AMD CPU speculation... and expert conjecture

[8350rocks]

One thing I've not found on any review site is a clear explanation of why 30W (or even 50W) additional TDP gains you so little in terms of moving from the A8 7600 to the A10 7850. The top end Richlands and Trinities were clearly ahead of the lower power SKUs, but this isn't the case for Kaveri. Kaveri's sweet spot definitely seems to be in the 35-45W range.

The long and short answer is...you are seeing the process differences between SOI vs BULK.

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

 

[blackkstar]

One thing I've not found on any review site is a clear explanation of why 30W (or even 50W) additional TDP gains you so little in terms of moving from the A8 7600 to the A10 7850. The top end Richlands and Trinities were clearly ahead of the lower power SKUs, but this isn't the case for Kaveri. Kaveri's sweet spot definitely seems to be in the 35-45W range.

The long and short answer is...you are seeing the process differences between SOI vs BULK.

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Intel users never admit that anything that comes from Intel is a failure. One of my favorite things to do to Intel fanboys is ask them why they aren't using an Intel Itanium on an Intel motherboard with an Intel dGPU. They usually don't have much of an answer.

But I do agree with you based on experience. Intel process is fragile and does not like a ton of voltage. AMD SOI doesn't care. I abused my poor Opty 165 and this FX and it still doesn't care. Every Intel I have owned has not faired so well.

I am really hoping AMD can show up and deliver on HEDT. Intel has been hitting "remove replace rejoice" really hard lately. First they swap TIM, gimp overclocking, then add better TIM and suddenly they are heroes. And now there are rumors they are abandoning FIVR. Another thing they backtracked on that benefits consumers but will be generally taken by users as a blessing.

Intel is really screwing up a lot lately but they have the luxury of not getting any negative press for it because they are still on top. Intel is spending all their time trying to force themselves into markets that are already filled while circle jerking HEDT by adding features that harm HEDT users and then removing them (or removing features and then adding them later).

The time is ripe for AMD to make a big move.

But I don't get why there is no hype for FX-7500. That's a lot of chip for 19w. A4-5000 is 15w and it's a quad at 1.5ghz and 128 GCN cores. FX 7500 is a quad with 2.1ghz base and 3.3ghz turbo with 6 GCN CUs.

I don't remember how many GCN cores are in a CU, but I'm quite sure FX 7500 has more.

 

[de5_Roy]

But I don't get why there is no hype for FX-7500. That's a lot of chip for 19w. A4-5000 is 15w and it's a quad at 1.5ghz and 128 GCN cores. FX 7500 is a quad with 2.1ghz base and 3.3ghz turbo with 6 GCN CUs.

I don't remember how many GCN cores are in a CU, but I'm quite sure FX 7500 has more.

gcn 1.1 has 64 shaders per CU. 6x 64 = 384 shaders. kabini has gcn 1.0 based 2x CU with 128 cores max. iirc gcn 1.1 has more prim/clock than gcn 1.0.
richland and trinity has decent clockrates too, but they couldn't hold those clockrates inside retail laptops with oem cooling. mobile kaveri shouldn't have that issue, but some reviewers found that dt kaveri (7850k) is prone to throttling. it'll come down to how oems design their laptops.

edit: fixed mistake.

 

[juanrga]

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

I still recall people in this forum claiming that Kaveri couldn't pass 3GHz because bulk was the poor thing in the universe. The facts are that Kaveri record is currently set on 6.1GHz. Average OC on air for Trinity, Richland and Kaveri are

A10-5800k: 4.5GHz (32nm FDSOI)
A10-6800k: 4.9GHz (32nm FDSOI)
A10-7850k: 4.5GHz (28nm bulk)

Everyone knew that Kaveri wouldn't break any OC world record, but bulk is much much better than what some claimed.

The reasons why AMD migrated from FDSOI to bulk are five (in no particular order):

- FDSOI is almost dead. Glofo canceled the 28nm FDSOI SHP process.
- bulk opens the foundry choices beyond being locked to glofo.
- bulk provides higher density needed for GCN.
- bulk provides clear migration path to FINFET.
- bulk reduces costs associated to migrating designs from/to dGPU.

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Again facts tell a different history... Average OC on air for Lynnfield, Sandy Bridge, and Ivy Bridge are

i7-880: 4.3GHz (45nm bulk)
i7-2600k: 5.1GHz (32nm bulk)
i7-3770k: 4.7GHz (22nm bulk finfet)

In the first place we can observe that 32nm bulk provides higher frequencies than 32nm FDSOI (see above). Thus bulk is not so bad as some pretend.

The reason why Ivy Bridge overclocks poor than sandy bridge has nothing to do with bulk neither with Intel abandoning HEDT.

The reason was that the introduction of finfets modified the shape of the voltage curve as shown in next figure

Around ~1.32V the 22nm FF curve crosses the 32nm bulk curve. This explains why Ivy Bridge introduced further gains in mobile but failed for overclockers. The interesting stuff is that 14nm FF doesn't have crossing point, which means that Broadwell will overclock much better than both Sandy and Ivy. If everything else goes the same, Broadwell-K would hit ~5.6Ghz on air.

22m FF was an intermediate process. 14nm FF will introduce again classical scaling.

