Question Rethinking AMD consumer CPU lineup in the situation of chiplet scarcity

[Kocicak]

If we look at the higher end of the consumer CPU lineup by AMD (beginning with Ryzen 3600), and the fact that AMD is unable to deliver a highly clocked 6+6 core CPU and effectively offers no high core count CPU at this point, and the fact that the high quality chiplets are better used in server CPUs, I believe it is time to radically rethink the lineup of consumer CPUs.

I believe the higher end of the consumer CPU lineup can comprise of just 4-5 processors.

1) 6C 3.9/4.2 65W - the same as current 3600 - price 200 USD
2) 6+6C 3.6/4.2 95W - made from two chiplets which are used in 6C - price 300 USD

These two CPUs can cover 75-90% of the market demand, they are made from low quality chiplets which can have up to two nonfunctional cores. These chiplets are unusable in server processors anyway. They are essentially a waste product of server CPU production.

The market demand can be nearly all covered by chiplets that are not needed for server CPU production !!!

3) 8C 4.0/4.7 95W - very high quality chiplet which could be used in server CPU - price 400-450 USD
4) 8+8C 3.8/4.7 135W - made from two chiplets which are used in 8C - price 750-800 USD

5) 8+8C-super ?/? ?W - made from "miraculous chiplets", intended for those who want something extraordinary and special - price 1000 - 1200 USD. Existence of this product depends on the existence and quantity of those "miraculous chiplets".

What do you think?

 

[Topweasel]

I'm sure, by this time next year, he will be able to buy 3900 for $300.

Don't tell yourself it can't be done. Ask yourself how can it be done.

It's been mentioned before to him by several people, at least once by myself.

 

[Thunder 57]

He wants a 90w 3900 and sold for as little as AMD can sell it for ($300). Pipedream. Its insane the mental hopes he is going through to sell himself that it needs to happen and that AMD so obviously sees that it needs to happen...

Now now now, let's get it right. A 105W 3900 for $249. What a joke! And apparently we're all shareholders. Heck, I wish I got some when it was $2-3 a share.

 

[amrnuke]

Now now now, let's get it right. A 105W 3900 for $249. What a joke! And apparently we're all shareholders. Heck, I wish I got some when it was $2-3 a share.

No, come on! He wants a 65W 3600x2 12C24T with an all-core boost of 4.8GHz with an in-package AIO cooler for $99.98. If AMD are just using their lowest-binned chiplets it won't really cost them anything!!!

 

[Thunder 57]

No, come on! He wants a 65W 3600x2 12C24T with an all-core boost of 4.8GHz with an in-package AIO cooler for $99.98. If AMD are just using their lowest-binned chiplets it won't really cost them anything!!!

I'm glad I wasn't drinking anything when I read this because I probably would've laughed it onto my monitor! Well done.

 

[Kocicak]

I just realised that 3600x2 would need to come with the best cooler, while 3600 comes with the worst cooler. It may be necessary to raise my suggested price $269 a bit to be able to add a better cooler. Personally I do not need a disco-cooler, plain old "dark cooler" is perfectly fine IMO.

What about $289?

Some time passed after I posted this thread and I hope that many more of you realised, what a perfect and simple product strategy this is:

1) Make one processor with low quality slow incomplete chiplet (=cheap) and make the other one by adding one more such chiplet.
2) Serve large part of the consumer market for the lowest platform (AM4) with these two CPUs made of "scrap" and sell them cheap.
3) Make third CPU with high quality fast complete chiplet for those people who need speed and make the fourth one by adding one more such chiplet for those who want the absolute best. Make people pay for this high quality stuff.

EASY. SIMPLE. LOGICAL. WONDERFULL.

Use your brains before posting in this thread, ok? This is for smart people only.



Insults are not allowed.


esquared
Anandtech Forum Director

 

[DrMrLordX]

EASY. SIMPLE. LOGICAL. WONDERFULL.

No. Please stop.

 

[NostaSeronx]

You don't understand @Kocicak

Ryzen is about rising prices, rising capabilities, rising demand!

Rather than talk about this cheapo garbo Ryzen 3600x2, you should think about the Ryzen 5890X on AM6 with four 12-core dies and a quad-channel DDR5/DDR6 io die. That can execute multiple AVX1024 instructions per cycle, with no idle dead area even with AVX128, AVX256, AVX512. At a simple price of $1999, only real Ryzen'ers would want that.

