AMD CPU speculation... and expert conjecture

[sarinaide]

I heard DT was dying around 2009, then it wasn't, DT is not declining in principle its declining as above because Intel has nothing useful to upgrade its own line to, and for some odd reason people are afraid on the alternative products. So the decline is largely Intels own doing but it is no way indicative that this is a diminishing trend.

The server market is diversified, there are clear blue prints of 8 core jaguar server based parts with co-processor support, basically a plug in PCI-e card with another 8 core processor, and I don't believe its limited to 8 core, think it will have 4-6-8-16 core models for low voltage servers which are more than copious at handling a lot of enterprises servers.

As AMD evolves its arches so will the standard server grade Opterons improve with it. And for normal home servers and lan servers a FX8150 (back then) was cheap and efficient at doing that job very well as a file server.

intel clearly have x86 performance advantage with a huge margin
they only need to clock their cpu higher (they have enough tdp room) and no amd can match their cpu performance at that price bracket (example i3@4ghz will trash any amd at current price in same price bracket fir majority of tasks)

means they can go for 95w i3 , 125w i5 & i7 , >150w i7 extreme
instead of releasing many models of same series they should opt for less models and more vfm product
for example : locked and unlocked i3/5/7 @4ghz base, these will be enough and since they will have a huge performance gap (like i3 vs i5) then none of them will interfare with each other's sales
but that will surely trash amd cpu


and their "HD" or gaming igpu fever is a foolish thing imo
because this is consuming die area and tdp
thus increasing cost of production and less profit margin

imo, they should leave gaming gpus to amd and nvidia
hd or hd2000 is enough for basic purpose , and for gaming purpose majority (excluding fanboys/girls) user prefer dedicated gpu

so in short
1. clock cpus higher and privide unlocked for each series.
2. decrease number of models in same series
3. limit gpu to hd or hd2000 to save die and tdp
4. leave graphics to amd/nvidia

1) Unlocked processors are a nightmare to intel RMA. Needless to say to provide a unlocked processor they also need to higher the TPD substantially or end up with instability.

2) Have you tried running benches for intel CPU's at 4 then 4.5, some have deminishing values, it probably then runs into architectural limitations of trying to ram down more data. The other aspect is a Intel CPU at 4-4.5ghz is as power hungry as a AMD processor and this is not what Intel wants.

3) Yes Intel HD is a big waste of die space but honestly if they didn't have it they will be paying AMD and Nvidia to produce output for all Mobility parts and AMD and Nvidia will have carte blanche in ripping Intel off for it, so HD is basically a fall back plan.

4) Yes but is your smartphone attending to your gaming needs or your content production needs or your HTPC needs. So in short no your smartphone is not replacing your desktop, its just ancillory.

5) Don't care about a krait, andreno, super mario, romeo romeo werefore art thou. Gaming on a phone is pointless and stupid I like my BF3/4 and I like my DayZ and I like to speak to people on my phone or at worst use chat apps, this doesn't need a hardcore phone and nor do I want to drop $1000 on a phone when I can build a much more useful PC or get a notebook.

 

[JAYDEEJOHN]

I think where ever Intel goes, they have competition
New markets
x86, we may see AMD creep up on them, which they have
x86 covers server, plus you also have nVidia and yes ARM of course

BUT
If all these other companies, QC,AMD,Samsung et al are competing, Intel has invested huge amounts of monies into their fabs
They didnt exactly forcast for the downturn, nor the shift in markets, or we would be seeing a better showing from them.

So, they have to keep those fabs fed, their bread n butter is shrinking, they cant command in the newer markets, their entire structure is setup on certain ROI, with X amount of sales all at those projected ROIs.
Intel has 7 billion to work with, I believe, its less than what I thought theyd have at this time

If they do let others in their fabs, it will be interesting to see if theres any takers, as some are hinting those costs are very very high for their process, so between the customers they already have, the competitors, at high high pricing, it seems to me there arent alot of takers going to use their fabs.

Thats the point here, not the status quo
Those days are behind Intel

 

[JAYDEEJOHN]

http://www.xbitlabs.com/news/other/display/20130403233037_The_End_of_Moore_s_Law_Is_on_Horizon_AMD.html

 

[gamerk316]

1) Unlocked processors are a nightmare to intel RMA. Needless to say to provide a unlocked processor they also need to higher the TPD substantially or end up with instability.