 

[logainofhades]

Intel users never admit that anything that comes from Intel is a failure. .

I am an Intel user and I say crapburst was an epic fail. Sadly, AMD tried to imitate that fail with faildozer. Somewhere along the line AMD lost their way and made the same mistakes they used to bash Intel for.

 

[8350rocks]

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

I still recall people in this forum claiming that Kaveri couldn't pass 3GHz because bulk was the poor thing in the universe. The facts are that Kaveri record is currently set on 6.1GHz. Average OC on air for Trinity, Richland and Kaveri are

A10-5800k: 4.5GHz (32nm FDSOI)
A10-6800k: 4.9GHz (32nm FDSOI)
A10-7850k: 4.5GHz (28nm bulk)

Everyone knew that Kaveri wouldn't break any OC world record, but bulk is much much better than what some claimed.

The reasons why AMD migrated from FDSOI to bulk are five (in no particular order):

- FDSOI is almost dead. Glofo canceled the 28nm FDSOI SHP process.
- bulk opens the foundry choices beyond being locked to glofo.
- bulk provides higher density needed for GCN.
- bulk provides clear migration path to FINFET.
- bulk reduces costs associated to migrating designs from/to dGPU.

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Again facts tell a different history... Average OC on air for Lynnfield, Sandy Bridge, and Ivy Bridge are

i7-880: 4.3GHz (45nm bulk)
i7-2600k: 5.1GHz (32nm bulk)
i7-3770k: 4.7GHz (22nm bulk finfet)

In the first place we can observe that 32nm bulk provides higher frequencies than 32nm FDSOI (see above). Thus bulk is not so bad as some pretend.

The reason why Ivy Bridge overclocks poor than sandy bridge has nothing to do with bulk neither with Intel abandoning HEDT.

The reason was that the introduction of finfets modified the shape of the voltage curve as shown in next figure

Around ~1.32V the 22nm FF curve crosses the 32nm bulk curve. This explains why Ivy Bridge introduced further gains in mobile but failed for overclockers. The interesting stuff is that 14nm FF doesn't have crossing point, which means that Broadwell will overclock much better than both Sandy and Ivy. If everything else goes the same, Broadwell-K would hit ~5.6Ghz on air.

22m FF was an intermediate process. 14nm FF will introduce again classical scaling.

1 SOI is NOT nearly dead...let me educate you:

1.) IBM still has 22nm FD-SOI SHP and so does STMicro, GF stopped because whatever they were doing was borked and AMD was not getting drawn into another screw up.

2.) SOI has STMicro (in small batches), Samsung, GF, IBM...hardly "locked" into choices...

3.) This is fallacy...the DIE SHRINK provided higher densities needed for GCN.

4.) FINFET on SOI is all there will be past ~7nm...who lied to you? Otherwise they will not have the density to go to the smaller transistors past 10nm on bulk, couple that with poor leakage properties and you get more than just physics issues.

5.) Cost vs. FD-SOI differences are minimal...UNLESS...you have yield issues...which is what I SUPREMELY SUSPECT was going on with GF @ 28nm FD-SOI. Now, I have not had confirmation outright...but it was all but spelled out to me before.

Too bad you know nothing of what you speak. Bulk is typically LOWER density than SOI...but an engineer knows that...not some speculative twit on TH.com website.

 

[The Q6660 Inside]

Intel users never admit that anything that comes from Intel is a failure. .

I am an Intel user and I say crapburst was an epic fail. Sadly, AMD tried to imitate that fail with faildozer. Somewhere along the line AMD lost their way and made the same mistakes they used to bash Intel for.

Bulldozer was just the first (and somewhat flawed) version, sort of like Agena and Piledriver would be Deneb, Steamroller is Thuban, etc.

 

[cdrkf]

Intel users never admit that anything that comes from Intel is a failure. .

I am an Intel user and I say crapburst was an epic fail. Sadly, AMD tried to imitate that fail with faildozer. Somewhere along the line AMD lost their way and made the same mistakes they used to bash Intel for.

I kinda get the analogy however the 'goal' of bulldozer if you like is very different to netburst. Netbursts design was to achieve the highest possible frequency on a single core. Bulldozer's goal was to cram as many cores onto a single bit of silicon within a specific TDP range. Unfortunately in the pursuit of more cores they decided to keep the individual cores 'light' making the bet that software as a whole would be multi-threaded enough to get a big boost of performance from the extra threads despite the lower instruction / clock of the individual cores. It proved to be a bad bet (although I would point out that AMD's 8 core fx chips are ageing quite well as software is slowly starting to use the threads).

 

[juanrga]

Spoiler

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

I still recall people in this forum claiming that Kaveri couldn't pass 3GHz because bulk was the poor thing in the universe. The facts are that Kaveri record is currently set on 6.1GHz. Average OC on air for Trinity, Richland and Kaveri are

A10-5800k: 4.5GHz (32nm FDSOI)
A10-6800k: 4.9GHz (32nm FDSOI)
A10-7850k: 4.5GHz (28nm bulk)

Everyone knew that Kaveri wouldn't break any OC world record, but bulk is much much better than what some claimed.

