AMD CPU speculation... and expert conjecture

[Fidgetmaster]

http://i1264.photobucket.com/albums/jj500/jspinks2388/FX-430047ghzBF4_zps0cebad21.png?t=1406531934

And they say it couldn't be done on a lowly 760 board haha,

I wanna go further but I'm already really pushing it haha

Okay I will be honest its starting to give my 920 a run for its money at this speed haha and all fanboy crap aside, its competing with my 920...and no I don't want to admit that haha

 

[cdrkf]

I think a 3 module, hex core APU isn't such a bad idea to be honest. I guess the issue with APU is they have a set budget of transistors, so +1 cpu module = -1 (or 2) GPU clusters. On the other hand, given the mix of workloads, for many people 384 GPU cores + 6 CPU would be a better mix than 512 + 4. I guess the problem is that you have to produce an additional die to create that config although I think the 2 would be roughly the same size.

Obviously the 'leak' is fake, but I like the concept- it would actually give the FMx sockets access to a bit more CPU muscle when using a dGPU. As it stands I'd never recommend anyone buy an FMx based system if they want to game due to the lack of any kind of upgrade, whereas with a hex core being available- might make more sense as you could get a lower model APU and use the iGPU initially, and have at least a reasonable upgrade path in future.

 

[szatkus]

I think a 3 module, hex core APU isn't such a bad idea to be honest. I guess the issue with APU is they have a set budget of transistors, so +1 cpu module = -1 (or 2) GPU clusters. On the other hand, given the mix of workloads, for many people 384 GPU cores + 6 CPU would be a better mix than 512 + 4. I guess the problem is that you have to produce an additional die to create that config although I think the 2 would be roughly the same size.

Obviously the 'leak' is fake, but I like the concept- it would actually give the FMx sockets access to a bit more CPU muscle when using a dGPU. As it stands I'd never recommend anyone buy an FMx based system if they want to game due to the lack of any kind of upgrade, whereas with a hex core being available- might make more sense as you could get a lower model APU and use the iGPU initially, and have at least a reasonable upgrade path in future.

Fake is at 20nm. 6+768 or 1024 would be possible.

 

[cdrkf]

I think a 3 module, hex core APU isn't such a bad idea to be honest. I guess the issue with APU is they have a set budget of transistors, so +1 cpu module = -1 (or 2) GPU clusters. On the other hand, given the mix of workloads, for many people 384 GPU cores + 6 CPU would be a better mix than 512 + 4. I guess the problem is that you have to produce an additional die to create that config although I think the 2 would be roughly the same size.

Obviously the 'leak' is fake, but I like the concept- it would actually give the FMx sockets access to a bit more CPU muscle when using a dGPU. As it stands I'd never recommend anyone buy an FMx based system if they want to game due to the lack of any kind of upgrade, whereas with a hex core being available- might make more sense as you could get a lower model APU and use the iGPU initially, and have at least a reasonable upgrade path in future.

Fake is at 20nm. 6+768 or 1024 would be possible.

Yeah true, I was thinking more about what they could do at 28nm if they re-spun Kaveri. I'm not sure we'll see any 20nm stuff next year but I'd love to be proven wrong

 

[juanrga]

@wh3resmycar, I wonder why I am asked the same questions again and again and again and again... Moreover, you mention Volta, have you checked my signature?

@colinp, you said that Excavator and Carrizo were canceled and I said were not. You were half-right and I was half-wrong: Excavator and Carrizo are coming to mobile. And your point is?

@8350rocks, highest margins in a very tiny volume market mean tiny money.

 

[juanrga]

@cdrkf, Kaveri mobile supports up to DDR3-2133 (check the FX-7600P), but I get your point and agree. Issue here will be cost. We will se lots of 20nm stuff next year, but not 4400MHz ;-)

@con635, heterogeneity is the new paradigm in chip design. AMD is pioneer with HSA, but ARM, Intel, Nvidia, IBM, and everyone else is designing new chips around heterogeneity. ARM will use HSA. Intel has its "neo-heterogeneity" approach. Nvidia relies on UVM CUDA, and IBM gave a recent talk about migrating from its traditional CPUs to heterogeneous processors with coprocessors on die and unified memory. Note the strong similarity between Nvidia UVM and AMD hUMA

The same person who said me that Kaveri will be extended to 2015 and part of 2016, said me that new HSA announcements are in the menu. I guess those will be announcements of new HSA enabled applications. I would expect apps from Adobe.

 

[juanrga]

@blackkstar, I mentioned time ago when and how FX desktop will do a comeback. I said the process node used, the number of cores, caches, memories, and other stuff. E.g. I said that minimum number of cores will be 8 (according to my sources).

JPR is one of the analysts that showed that discrete GPU market is dropping. His last 2014 reports show a new decrease in shipments of dGPUs

http://www.fudzilla.com/home/item/34800-discrete-gpu-shipments-down-in-q1
http://www.fudzilla.com/home/item/34764-intel-gains-gpu-market-share-again

Shipments of discrete GPUs were down 11.5 sequentially. Part of the drop can be attributed to seasonal trends, but shipments were also down 14.3 from last year. JPR says the long-term trend for discrete GPUs is not good. The firm forecasts a CAGR of 0.1 percent from 2014 to 2017.

Can you read the bold part? long-term trend for discrete GPUs is not good.