Nevermind that they kill profits. How much money do you think Intel lost because people OC'd the Q6600 over the years?

2) Have you tried running benches for intel CPU's at 4 then 4.5, some have deminishing values, it probably then runs into architectural limitations of trying to ram down more data. The other aspect is a Intel CPU at 4-4.5ghz is as power hungry as a AMD processor and this is not what Intel wants.

The performance aspect makes sense. Assuming you have a fast enough CPU where there aren't any major application bottlenecks, a faster CPU increases performance simply because the application has a shorter time period where it isn't running (more cycles = more periods for threads to run). Performance increases diminish with clock speed increases; nothing new here.

3) Yes Intel HD is a big waste of die space but honestly if they didn't have it they will be paying AMD and Nvidia to produce output for all Mobility parts and AMD and Nvidia will have carte blanche in ripping Intel off for it, so HD is basically a fall back plan.

Remember, that GPU's are much better processors for certain types of tasks. All parallel workloads are moving to GPGPU's now, so Intel has to get in the game, or risk being left behind.

In theory, if GPU clockspeeds start to approach the 2GHz range, you could theoretically run an OS on a GPU like architecture without too much performance loss in most general tasks. That would be disastrous for Intel.

4) Yes but is your smartphone attending to your gaming needs or your content production needs or your HTPC needs. So in short no your smartphone is not replacing your desktop, its just ancillory.

Today? No. But if internet latency drops enough, the cloud gaming becomes viable.

5) Don't care about a krait, andreno, super mario, romeo romeo werefore art thou. Gaming on a phone is pointless and stupid I like my BF3/4 and I like my DayZ and I like to speak to people on my phone or at worst use chat apps, this doesn't need a hardcore phone and nor do I want to drop $1000 on a phone when I can build a much more useful PC or get a notebook.

But for the general user, Angry Birds is good enough entertainment. Most people don't want games that are "too" involving. They want to sit down for 15 minutes, play a game, and stop. How many PC games let you do that easily? [Heck, how many PC games have an autosave over that period?]

 

[griptwister]

IMO, Haswell on the z87 side was just an excuse to fix the heat spreader which should have been fixed a long time ago. It is also an excuse to bump up the prices $50 for the better graphics. Which we all know is total bull, Intel knows that Intel's HD graphics are not their focus. They just wanted an excuse to release a new series of CPUs and make a bigger profit than the previous series. You also notice the prices on the 3570K are only $70 away from the i7 now?

Also, I find this statement interesting. As Ivy Bridge uses more power when overclocked that Sandy B.

http://www.xbitlabs.com/news/other/display/20130403233037_The_End_of_Moore_s_Law_Is_on_Horizon_AMD.html

 

[sarinaide]

I believe even those here backing team blue with Sandy systems will deep down and honestly say they still have nothing to upgrade to, a generation that is not even 10% faster than its 2 year old architecture is not a upgrade. The IMC is the only thing that has really improved even GT2 is barely pushing 15% over HD4000.

Dropping the die size adds complexity and cost, intel is seeing it first hand yet they seem resigned to just releasing Haswell and moving onto the 14nm Broadwell for 2015 and hopefully better utilise the time on Broadwell. You can almost say that 22nm kind of failed its just moving on admitting defeat. To me its not the right approach neither was socket 1150, but as above its a money spinner and people will be swayed to buying haswell no matter how bad it actually is.

Anyways before this becomes Intel speculation 2.0 will move on.

Speaking with a expert in the field contrary to the belief that Steam roller represents a new arch, it is implicitly stated that steam roller is the last of the bulldozer arch with Excavator representing a new arch.

Excavator speculation;

Twin Hyper Transport 4.0 busses
Quad Channel DDR4 support
RCM
2,4,6,8 core variants on DT/Mobile
Each core will have a unshared FPU
4 stages of cache, I am assuming this is like Intel and refers to stacking.
As to die size, most speculating 20nm.

No information on unified sockets and thus no information on a Radeon SIMD processor, though there is armchair speculation on this that the complexity of the iGPU will be for GPGPU calculation capable of matching current high end parts, that is a lot of FLOPS.