The reasons why AMD migrated from FDSOI to bulk are five (in no particular order):

- FDSOI is almost dead. Glofo canceled the 28nm FDSOI SHP process.
- bulk opens the foundry choices beyond being locked to glofo.
- bulk provides higher density needed for GCN.
- bulk provides clear migration path to FINFET.
- bulk reduces costs associated to migrating designs from/to dGPU.

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Again facts tell a different history... Average OC on air for Lynnfield, Sandy Bridge, and Ivy Bridge are

i7-880: 4.3GHz (45nm bulk)
i7-2600k: 5.1GHz (32nm bulk)
i7-3770k: 4.7GHz (22nm bulk finfet)

In the first place we can observe that 32nm bulk provides higher frequencies than 32nm FDSOI (see above). Thus bulk is not so bad as some pretend.

The reason why Ivy Bridge overclocks poor than sandy bridge has nothing to do with bulk neither with Intel abandoning HEDT.

The reason was that the introduction of finfets modified the shape of the voltage curve as shown in next figure

Around ~1.32V the 22nm FF curve crosses the 32nm bulk curve. This explains why Ivy Bridge introduced further gains in mobile but failed for overclockers. The interesting stuff is that 14nm FF doesn't have crossing point, which means that Broadwell will overclock much better than both Sandy and Ivy. If everything else goes the same, Broadwell-K would hit ~5.6Ghz on air.

22m FF was an intermediate process. 14nm FF will introduce again classical scaling.

1 SOI is NOT nearly dead...let me educate you:

1.) IBM still has 22nm FD-SOI SHP and so does STMicro, GF stopped because whatever they were doing was borked and AMD was not getting drawn into another screw up.

2.) SOI has STMicro (in small batches), Samsung, GF, IBM...hardly "locked" into choices...

3.) This is fallacy...the DIE SHRINK provided higher densities needed for GCN.

4.) FINFET on SOI is all there will be past ~7nm...who lied to you? Otherwise they will not have the density to go to the smaller transistors past 10nm on bulk, couple that with poor leakage properties and you get more than just physics issues.

5.) Cost vs. FD-SOI differences are minimal...UNLESS...you have yield issues...which is what I SUPREMELY SUSPECT was going on with GF @ 28nm FD-SOI. Now, I have not had confirmation outright...but it was all but spelled out to me before.

Too bad you know nothing of what you speak. Bulk is typically LOWER density than SOI...but an engineer knows that...not some speculative twit on TH.com website.

After the occasional laugh, let us begin by looking at data again:

1) In case you are been living under a rock (PUN intended) let me inform you that IBM is abandoning the foundry business.

2) Let us see the foundries choices for the 14nm node:

Global Foundries ==> FinFET on bulk
[strike]IBM ==> FinFET on SOI[/strike]
Intel ==> FinFET on bulk
Samsung ==> FinFET on bulk
STMicro ==> FDSOI
TSMC ==> FinFET on bulk
UMC ==> FinFET on bulk

Huh! FDSOI is very popular true? it accounts for about 2.1% of foundries total volume production. FinFET on bulk is 96.2%. There are technical reasons why FDSOI is rejected massively by engineers/scientists.

Moreover, STMicro FDSOI process is designed for very low-power devices such as phones.

For your information, AMD's new K12 core is being designed for FINFET on bulk; Keller is very happy with FINFET on bulk.

3) 32nm ==> 28nm gives up to 31% higher density. Still, Kaveri (28nm) is a 85% more dense than Richland (32 nm). This is why 28nm FDSOI was rejected to Kaveri. The needed density for GCN was introduced by the custom SHP bulk process. AMD already explained that they selected 28nm SHP bulk for Kaveri because:

Kaveri's 28 nm SHP bulk silicon process optimizes density at the expense of higher clock rates

http://www.tomshardware.com/reviews/fx-7600p-kaveri-apu,3842.html

http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k

4) FDSOI has manufacturing and scaling problems. This is why almost any foundry has migrated to FINFETs. 10nm FDSOI requires ~4nm channel thickness, which is too close to quantum confinement limit. Theoretical FDSOI 7nm hits quantum confinement limit for the channel thickness. This is why several experts consider that FDSOI cannot be scaled beyond 10nm and is a dead-road.

However, FINFET scales down: 10nm FINFET ==> 7nm FINFET ==> 5nm FINFET.

5) I was discussing the cost of migrating the design. Yield issues for a given foundry process has nothing to do with what I was saying.

 

[colinp]

That's right, woodchuck-chuckers - it's Groundhog Day!

 

[juanrga]

Bulldozer was just the first (and somewhat flawed) version, sort of like Agena and Piledriver would be Deneb, Steamroller is Thuban, etc.

Piledriver and Steamroller fixed some flaws of Bulldozer design (specially Steamroller which is a step backward from the CMT design). But to be fair, the entire arch is unfixable. This is why AMD is not releasing any Warsaw CPU based in Steamroller modules for servers and also why the new K12 core is a clean design that abandon CMT.

I kinda get the analogy however the 'goal' of bulldozer if you like is very different to netburst. Netbursts design was to achieve the highest possible frequency on a single core. Bulldozer's goal was to cram as many cores onto a single bit of silicon within a specific TDP range. Unfortunately in the pursuit of more cores they decided to keep the individual cores 'light' making the bet that software as a whole would be multi-threaded enough to get a big boost of performance from the extra threads despite the lower instruction / clock of the individual cores. It proved to be a bad bet (although I would point out that AMD's 8 core fx chips are ageing quite well as software is slowly starting to use the threads).