Tell AMD to bring out the real guns with real value than this chicken garbo. There isn't enough protein in these processors and they are being weighted down by my gains yo. We've been promised 10 GHz and ultra-super-duper-wide cores, enough is enough vote with your wallets my bro-ananders.

 

[TheGiant]

I just realised that 3600x2 would need to come with the best cooler, while 3600 comes with the worst cooler. It may be necessary to raise my suggested price $269 a bit to be able to add a better cooler. Personally I do not need a disco-cooler, plain old "dark cooler" is perfectly fine IMO.

What about $289?

Some time passed after I posted this thread and I hope that many more of you realised, what a perfect and simple product strategy this is:

1) Make one processor with low quality slow incomplete chiplet (=cheap) and make the other one by adding one more such chiplet.
2) Serve large part of the consumer market for the lowest platform (AM4) with these two CPUs made of "scrap" and sell them cheap.
3) Make third CPU with high quality fast complete chiplet for those people who need speed and make the fourth one by adding one more such chiplet for those who want the absolute best. Make people pay for this high quality stuff.

EASY. SIMPLE. LOGICAL. WONDERFULL.

Use your brains before posting in this thread, ok? This is for smart people only.

I get your point that AMD didnt make a proper job in products, but you are on the wrong forum- here you enter the excel formula =if(A1>1000;if(A2="Intel";"Piss";If(A2="AMD";"enjoy";"consider"));if(A2="Intel";"Piss";If(A2="AMD";"enjoy";"consider")))

 

[Thunder 57]

Use your brains before posting in this thread, ok? This is for smart people only.

Smart people only, eh? 3600 on Amazon. Which would AMD rather have, $289 or $389?

Considering the recent news about AMD, they would be rather wise not to entertain your idea. Sure it would be nice if the whole lineup was half the price, but these things cost a lot of money to develop and it's very high risk, high reward. If you end up developing a turd like Bulldozer you are screwed for at least five years.

Besides, they are priced very competitively, I don't get your fascination with this pipe dream of yours.

 

[Kocicak]

It would be even possible to make one more CPU between 8C and 8+8C (all HQ), and that would be HQ8 + LQ6. The low quality 6C would activate just when the CPU would need to run heavilly multithreaded load. You would get all the benefits - speed and efficiency of HQ 8C chiplet, while having more multithreaded grunt when needed without having to pay for the second expensive HQ 8C.

So the lineup would then be:

LQ 6C - $200
LQ 6C + LQ 6C - $300
HQ 8C - $400
HQ 8C + LQ 6C - $500
HQ 8C + HQ 8C - $750

 

[Topweasel]

It would be even possible to make one more CPU between 8C and 8+8C (all HQ), and that would be HQ8 + LQ6. The low quality 6C would activate just when the CPU would need to run heavilly multithreaded load. You would get all the benefits - speed and efficiency of HQ 8C chiplet, while having more multithreaded grunt when needed without having to pay for the second expensive HQ 8C.

So the lineup would then be:

LQ 6C - $200
LQ 6C + LQ 6C - $300
HQ 8C - $400
HQ 8C + LQ 6C - $500
HQ 8C + HQ 8C - $750

Several problems with your thought process. 1 There is no such thing as a LQ die. People like to pretend there is a big diffence and one is getting worse dies than others. The real difference in max clocks is going to be in the 100-200Mhz area with a little bit on the outside on both sides of that. Thats where the 5% TR stuff comes from. 90% of the dies that come out are going to be good enough for Epyc. The ones not become the X lineup in Ryzen or Threadripper. The characteristics that max a good ryzen chip are the opposite of what makes a good Epyc chip. There wouldn't be really any 6c capable chips that wouldn't fall into one of those bins. So they would be losing margin and cannibalizing sales of other chips by offering your HQLQ and LQLQ combinations. 2. More than increased IPC and increased clocks. Zen 2 is about bringing uniformity to performance at every part of the scale. This is mostly a server side reason. But the whole idea with the IO die is to make performance quicker and more predictable using multiple core chips a HQLQ design kills that. 3. AMD already xnayed unbalanced CCX's though technically possible. Besides using dies that can't run as fast. I seriously doubt AMD will give difference cor count dies per package a second thought.

You can stop with this its a pipedream. There is room for more products in the stack. It's not going to be the ones you are looking for.