 

[truegenius]

I believe even those here backing team blue with Sandy systems will deep down and honestly say they still have nothing to upgrade to

i too like i5-2500k over i5-3570k
before ivy was launched i thought that ivy's smaller node will be good but it wasn't

Excavator speculation;
Twin Hyper Transport 4.0 busses
Quad Channel DDR4 support
RCM
2,4,6,8 core variants on DT/Mobile
Each core will have a unshared FPU
4 stages of cache, I am assuming this is like Intel and refers to stacking.
As to die size, most speculating 20nm.

hyper transport 4.0 ! that means that Ex-[strike]Girfriend[/strike] will have pcie v3.0 (maybe upto 32 lanes)
i told ya ! they need faster ht for pcie3.0 . [strike]f@#$[/strike] yeah B)

quad channel !

unshared fpu ! means amd is admitting that bd was a crap design
bulldozer = bullcrap (this line may recieve a lot of heat )

at least exy will represent a good upgrade from my current 1090t

 

[jdwii]

I believe even those here backing team blue with Sandy systems will deep down and honestly say they still have nothing to upgrade to, a generation that is not even 10% faster than its 2 year old architecture is not a upgrade. The IMC is the only thing that has really improved even GT2 is barely pushing 15% over HD4000.

Dropping the die size adds complexity and cost, intel is seeing it first hand yet they seem resigned to just releasing Haswell and moving onto the 14nm Broadwell for 2015 and hopefully better utilise the time on Broadwell. You can almost say that 22nm kind of failed its just moving on admitting defeat. To me its not the right approach neither was socket 1150, but as above its a money spinner and people will be swayed to buying haswell no matter how bad it actually is.

Anyways before this becomes Intel speculation 2.0 will move on.

Speaking with a expert in the field contrary to the belief that Steam roller represents a new arch, it is implicitly stated that steam roller is the last of the bulldozer arch with Excavator representing a new arch.

Excavator speculation;

Twin Hyper Transport 4.0 busses
Quad Channel DDR4 support
RCM
2,4,6,8 core variants on DT/Mobile
Each core will have a unshared FPU
4 stages of cache, I am assuming this is like Intel and refers to stacking.
As to die size, most speculating 20nm.

No information on unified sockets and thus no information on a Radeon SIMD processor, though there is armchair speculation on this that the complexity of the iGPU will be for GPGPU calculation capable of matching current high end parts, that is a lot of FLOPS.

^ Thats what everyone wants i think even Amd's engineers(also didn't Amd hire an ex engineer who people say is one of the best) Again i still want people to answer who is increasing performance more per year Amd or Intel.
Also is anyone else sad that Video cards will only come out new every other year. I used to wait every other year to get a video card since usually they shrink one year and then fine tune the next year(the time i bought the GPU). Looking to upgrade my 6950 when they come out with the radeon 8000 series(if i have the money)

 

[8350rocks]

GCN 2.0 seems to be AMD's priority for the next year or 2 in the GPU market, they are talking it up quite a bit, and the HD 8XXX series is going to be interesting to watch. I haven't seen a really hard number on the performance gains over GCN 1.0

I also think that Excavator will end up on either GloFo or TSMC 20nm SOI...AMD seems to be aiming squarely at that. Though Steamroller WILL be a slight die shrink...(28nm process) I am nearly certain at this point that Excavator will entail a new socket. Now, the interesting part of that is, FM3 is supposedly coming with Kaveri, if that's the case it means that the Kaveri (steamroller 28nm process) die would likely be the same size as an excavator based APU die. That means LOTS more room for shaders/cores and other GPU components on the die, which would likely mean we really would see HD 7770 equivalent iGPUs from AMD in the Excavator era.

Also, steamroller is supposed to be a 30% improvement in performance over piledriver, if they roll out a dieshrink and get similar performance gains at 20nm versus the shrink to 28nm...that would be incredibly impressive...

Dare I ask it...? Could we theroretically see a return to the days of the Athlon performance increase @ excavator?? Haswell and broadwell are not supposed to be huge differences in performance, and I think intel is actually considering going to SOI over their current process because of the issues they're having with such small processes. Meanwhile GloFo just announced that 20nm SOI using a new process they've pioneered is ramping up for full delivery to customers for development.