I agree on that 8 core FX chips are aging well, this is why AMD extended the current FX line up to 2015. The new gen games are scaling up well. This is a new example of scaling up to 8 cores:

 

[h2323]

Intel users never admit that anything that comes from Intel is a failure. .

I am an Intel user and I say crapburst was an epic fail. Sadly, AMD tried to imitate that fail with faildozer. Somewhere along the line AMD lost their way and made the same mistakes they used to bash Intel for.

I kinda get the analogy however the 'goal' of bulldozer if you like is very different to netburst. Netbursts design was to achieve the highest possible frequency on a single core. Bulldozer's goal was to cram as many cores onto a single bit of silicon within a specific TDP range. Unfortunately in the pursuit of more cores they decided to keep the individual cores 'light' making the bet that software as a whole would be multi-threaded enough to get a big boost of performance from the extra threads despite the lower instruction / clock of the individual cores. It proved to be a bad bet (although I would point out that AMD's 8 core fx chips are ageing quite well as software is slowly starting to use the threads).

Exactly, good post, not much else to it that is of value at this point.

 

[h2323]

That's right, woodchuck-chuckers - it's Groundhog Day!

LOL, ya think... these two argue back and forth back and forth about the same point over and over and over again. Great post for the release day of Edge of Tomorrow.

Lets rename the forum "juan and 8350 argue and say the same thing with some AMD news sometimes"

Oh by the way AMD owns the graphics card charts, pretty surprising from Tom's

 

[8350rocks]

Spoiler

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

I still recall people in this forum claiming that Kaveri couldn't pass 3GHz because bulk was the poor thing in the universe. The facts are that Kaveri record is currently set on 6.1GHz. Average OC on air for Trinity, Richland and Kaveri are

A10-5800k: 4.5GHz (32nm FDSOI)
A10-6800k: 4.9GHz (32nm FDSOI)
A10-7850k: 4.5GHz (28nm bulk)

Everyone knew that Kaveri wouldn't break any OC world record, but bulk is much much better than what some claimed.

The reasons why AMD migrated from FDSOI to bulk are five (in no particular order):

- FDSOI is almost dead. Glofo canceled the 28nm FDSOI SHP process.
- bulk opens the foundry choices beyond being locked to glofo.
- bulk provides higher density needed for GCN.
- bulk provides clear migration path to FINFET.
- bulk reduces costs associated to migrating designs from/to dGPU.

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Again facts tell a different history... Average OC on air for Lynnfield, Sandy Bridge, and Ivy Bridge are

i7-880: 4.3GHz (45nm bulk)
i7-2600k: 5.1GHz (32nm bulk)
i7-3770k: 4.7GHz (22nm bulk finfet)

In the first place we can observe that 32nm bulk provides higher frequencies than 32nm FDSOI (see above). Thus bulk is not so bad as some pretend.

The reason why Ivy Bridge overclocks poor than sandy bridge has nothing to do with bulk neither with Intel abandoning HEDT.

The reason was that the introduction of finfets modified the shape of the voltage curve as shown in next figure

Around ~1.32V the 22nm FF curve crosses the 32nm bulk curve. This explains why Ivy Bridge introduced further gains in mobile but failed for overclockers. The interesting stuff is that 14nm FF doesn't have crossing point, which means that Broadwell will overclock much better than both Sandy and Ivy. If everything else goes the same, Broadwell-K would hit ~5.6Ghz on air.

22m FF was an intermediate process. 14nm FF will introduce again classical scaling.

1 SOI is NOT nearly dead...let me educate you:

1.) IBM still has 22nm FD-SOI SHP and so does STMicro, GF stopped because whatever they were doing was borked and AMD was not getting drawn into another screw up.

2.) SOI has STMicro (in small batches), Samsung, GF, IBM...hardly "locked" into choices...

3.) This is fallacy...the DIE SHRINK provided higher densities needed for GCN.

4.) FINFET on SOI is all there will be past ~7nm...who lied to you? Otherwise they will not have the density to go to the smaller transistors past 10nm on bulk, couple that with poor leakage properties and you get more than just physics issues.

5.) Cost vs. FD-SOI differences are minimal...UNLESS...you have yield issues...which is what I SUPREMELY SUSPECT was going on with GF @ 28nm FD-SOI. Now, I have not had confirmation outright...but it was all but spelled out to me before.

Too bad you know nothing of what you speak. Bulk is typically LOWER density than SOI...but an engineer knows that...not some speculative twit on TH.com website.

After the occasional laugh, let us begin by looking at data again:

1) In case you are been living under a rock (PUN intended) let me inform you that IBM is abandoning the foundry business.

2) Let us see the foundries choices for the 14nm node:

Global Foundries ==> FinFET on bulk
[strike]IBM ==> FinFET on SOI[/strike]
Intel ==> FinFET on bulk
Samsung ==> FinFET on bulk
STMicro ==> FDSOI
TSMC ==> FinFET on bulk
UMC ==> FinFET on bulk

Huh! FDSOI is very popular true? it accounts for about 2.1% of foundries total volume production. FinFET on bulk is 96.2%. There are technical reasons why FDSOI is rejected massively by engineers/scientists.