 

[DrMrLordX]

I get your point that AMD didnt make a proper job in products, but you are on the wrong forum- here you enter the excel formula =if(A1>1000;if(A2="Intel";"Piss";If(A2="AMD";"enjoy";"consider"));if(A2="Intel";"Piss";If(A2="AMD";"enjoy";"consider")))

Untrue, and basically off-topic.

It would be even possible to make one more CPU

Not for us, it isn't. Why don't you give all these ideas to AMD instead of telling us about them? You have already discussed your dissatisfaction with AMD's current product lineup and boost behavior at length. What more do you hope to accomplish here?

 

[Kocicak]

Several problems with your thought process. 1 There is no such thing as a LQ die. People like to pretend there is a big diffence and one is getting worse dies than others. The real difference in max clocks is going to be in the 100-200Mhz area with a little bit on the outside on both sides of that. Thats where the 5% TR stuff comes from. 90% of the dies that come out are going to be good enough for Epyc. The ones not become the X lineup in Ryzen or Threadripper. The characteristics that max a good ryzen chip are the opposite of what makes a good Epyc chip. There wouldn't be really any 6c capable chips that wouldn't fall into one of those bins. So they would be losing margin and cannibalizing sales of other chips by offering your HQLQ and LQLQ combinations.

There is 600 MHz difference between Ryzen 3500 and 3950X in specs. I bet anybody who spent some time overclocking will not confirm, that there are no better or worse dies. There are even some companies, which sort CPUs for people, who want to get better quality guaranteed. Have you ever heard about silicon lottery???

Where did you get those percents from? Are you just guessing or speculating?

You know the bins, how full they are, how AMD calculates prices of components (e.g. chiplets according to their various characteristics) and prices of finished products? Do you know sales figures and profit of differents products? You do not.

You just presume, that what I am proposing would not be financially benefitial for AMD. I presume, that it would not hurt them and could be benefitial not only financially, but in other ways, for example simplifying materials logistics, freeing up more HQ materials for other use etc. Your presumptions are not better than mine.

It returns again, I think like a consumer who wants to get the best valuie for his money. I do not think like a AMD shareholder. I am not emotionally attached to them.

2. More than increased IPC and increased clocks. Zen 2 is about bringing uniformity to performance at every part of the scale. This is mostly a server side reason. But the whole idea with the IO die is to make performance quicker and more predictable using multiple core chips a HQLQ design kills that.

You just speculate again. You can yourself tune frequencies of individual cores. That thing is highly configurable. Cores on a single chiplet behave differently. I had two quick cores and one helper core in the 3900X. What would happen, if the cpu by accident let the weak core run fast? It would crash. It knows, how fast each core can run and how to use them.

3. AMD already xnayed unbalanced CCX's though technically possible. Besides using dies that can't run as fast. I seriously doubt AMD will give difference cor count dies per package a second thought.

Umm, do you know, that you can turn cores off? That thing still works with some cores turned off. Miracle.

You can stop with this its a pipedream.

Why? Putting a second chiplet in a 3600 is easy and cheap to do for AMD, and I know I would want such a product as a customer, so why should I stop? I do not need HQ chiplets as those in 3900X. I am perfectly fine with the stuff that is in 3600.

BTW there are some special departments in companies which try to find out, what customers want. Have you by chance ever heard about those???

 

[maddie]

I keep thinking of captain Ahab.

 

[Thunder 57]

...Umm, do you know, that you can turn cores off? That thing still works with some cores turned off. Miracle.

Totally not the same thing. An 8 core CCD would need to be paired with another 8 core CCD. I thought this the place for smart people though? How much longer until this thread gets locked too? You've made your points, others have done the same, in multiple threads now. We have reached an impasse. Please stop.

 

[DarthKyrie]

Untrue, and basically off-topic.



Not for us, it isn't. Why don't you give all these ideas to AMD instead of telling us about them? You have already discussed your dissatisfaction with AMD's current product lineup and boost behavior at length. What more do you hope to accomplish here?

To get people to buy Intel CPUs instead of AMD CPUs.

 

[Topweasel]

There is 600 MHz difference between Ryzen 3500 and 3950X in specs. I bet anybody who spent some time overclocking will not confirm, that there are no better or worse dies. There are even some companies, which sort CPUs for people, who want to get better quality guaranteed. Have you ever heard about silicon lottery???