 

[lilcinw]

bulldozer = bullcrap (this line may recieve a lot of heat )

And Exy = Sexy?

Also, steamroller is supposed to be a 30% improvement in performance over piledriver, if they roll out a dieshrink and get similar performance gains at 20nm versus the shrink to 28nm...that would be incredibly impressive...

From what I recall of the old AMD slides the 30% was still comparing SR to BD not PD.

An unshared FPU would be nice. It should make for a better CPU for Folding@Home. An SIMD array with DDR4 access would probably be even better but it might be difficult to utilize properly (F@H isn't always on the ball when it comes to new tech). Hopefully HSA produces something significant rather than gimmicks that only appear at the occasional conference.

 

[8350rocks]

bulldozer = bullcrap (this line may recieve a lot of heat )

And Exy = Sexy?

Also, steamroller is supposed to be a 30% improvement in performance over piledriver, if they roll out a dieshrink and get similar performance gains at 20nm versus the shrink to 28nm...that would be incredibly impressive...

From what I recall of the old AMD slides the 30% was still comparing SR to BD not PD.

An unshared FPU would be nice. It should make for a better CPU for Folding@Home. An SIMD array with DDR4 access would probably be even better but it might be difficult to utilize properly (F@H isn't always on the ball when it comes to new tech). Hopefully HSA produces something significant rather than gimmicks that only appear at the occasional conference.

SIMD would be interesting...though I am pretty certain their GPUs run on that type of instruction setup(I could not easily find confirmation...), that would be an interesting junction of CPU/GPU tech...

 

[JAYDEEJOHN]

Remember, tho nodes may be slowing design doesnt have to.
In the past AMD was heavily dependent on process maturation in their cpus, and still will be to some extent, but they wont be jumping on the earliest time for node ramp as seen in the past, which puts their cadence backwards if you will, new process, new process new design, to mature process new design is what I see, slower but not halved

 

[JAYDEEJOHN]

http://www.aida64.com/whatsnew/aida64-march-update-AMD-Richland-APU-Windows-8
Preliminary support for AMD “Beema” APUs and GDDR5 system memory.

 

[palladin9479]

Twin Hyper Transport 4.0 busses
Quad Channel DDR4 support
RCM
2,4,6,8 core variants on DT/Mobile
Each core will have a unshared FPU
4 stages of cache, I am assuming this is like Intel and refers to stacking.
As to die size, most speculating 20nm.

Well seeing as I've heard nothing about "Hypertransport 4" that the entire design of HT makes "twin buss" a rather laughable idea.

DDR4 isn't "channeled" like current memory buss's, each chip / module has a direct line to CPU instead of being connected together in serial. So technically it would be eight to sixteen "channels" depending on how they implement it.

Each Core does have it's own 128-bit FPU though the design's don't really explain that well. It has a single scheduler that push's data into whichever path is needing to go in unless a 256-bit instruction comes along, then it combined them into a single unit. Also considering the design of their CPU I highly doubt this is going to happen.

The entire BD (PD / SR) design is to save die space and thus implement as many uniquely addressable integer units as possible. Their not sharing execution units, their sharing schedulers and cache.

Also found an interesting hobby board while looking to upgrade my home server. Seems VIA isn't out of the picture, they've just shifted focus's.

http://www.viaembedded.com/en/

Most of that is WAY too expensive for home use unless your trying to do Jarvis like home automation. One product kinda poked it's head out, the VE-900 seems to be a nice cheap product for home labs and other experimental setups. $89 for a board + 1.4Ghz dual core 64-bit CPU that has everything else integrated. Slap it into a SFF Mini-ITX case with a little 80~90W picoPSU and you got a cheap experimental box.

 

[jdwii]

GCN 2.0 seems to be AMD's priority for the next year or 2 in the GPU market, they are talking it up quite a bit, and the HD 8XXX series is going to be interesting to watch. I haven't seen a really hard number on the performance gains over GCN 1.0

I also think that Excavator will end up on either GloFo or TSMC 20nm SOI...AMD seems to be aiming squarely at that. Though Steamroller WILL be a slight die shrink...(28nm process) I am nearly certain at this point that Excavator will entail a new socket. Now, the interesting part of that is, FM3 is supposedly coming with Kaveri, if that's the case it means that the Kaveri (steamroller 28nm process) die would likely be the same size as an excavator based APU die. That means LOTS more room for shaders/cores and other GPU components on the die, which would likely mean we really would see HD 7770 equivalent iGPUs from AMD in the Excavator era.