Moreover, STMicro FDSOI process is designed for very low-power devices such as phones.

For your information, AMD's new K12 core is being designed for FINFET on bulk; Keller is very happy with FINFET on bulk.

3) 32nm ==> 28nm gives up to 31% higher density. Still, Kaveri (28nm) is a 85% more dense than Richland (32 nm). This is why 28nm FDSOI was rejected to Kaveri. The needed density for GCN was introduced by the custom SHP bulk process. AMD already explained that they selected 28nm SHP bulk for Kaveri because:

Kaveri's 28 nm SHP bulk silicon process optimizes density at the expense of higher clock rates

http://www.tomshardware.com/reviews/fx-7600p-kaveri-apu,3842.html

http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k

4) FDSOI has manufacturing and scaling problems. This is why almost any foundry has migrated to FINFETs. 10nm FDSOI requires ~4nm channel thickness, which is too close to quantum confinement limit. Theoretical FDSOI 7nm hits quantum confinement limit for the channel thickness. This is why several experts consider that FDSOI cannot be scaled beyond 10nm and is a dead-road.

However, FINFET scales down: 10nm FINFET ==> 7nm FINFET ==> 5nm FINFET.

5) I was discussing the cost of migrating the design. Yield issues for a given foundry process has nothing to do with what I was saying.

Who lied to you? The SOI consortium has a new technology from STMIcro, called UTBB, which stands for Ultra Thin Body and Box Fully Depleted Silicon On Insulator.

Additionally, what you are talking about is also untrue, because the only reason bulk is considered viable past ~10nm at all is because of FinFETs, however, you can still do the FinFETs on SOI.

http://www.soiconsortium.org/fully-depleted-soi/presentations/october-2013/SEH%20-%20Shanghai%20SOI%20Summit%20-%20Oct%202013.pdf

http://www.soiconsortium.org/fully-depleted-soi/presentations/october-2013/IBM%20-%20Shanghai%20SOI%20Summit%20-%20Oct%202013.pdf

There are presentations from the Consortium gathering last year talking about processes for FinFET on FD-SOI @ 7nm already...

Juan, give it up...you are the only one claiming that bulk is superior. In a room full of coders, IT staff, and engineers...you...without any formal education on the subject...are the only claiming that....

I am starting to see a trend...

Maybe you will too...

EDIT: In case you are lazy...

）
Quality Achieved
SEH has already achieved required wafer quality for FIN
‐
SOI and FD
‐
SOI.
For FD
‐
SOI, thickness range =< 1nm
   
Micro roughness is about 0.1nm RMS in AFM measurement.

=< 1nm...hmm...yep, looks like that 7nm node will be no problem...

There is, well you know...such thing as research...try it out. You could even pay for an education so that you might actually know half as much as you think.

Now please shut up about SOI. You know nothing.

 

[juanrga]

Spoiler

Spoiler

SOI scales very well upward and has much lower leakage. That enables you to turn the power up to 11 and squeeze more out if you are one of those MOAR POWAR! kind of guys.

Bulk scales poorly and has high leakage. That means when you turn the power up to 11, you get lots of power draw and less performance gains relatively. The difference in performance gain margins is quite telling of the differences between the 2 processes...I foretold we would see this the day the announcement came.

This is one of the many things I attempted to point out to someone who was singing praises for bulk process for months. I believe the comments were something along the lines of "the process will not make that big an impact".

I still recall people in this forum claiming that Kaveri couldn't pass 3GHz because bulk was the poor thing in the universe. The facts are that Kaveri record is currently set on 6.1GHz. Average OC on air for Trinity, Richland and Kaveri are

A10-5800k: 4.5GHz (32nm FDSOI)
A10-6800k: 4.9GHz (32nm FDSOI)
A10-7850k: 4.5GHz (28nm bulk)

Everyone knew that Kaveri wouldn't break any OC world record, but bulk is much much better than what some claimed.

The reasons why AMD migrated from FDSOI to bulk are five (in no particular order):

- FDSOI is almost dead. Glofo canceled the 28nm FDSOI SHP process.
- bulk opens the foundry choices beyond being locked to glofo.
- bulk provides higher density needed for GCN.
- bulk provides clear migration path to FINFET.
- bulk reduces costs associated to migrating designs from/to dGPU.

However, look at Intel, their desktop designs have not really improved much beyond high single digits on bulk process EVEN WITH DIE SHRINKS...but the ULP stuff has gotten a bit better than it was...

That is all from chasing the mobile squirrel and losing sight of HEDT, where people shill out big money for big systems and care nothing for power draw.

The problem is, these days, Johnny-college-student-art-major, and Susie-social-media-busybody are not Joe Q. Gamer. Since they far outnumber Joe Q Gamer, we got the shaft, and you see everyone trying to emulate, in a more successful manner, the failure that is Ultrabooks...even though Intel would never openly admit they were a failure.

Again facts tell a different history... Average OC on air for Lynnfield, Sandy Bridge, and Ivy Bridge are

i7-880: 4.3GHz (45nm bulk)
i7-2600k: 5.1GHz (32nm bulk)
i7-3770k: 4.7GHz (22nm bulk finfet)

In the first place we can observe that 32nm bulk provides higher frequencies than 32nm FDSOI (see above). Thus bulk is not so bad as some pretend.

