SoCs? How about large ASICS? Has TSMC released any data?

 

[coercitiv]

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.

This is my take on it in a nutshell. I didn't post here since I considered it OT for this thread, but time has it's ways.