Where did you get those percents from? Are you just guessing or speculating?

You know the bins, how full they are, how AMD calculates prices of components (e.g. chiplets according to their various characteristics) and prices of finished products? Do you know sales figures and profit of differents products? You do not.

You just presume, that what I am proposing would not be financially benefitial for AMD. I presume, that it would not hurt them and could be benefitial not only financially, but in other ways, for example simplifying materials logistics, freeing up more HQ materials for other use etc. Your presumptions are not better than mine.

It returns again, I think like a consumer who wants to get the best valuie for his money. I do not think like a AMD shareholder. I am not emotionally attached to them.

You obviously don't know how binning works and how it applies to AMD and their dies. 2 Things the 3500 is meant to be the low power cheap Chinese OEM CPU. Can't have that at default settings competing with the 3600 and 3700. That's not the limit of the die. It's the clocks they set to make sure that on both single and multicore workloads the more expensive chip is always better. Manual overclock and it will be within 200MHz of the top chips clocks. But beyond segmentation (learn up on this before you continue because its the big reason your idea's will never see the light of day. Things that make a die ultra power efficient are the same things that hinder clockspeed. Intel can sell their top desktop cpu's with really great clocks, because even for a laptop CPU (which their Desktop CPU's are really designed to be) they make so many that the bad horribly efficient dies (the ones that can clock up to 5 GHz) even at 2-3% are enough to supply the market with 5GHz CPU's. Intels problem is the middle ground chips, not insanely fast still not efficient for mobile (because they have to be really really efficient). There is room in their stack (and the 8400 was a perfect example of this) for "bad dies". AMD doesn't have to be ultra efficient. Just really efficient. They don't have the production numbers to push into the real limits of their chips. Its why on their CPU's and GPU's they are considered undervolting champions. Even though some if not most of their chips can hit the clocks at less power settings they don't have that luxury and tend to send more volts than needed by a lot to make sure they get at least a certain amount of chips at that yield. My Radeon VII uses almost 100 less watts than rated at "auto undervolt" at no measurable performance loss. That's what is happening with the X chips. They accept more power better and is at a trade off where they get the amount of them they want at the performance watermark they set. On the other hand. Unlike Intel with their mobile chips AMD has a lot of power room in their server chips (and with that TR). This means that they don't have to be the best of the best in terms of efficiency. Which means AMD is very very very very unlikely to find any chip "LQ" and what you rate as LQ, actually would sell for twice as much as it does per core on the non-X parts. AMD doesn't really have a 8400 class CPU, not for speed. Maybe the 3500 in terms of cores or L3, but not speed.

ou just speculate again. You can yourself tune frequencies of individual cores. That thing is highly configurable. Cores on a single chiplet behave differently. I had two quick cores and one helper core in the 3900X. What would happen, if the cpu by accident let the weak core run fast? It would crash. It knows, how fast each core can run and how to use them.

It's highly configurable on desktop and AMD's turbo is very highly tuned. But until MS has completely caught up with golden cores this is a non-starter. But even outside PBO and AutoOC. AMD basically has all cores run very similar. But this one you are the closest to possible. AMD would generally be against it just because each of the cores are different and can be clocked seperately we don't have all the secret sauce info on how the IO and CPU works in general and how it will act with two wildly different dies in performance. Long term this is like a Big.little setup and I have a feeling specially with mobile products AMD will have to get that going. But that doesn't mean that Zen 2 as established on TR/Epyc or Ryzen 3k is really comfortable with this.

Umm, do you know, that you can turn cores off? That thing still works with some cores turned off. Miracle.

AMD has been very shy of unbalanced CCX's and has been on CCD's. It potentially causes a performance missmatch that can make the performance unpredictable. Which is specifically what they have been avoiding and what Zen 2 was designed for. We haven't seen it because they aren't going to do it, even if techinically its possible.

Why? Putting a second chiplet in a 3600 is easy and cheap to do for AMD, and I know I would want such a product as a customer, so why should I stop? I do not need HQ chiplets as those in 3900X. I am perfectly fine with the stuff that is in 3600.

BTW there are some special departments in companies which try to find out, what customers want. Have you by chance ever heard about those???

Yeah and they also have departments that make sure their product stack projects their corporate image and other departments about how to offer their CPU's to keep up with demand while maximizing profitability. What you are asking for is comparable to asking Ferrari to make a 488 Pista with a 458 Italia engine and pricing it like that of a Corvette. It's not a realistic thought process, even if in some way technically feasible.