Also, steamroller is supposed to be a 30% improvement in performance over piledriver, if they roll out a dieshrink and get similar performance gains at 20nm versus the shrink to 28nm...that would be incredibly impressive...

Dare I ask it...? Could we theroretically see a return to the days of the Athlon performance increase @ excavator?? Haswell and broadwell are not supposed to be huge differences in performance, and I think intel is actually considering going to SOI over their current process because of the issues they're having with such small processes. Meanwhile GloFo just announced that 20nm SOI using a new process they've pioneered is ramping up for full delivery to customers for development.

Again Amd said 15% boost in performance per watt(although it looks more than that)

 

[mayankleoboy1]

http://www.aida64.com/whatsnew/aida64-march-update-AMD-Richland-APU-Windows-8
Preliminary support for AMD “Beema” APUs and GDDR5 system memory.

From the same article : "“Haswell-ULT” SoC,"
So they are making Haswell based SoC, presumably used in smartphones.

 

[JAYDEEJOHN]

Yea, it not like w1zzard over at TPU, where he fudges a few things that could be discovered.
Also the DDR4 as well

 

[bobbybamf12]

Okay this is how its going to go down. AMD going to say that steamroller going to get 30% performance increased. Then when the product comes out you will see that they just up the clock speed and tweak some things. AMD lost all their credibility after bulldozer and they shouldn't even put a percentage on how much better steamroller is then piledriver. All i care about is ipc! I couldn't careless about a cpu hitting 4.5ghz stock. I can overclock the cpu myself, of course it would be sweet hitting 4.5ghz on 125 watts but still won't make the cpu worth it. Hopefully there focus is on make a huge jump in ipc rather then adding more mhz to clock speeds cause that will only get them so far.

o' rlly? What is the base for your predictions? Keep in mind there is a new process and there is a new architecture and they probably want to compete with Intel on single threading because that's where they have to improve. Keep in mind AMD has a smart group of people working with them. They recognized their issues with single GPUs and were able to fix those, who knows, they might fix their single threaded performance due to high demand. That's the only thing holding them back.

Yeah well i do wanna see AMD succeed and all but if they don't fixed their ipc on their cpu's it will hurt them majorly. If I am correct about steamroller then AMD will probably will ruin their chance to ever make a comeback in the high end cpu market. Most people that are fairly knowledgeable about cpu's will look at AMD ipc performance over how high the ghz are. People will spend money for 2500/3570 even though it has 4 less threads and cost more money. Steamroller ipc will have to be within 15% of sandy bridge ipc to keep a good chunk of the cpu market.

I guess we will have to see how it is once it releases.

btw i give you best answer on my thread. lol

 

[jdwii]

Just found a person that is pretty funny seems to hate logan a lot.

http://www.youtube.com/watch?NR=1&v=AbMYV8Djt7k&feature=fvwp


Don't know why i like them so much i like when he compares a CPU to in engine but so far my favorite part is when he talks about the processes in the task messenger i'm probably going to subscribe to his channel i find this guy to be quite hilarious in a stupid way. Go to like 27 min in i laughed pretty hard lol

 

[truegenius]

Just found a person that is pretty funny seems to hate logan a lot.

http://www.youtube.com/watch?NR=1&v=AbMYV8Djt7k&feature=fvwp


Don't know why i like them so much i like when he compares a CPU to in engine but so far my favorite part is when he talks about the processes in the task messenger i'm probably going to subscribe to his channel i find this guy to be quite hilarious in a stupid way. Go to like 27 min in i laughed pretty hard lol

you are doing this to make me jealous ain't you !
because I have super slow net ( <10KBps ) so I can't watch any video at any quality

 

[de5_Roy]

you are doing this to make me jealous ain't you !
because I have super slow net ( <10KBps ) so I can't watch any video at any quality

i don't even have flashplayer!
.... that's not as bad as having slow net connection but it must count in some way....