The reason why Ivy Bridge overclocks poor than sandy bridge has nothing to do with bulk neither with Intel abandoning HEDT.

The reason was that the introduction of finfets modified the shape of the voltage curve as shown in next figure

Around ~1.32V the 22nm FF curve crosses the 32nm bulk curve. This explains why Ivy Bridge introduced further gains in mobile but failed for overclockers. The interesting stuff is that 14nm FF doesn't have crossing point, which means that Broadwell will overclock much better than both Sandy and Ivy. If everything else goes the same, Broadwell-K would hit ~5.6Ghz on air.

22m FF was an intermediate process. 14nm FF will introduce again classical scaling.

1 SOI is NOT nearly dead...let me educate you:

1.) IBM still has 22nm FD-SOI SHP and so does STMicro, GF stopped because whatever they were doing was borked and AMD was not getting drawn into another screw up.

2.) SOI has STMicro (in small batches), Samsung, GF, IBM...hardly "locked" into choices...

3.) This is fallacy...the DIE SHRINK provided higher densities needed for GCN.

4.) FINFET on SOI is all there will be past ~7nm...who lied to you? Otherwise they will not have the density to go to the smaller transistors past 10nm on bulk, couple that with poor leakage properties and you get more than just physics issues.

5.) Cost vs. FD-SOI differences are minimal...UNLESS...you have yield issues...which is what I SUPREMELY SUSPECT was going on with GF @ 28nm FD-SOI. Now, I have not had confirmation outright...but it was all but spelled out to me before.

Too bad you know nothing of what you speak. Bulk is typically LOWER density than SOI...but an engineer knows that...not some speculative twit on TH.com website.

After the occasional laugh, let us begin by looking at data again:

1) In case you are been living under a rock (PUN intended) let me inform you that IBM is abandoning the foundry business.

2) Let us see the foundries choices for the 14nm node:

Global Foundries ==> FinFET on bulk
[strike]IBM ==> FinFET on SOI[/strike]
Intel ==> FinFET on bulk
Samsung ==> FinFET on bulk
STMicro ==> FDSOI
TSMC ==> FinFET on bulk
UMC ==> FinFET on bulk

Huh! FDSOI is very popular true? it accounts for about 2.1% of foundries total volume production. FinFET on bulk is 96.2%. There are technical reasons why FDSOI is rejected massively by engineers/scientists.

Moreover, STMicro FDSOI process is designed for very low-power devices such as phones.

For your information, AMD's new K12 core is being designed for FINFET on bulk; Keller is very happy with FINFET on bulk.

3) 32nm ==> 28nm gives up to 31% higher density. Still, Kaveri (28nm) is a 85% more dense than Richland (32 nm). This is why 28nm FDSOI was rejected to Kaveri. The needed density for GCN was introduced by the custom SHP bulk process. AMD already explained that they selected 28nm SHP bulk for Kaveri because:

Kaveri's 28 nm SHP bulk silicon process optimizes density at the expense of higher clock rates

http://www.tomshardware.com/reviews/fx-7600p-kaveri-apu,3842.html

http://www.anandtech.com/show/7677/amd-kaveri-review-a8-7600-a10-7850k

4) FDSOI has manufacturing and scaling problems. This is why almost any foundry has migrated to FINFETs. 10nm FDSOI requires ~4nm channel thickness, which is too close to quantum confinement limit. Theoretical FDSOI 7nm hits quantum confinement limit for the channel thickness. This is why several experts consider that FDSOI cannot be scaled beyond 10nm and is a dead-road.

However, FINFET scales down: 10nm FINFET ==> 7nm FINFET ==> 5nm FINFET.

5) I was discussing the cost of migrating the design. Yield issues for a given foundry process has nothing to do with what I was saying.

Who lied to you? The SOI consortium has a new technology from STMIcro, called UTBB, which stands for Ultra Thin Body and Box Fully Depleted Silicon On Insulator.

Additionally, what you are talking about is also untrue, because the only reason bulk is considered viable past ~10nm at all is because of FinFETs, however, you can still do the FinFETs on SOI.

http://www.soiconsortium.org/fully-depleted-soi/presentations/october-2013/SEH%20-%20Shanghai%20SOI%20Summit%20-%20Oct%202013.pdf

http://www.soiconsortium.org/fully-depleted-soi/presentations/october-2013/IBM%20-%20Shanghai%20SOI%20Summit%20-%20Oct%202013.pdf

There are presentations from the Consortium gathering last year talking about processes for FinFET on FD-SOI @ 7nm already...

Juan, give it up...you are the only one claiming that bulk is superior. In a room full of coders, IT staff, and engineers...you...without any formal education on the subject...are the only claiming that....

I am starting to see a trend...

Maybe you will too...

EDIT: In case you are lazy...

）
Quality Achieved
SEH has already achieved required wafer quality for FIN
‐
SOI and FD
‐
SOI.
For FD
‐
SOI, thickness range =< 1nm
   
Micro roughness is about 0.1nm RMS in AFM measurement.

=< 1nm...hmm...yep, looks like that 7nm node will be no problem...

There is, well you know...such thing as research...try it out. You could even pay for an education so that you might actually know half as much as you think.