 

[amd6502]

What do guys think of an ultimate type Threatripper where you have one high quality CCD running 16t/8c ~3.8-4.5ghz (base-boost) and a pair of CCDs running 8t/8c (SMT-off) each with more modest frequencies of say 3.8-4.1ghz (using a bit of overvolting relative to the high quality CCD).

It would be like a 24c/32t gaming beast imho. I would think the bin quality for the accessory CCDs would be medium like 3600 ryzen.

 

[Thunder 57]

What do guys think of an ultimate type Threatripper where you have one high quality CCD running 16t/8c ~3.8-4.5ghz (base-boost) and a pair of CCDs running 8t/8c (SMT-off) each with more modest frequencies of say 3.8-4.1ghz (using a bit of overvolting relative to the high quality CCD).

It would be like a 24c/32t gaming beast imho. I would think the bin quality for the accessory CCDs would be medium like 3600 ryzen.

I think that nobody wants a "24 core gaming beast". And why would you disable SMT? I have no idea how you came up with this one to be honest.

 

[TheGiant]

I think that nobody wants a "24 core gaming beast". And why would you disable SMT? I have no idea how you came up with this one to be honest.

wait until TR3 line comes out, you will find soon enough how it is needed

 

[Thunder 57]

wait until TR3 line comes out, you will find soon enough how it is needed

No doubt, I'm just saying that nobody needs a 24 core beast for strictly gaming. If you use it for work and game on the side then great. I doubt AMD would mix CCD's with different max frequencies though. And yes, let's wait and see the real lineup before we speculate about another "that would be great" CPU.

 

[DrMrLordX]

To get people to buy Intel CPUs instead of AMD CPUs.

What? That's unpossible!!!

 

[TheGiant]

Untrue, and basically off-topic.

this is the most true statement here, or can't see the forest for the trees?

To get people to buy Intel CPUs instead of AMD CPUs.

need any more? its not about what, but about who here for a long time

Kocicak has a point and the results of his point and AMDs performance outside the product itself and its value, which is excellect is horrible, if not amateur
as good as they are in technical part, bad is everything else
you can see it in their financial results

 

[DrMrLordX]

this is the most true statement here, or can't see the forest for the trees?

This forum is not AMD-biased. It has spent most of its history being Intel-biased, since Intel has had the faster CPUs throughout that period. Every time AMD pulls ahead, heads start to explode.

Kocicak has a point

Seriously? No, he doesn't. He wants AMD to undercut their own pricing. I don't see him begging Intel to sell downbinned 8c/16t parts for $250 or less to combat the 3700x. He's just fine with their current product lineup. The 9700k @ $360 street price is a perfectly good alternative! Feh.

Where are the "Intel lies about their power usage!!!!" controversy threads? There aren't any.

Where are the "Intel betrayed me, where are my 10nm dekstop CPUS!!!!!" threads? There aren't any.

Where are the "Intel won't unlock all their parts!!!!!" controversy threads? There aren't any.

Where are the "Intel CPUs can't run AVX2 at full clockspeed!!!" threads? Nope, none of those either.

Instead, what are his complaint threads about? AMD charges too much for 12c/24t CPUs on the desktop, aaaaaand they categorize their boost clocks improperly. Give me a break. You want more pro-Intel posting around here? Tell Intel to release Alder Lake-S in Q3 2020. Too bad they can't do it. So long as AMD has the better hardware, you'll see pro-AMD posting, albeit with the usual sniping from people who can't stand seeing AMD do well.

 

[amd6502]

I think that nobody wants a "24 core gaming beast". And why would you disable SMT? I have no idea how you came up with this one to be honest.

1a. to balance out the lack of elite frequencies by avoiding shared resources (L3,L2,L1, decode, execution pipes, and power budget) on the inferior dies.

1b. overvolted (low bin) dies will thermally behave much nicer running SMT-off than running at full load with SMT2. They will need to be overvolted a good bit if one wants to match the top base frequencies between elite and low bin.

2. to still get quite decent perf/watt by leveraging the high quality die for most the threads (for most loads).

3. because 32 threads is more than plenty for most, even among TR4/4+ users.

4. because lower quality bins are far more abundant and cheap than a die from an elite bin.