AMD Shows Convertible PC Designs, Demos “Temash” and “Kabini” in Action.
AMD Demonstrates Next-Gen Elite Mobility APUs
http://www.xbitlabs.com/news/mobile/display/20130404185437_AMD_Shows_Convertible_PC_Designs_Demos_Temash_and_Kabini_in_Action.html

http://www.xbitlabs.com/news/other/display/20130404143508_TSMC_Is_Ahead_of_Its_Own_20nm_Roadmap_Report.html
i didn't understand this part:

Given the fact that TSMC’s 20nm process technology will come only in one version and will neither be tailored for performance nor power consumption, but will attempt to address completely different needs, not all customers of TSMC will use it.

what are the other factors? yields?

 

[palladin9479]

Just found a person that is pretty funny seems to hate logan a lot.

http://www.youtube.com/watch?NR=1&v=AbMYV8Djt7k&feature=fvwp


Don't know why i like them so much i like when he compares a CPU to in engine but so far my favorite part is when he talks about the processes in the task messenger i'm probably going to subscribe to his channel i find this guy to be quite hilarious in a stupid way. Go to like 27 min in i laughed pretty hard lol

Most of what they were saying was flat wrong. It was in response to the review TS did where they found the FX8350 performed better then anticipated. Everyone got extremely "butt hurt" over that, was hilarious to see peoples reactions.

 

[jdwii]

Just found a person that is pretty funny seems to hate logan a lot.

http://www.youtube.com/watch?NR=1&v=AbMYV8Djt7k&feature=fvwp


Don't know why i like them so much i like when he compares a CPU to in engine but so far my favorite part is when he talks about the processes in the task messenger i'm probably going to subscribe to his channel i find this guy to be quite hilarious in a stupid way. Go to like 27 min in i laughed pretty hard lol

you are doing this to make me jealous ain't you !
because I have super slow net ( <10KBps ) so I can't watch any video at any quality

Sorry to hear that even at home i can use hughesnet which is OK at night with 1mb download speed usually. Hope others watch this guy though he is funny.

 

[jdwii]

Just found a person that is pretty funny seems to hate logan a lot.

http://www.youtube.com/watch?NR=1&v=AbMYV8Djt7k&feature=fvwp


Don't know why i like them so much i like when he compares a CPU to in engine but so far my favorite part is when he talks about the processes in the task messenger i'm probably going to subscribe to his channel i find this guy to be quite hilarious in a stupid way. Go to like 27 min in i laughed pretty hard lol

Most of what they were saying was flat wrong. It was in response to the review TS did where they found the FX8350 performed better then anticipated. Everyone got extremely "butt hurt" over that, was hilarious to see peoples reactions.

Yeah what i enjoy about the videos(which i put in the background while i do other things) is the guy behind the camera, even the guy talking ignores what he says have the time.

 

[mayankleoboy1]

I believe even those here backing team blue with Sandy systems will deep down and honestly say they still have nothing to upgrade to, a generation that is not even 10% faster than its 2 year old architecture is not a upgrade. The IMC is the only thing that has really improved even GT2 is barely pushing 15% over HD4000.

Dropping the die size adds complexity and cost, intel is seeing it first hand yet they seem resigned to just releasing Haswell and moving onto the 14nm Broadwell for 2015 and hopefully better utilise the time on Broadwell. You can almost say that 22nm kind of failed its just moving on admitting defeat. To me its not the right approach neither was socket 1150, but as above its a money spinner and people will be swayed to buying haswell no matter how bad it actually is.

22nm worked quite good in the Notebook CPU and iGPU space, specially the perf/watt and perf numbers. A lot of that is due to Turbo clocking more cores, at higher speeds, for longer periods of workload.
Desktops? Who cares about Desktops? Not Intel

Speaking with a expert in the field contrary to the belief that Steam roller represents a new arch, it is implicitly stated that steam roller is the last of the bulldozer arch with Excavator representing a new arch.

Interesting. Not really surprising, as the Excavator is/was supposed to be "extreme fusion of CPU and GPU".

No information on unified sockets and thus no information on a Radeon SIMD processor, though there is armchair speculation on this that the complexity of the iGPU will be for GPGPU calculation capable of matching current high end parts, that is a lot of FLOPS.

Depends on what is considered as "high end" parts. Even at 20nm and GCN2.0 , i dont think that the highest end APU can match a HD7850 level perf.