Now please shut up about SOI. You know nothing.

Lovely attempts to try to discuss about stuff that I didn't say and to mix FDSOI with FINSOI. I will ignore those attempts and your new ad hominem. I will focus again on facts.

Last paper on UTBB FDSOI on 14nm node reports a Tsi of 6nm and the same quantum confinement problem beyond 10nm that I mentioned before...

Even IBM researched both FINFET and FDSOI and concluded that FDSOI has scaling issues:

Although the FinFET and FDSOI devices share many design considerations there are also important
distinctions. The advantage of the FDSOI device is its simplicity. Unlike the FinFET it is transparent to the designer, it is also geometrically simpler, and the critical Tsi, silicon body thickness, is defined through a planar process. The disadvantage is in a thinner Tsi required to make the FDSOI device turn off as well as a FinFET. Due to fringing fields in the BOX, the Tsi in FDSOI device has to be more than half as thin as Tsi in a FinFET. As a result the extension resistance is a more significant problem, the fact that resistivity of thin films increases with decreasing thickness due to surface effects only exacerbates the problem further. In addition the thinnerTsi makes quantum confinement effects and surface
scattering more pronounced in the FDSOI device.

As I mentioned above FDSOI is almost dead (Global Foundries, IBM, Intel, Samsung, TSMC, UMC all them reject FDSOI for 14nm). You can continue spreading nonsense and lies about FDSOI (like when you said us that Kaveri was being delayed because was being made in a new 28nm FDSOI process at Glofo that only exists in your imagination) but reality will not change.

What matters is that AMD chose bulk for 28/20nm and FINFET on bulk for 14nm.

 

[blackkstar]

The discussion about AMD losing things by going bulk was exaggerated, but it is true. The general consensus on Kaveri desktop is that the IPC gains were wiped out by the fact that Kaveri can't clock as high. I don't see too many architectural changes that would have caused that. It seems far more likely to be related to the process instead.

Yeah, there are good OCing Kaveri's out there. But the fact remains that at the end of the day we had to see a stock clockspeed reduction thanks to going bulk and that the point of diminishing returns when raising TDP is much lower.

AMD lost a lot going from Richland to Kaveri in terms of clock speeds. It wasn't some catastrophic unable to break 3ghz type of event, but it was enough to hurt overall CPU performance.

 

[etayorius]

If Kaveri at least would had matched Richland Base and Turbo Frequency, 7850k would have been a very, very different story, that 10% would have been a complete Game Changer.

 

[juanrga]

The discussion about AMD losing things by going bulk was exaggerated, but it is true. The general consensus on Kaveri desktop is that the IPC gains were wiped out by the fact that Kaveri can't clock as high. I don't see too many architectural changes that would have caused that. It seems far more likely to be related to the process instead.

Yeah, there are good OCing Kaveri's out there. But the fact remains that at the end of the day we had to see a stock clockspeed reduction thanks to going bulk and that the point of diminishing returns when raising TDP is much lower.

AMD lost a lot going from Richland to Kaveri in terms of clock speeds. It wasn't some catastrophic unable to break 3ghz type of event, but it was enough to hurt overall CPU performance.

Evidently part of the IPC increment introduced by Steamroller architectural improvements was dropped by the lower base clock. Nobody is negating that! What is being discussed here is why AMD underclocked Kaveri.

We know that Kaveri average OC on air is 4.5GHz. We have seen Kaveri benchmarked @4.7GHz:

http://www.hardcoreware.net/kaveri-7850k-overclocked-benchmarks/

Anand showed that Kaveri can be OC up to ~4.2GHz without changing the voltage (1.25V) and with almost same power consumption 147W vs 148W

We also know that early talks and docs from AMD mentioned explicitly that Kaveri was a 100W APU originally.

Thus the question is why did AMD downrated Kaveri to 95W and reduced the base frequency from the expected 4.0GHz (check my BSN article) to final 3.7GHz? Why did AMD do that when Kaveri can hit 4.5/4.7GHz on air?

This is the million dollar question that nobody is answering!

 

[de5_Roy]

If Kaveri at least would had matched Richland Base and Turbo Frequency, 7850k would have been a very, very different story, that 10% would have been a complete Game Changer.

"current" kaveri's focus is on mobile. check the mobile kaveri clockrate specs and compare them to richland - kaveri has both clockrate and ipc advantage. then there's igpu uarch advantage, better power management advantage, igpu driver support advantage (7660D never got much love) - all with a half node shrink and different silicon substrate. game has already been changed. but when the foundry partner is glofo, there is always yield problems. that's why you get only one fully enabled sku. keep in mind that these haven't shows up at the hands of independent reviewers in retail laptops yet. amd hasn't released pricing info either.

 

[juanrga]

That's right, woodchuck-chuckers - it's Groundhog Day!

LOL, ya think... these two argue back and forth back and forth about the same point over and over and over again. Great post for the release day of Edge of Tomorrow.

Lets rename the forum "juan and 8350 argue and say the same thing with some AMD news sometimes"

Oh by the way AMD owns the graphics card charts, pretty surprising from Tom's

My proposal is to rename the forum to "bash about anything AMD related":

If AMD announces bulk, some here spread nonsense/FUD about bulk. E.g. "Kaveri will not hit 3GHz".

If AMD announces migration to APUs, some here spread nonsense/FUD about APUs. E.g. "AMD is abandoning HEDT".

If AMD announces Steamroller, some here spread nonsense/FUD about Steamroller. E.g. "Steamroller will be 45% slower than Piledriver".

If AMD announces HSA, some here spread nonsense/FUD about HSA. E.g. "There is no market for HSA in the real world."

If AMD announces MANTLE, some here spread nonsense/FUD about MANTLE. E.g. "nobody will use mantel", "AMD stole mantel from Microsoft/Nvidia".

If AMD announces strong interest in ARM, some here spread nonsense/FUD about ARM. E.g. "ARM cannot run even a print server", "ARM cannot scale up".

If AMD announces Skybridge, some here spread nonsense/FUD about Skybridge. E.g. "Which is why I can say, without seeing a single diagram or piece of information on the core design, that I know AMD's universal core approach is going to suck in performance terms.".

If AMD announces FINFET, some here spread nonsense/FUD about FINFET. E.g. "FINFET cannot scale beyond 10nm".

If AMD announces support for linux, some here spread nonsense/FUD about linux. E.g. "Android cannot display webpages correctly because linux sucks".

And so on and so on...

EDIT: Change it to "bash about anything AMD related or mention juan"

 

[juanrga]

But at what cost. A 125watt apu ? After 4.5 there is little gain on trinity and is only for show. llano actually scaled much better but with lower clocks.

Above table shows zero variation in voltage between 3.7GHz and 4.2Ghz. Kaveri with 4.0GHz base clock would be rated at ~100W instead of current 95W.

FD-SOI or bulk looks to be a none issue and a product of manufacturing preference.

http://www.goldstandardsimulations.com/index.php/news/blog_search/fd-soi-finfets-can-offer-better-performance-than-bulk-finfets/

I agree on that it is a question of choice. The zero percent (0%) of foundries have selected FD-SOI finfets, not only by the technical reasons given before, but also by economic reasons (reason why IBM is abandoning the foundry business).

 

[de5_Roy]

If AMD announces Skylake, some here spread nonsense/FUD about Skylake. E.g. "Which is why I can say, without seeing a single diagram or piece of information on the core design, that I know AMD's universal core approach is going to suck in performance terms.".

amd didn't announce skylake. you're the one spreading Nonsense and FUD. skylake is the codename for intel's 14nm brw successor.

 

[juanrga]

If Kaveri at least would had matched Richland Base and Turbo Frequency, 7850k would have been a very, very different story, that 10% would have been a complete Game Changer.

"current" kaveri's focus is on mobile. check the mobile kaveri clockrate specs and compare them to richland - kaveri has both clockrate and ipc advantage. then there's igpu uarch advantage, better power management advantage, igpu driver support advantage (7660D never got much love) - all with a half node shrink and different silicon substrate. game has already been changed. but when the foundry partner is glofo, there is always yield problems. that's why you get only one fully enabled sku. keep in mind that these haven't shows up at the hands of independent reviewers in retail laptops yet. amd hasn't released pricing info either.

Kaveri desktop has been released only in two quad-core SKU. Only one mobile SKU is dual-core whereas the other six mobile SKUs are quad-core.

The no existence of more dual-core salvage parts from quad core dies is clearly indicating extremely good yields.

 

[juanrga]

If AMD announces Skybridge, some here spread nonsense/FUD about Skybridge. E.g. "Which is why I can say, without seeing a single diagram or piece of information on the core design, that I know AMD's universal core approach is going to suck in performance terms.".

amd didn't announce skylake. you're the one spreading Nonsense and FUD. skylake is the codename for intel's 14nm brw successor.

It was clearly a typo. I did mean Skybridge. I edited and corrected it ^^. You would learn the difference between making an occasional typo and the systematic spreading of daily FUD/nonsense such as "it is going to suck in performance terms.".

Spoiler

E.g. this is dictionary entry for FUD:

Fear, Uncertainty and Doubt. Also known as scare tactics, either accomplished by threat or making the opponent doubt his standpoint. Not only used in lawsuits, but also in politics and military propaganda.

And this is entry for typo:

A misspeeling of a word when typing or texting

Its annoying when people point out your erorrs and correct you, because some times you don't focus on your grammar and punctuation like some fucking nerd or maybe you just didn't know how to speel the damn word.

Note the fine irony in this last definition

 

[juanrga]

AMD demonstrates first freesync monitor prototype

As predicted, MANTLE and Steambox already having an impact on consoles lifecycles. According to AMD, Sony and Microsoft "are already thinking about what comes next"

http://hereisthecity.com/en-gb/2014/06/05/amd-expect-ps4-and-xbox-one-life-cycles-will-be-shorter-than-bef/
http://www.gamezone.com/news/2014/06/06/xbox-one-ps4-lifecycle-probably-going-to-be-shorter-says-amd

I partially agree with the second article speculation about PS4/Xbox1 being probably the last consoles. If Steamboxes match expectations and if MANTLE finally arrives to linux, then I see no advantage on traditional console

I still recall when an 'expert' here pretended to convince me that "APUs are only for cheap gaming" Yeah sure!

AMD Aims to Revolutionize Commercial PCs With New AMD PRO A Series Line of APUs