John predicted this about 4 years ago and the cause are guess what, integrated graphics:

The impact in the total PC and related market on discrete GPUs due to the combination of devices being offered with integrated graphics (IGPs, EPGs, and HPUs) will break the historical rise of discrete GPU sales and put the category in decline.

Can you read the bold part? integrated graphics [...] will break the historical rise of discrete GPU sales and put the category in decline.

He predicted four years ago that discrete GPUs would be reduced to a tiny 10% market share by 2015. This is a niche market. This year is more or less the year when Intel also predicts the kill of dGPUs. I could agree 2015 is when things will start to look very interesting. 2015 is also the year when Intel will start killing Nvidia and AMD GPGUs with the new Knight Landing

However, I maintain my 2020 prediction. Everyone knows that dGPUs will be killed.

 

[Embra]

Another link, didn't see this posted.
AMD’s Carrizo APU A10-8890K CPU-Z Leaked – Six Excavator Cores with 95W TDP on the 20nm Node

Read more: http://wccftech.com/amds-carrizo-apu-a10-8890k-cpuz-hexacore-20nm/#ixzz38lO8fdXa

http://wccftech.com/amds-carrizo-apu-a10-8890k-cpuz-hexacore-20nm/

 

[cdrkf]

Another link, didn't see this posted.
AMD’s Carrizo APU A10-8890K CPU-Z Leaked – Six Excavator Cores with 95W TDP on the 20nm Node

Read more: http://wccftech.com/amds-carrizo-apu-a10-8890k-cpuz-hexacore-20nm/#ixzz38lO8fdXa

http://wccftech.com/amds-carrizo-apu-a10-8890k-cpuz-hexacore-20nm/

Yeah unfortunately already confirmed a fake.

 

[gamerk316]

@blackkstar, I mentioned time ago when and how FX desktop will do a comeback. I said the process node used, the number of cores, caches, memories, and other stuff. E.g. I said that minimum number of cores will be 8 (according to my sources).

JPR is one of the analysts that showed that discrete GPU market is dropping. His last 2014 reports show a new decrease in shipments of dGPUs

http://www.fudzilla.com/home/item/34800-discrete-gpu-shipments-down-in-q1
http://www.fudzilla.com/home/item/34764-intel-gains-gpu-market-share-again

Shipments of discrete GPUs were down 11.5 sequentially. Part of the drop can be attributed to seasonal trends, but shipments were also down 14.3 from last year. JPR says the long-term trend for discrete GPUs is not good. The firm forecasts a CAGR of 0.1 percent from 2014 to 2017.

Can you read the bold part? long-term trend for discrete GPUs is not good.

John predicted this about 4 years ago and the cause are guess what, integrated graphics:

The impact in the total PC and related market on discrete GPUs due to the combination of devices being offered with integrated graphics (IGPs, EPGs, and HPUs) will break the historical rise of discrete GPU sales and put the category in decline.

Can you read the bold part? integrated graphics [...] will break the historical rise of discrete GPU sales and put the category in decline.

He predicted four years ago that discrete GPUs would be reduced to a tiny 10% market share by 2015. This is a niche market. This year is more or less the year when Intel also predicts the kill of dGPUs. I could agree 2015 is when things will start to look very interesting. 2015 is also the year when Intel will start killing Nvidia and AMD GPGUs with the new Knight Landing

However, I maintain my 2020 prediction. Everyone knows that dGPUs will be killed.

I can not wait for the second half of 2015. Seriously, do you even read the propaganda you are spouting out, or do you honestly believe what you are saying?

 

[con635]

@blackkstar, I mentioned time ago when and how FX desktop will do a comeback. I said the process node used, the number of cores, caches, memories, and other stuff. E.g. I said that minimum number of cores will be 8 (according to my sources).

JPR is one of the analysts that showed that discrete GPU market is dropping. His last 2014 reports show a new decrease in shipments of dGPUs

http://www.fudzilla.com/home/item/34800-discrete-gpu-shipments-down-in-q1
http://www.fudzilla.com/home/item/34764-intel-gains-gpu-market-share-again

Shipments of discrete GPUs were down 11.5 sequentially. Part of the drop can be attributed to seasonal trends, but shipments were also down 14.3 from last year. JPR says the long-term trend for discrete GPUs is not good. The firm forecasts a CAGR of 0.1 percent from 2014 to 2017.

Can you read the bold part? long-term trend for discrete GPUs is not good.

John predicted this about 4 years ago and the cause are guess what, integrated graphics:

The impact in the total PC and related market on discrete GPUs due to the combination of devices being offered with integrated graphics (IGPs, EPGs, and HPUs) will break the historical rise of discrete GPU sales and put the category in decline.

Can you read the bold part? integrated graphics [...] will break the historical rise of discrete GPU sales and put the category in decline.

He predicted four years ago that discrete GPUs would be reduced to a tiny 10% market share by 2015. This is a niche market. This year is more or less the year when Intel also predicts the kill of dGPUs. I could agree 2015 is when things will start to look very interesting. 2015 is also the year when Intel will start killing Nvidia and AMD GPGUs with the new Knight Landing

However, I maintain my 2020 prediction. Everyone knows that dGPUs will be killed.

I can not wait for the second half of 2015. Seriously, do you even read the propaganda you are spouting out, or do you honestly believe what you are saying?

Whats so hard to believe? They are going to be integrated soon like everything,just a matter of time, juans not saying we'll all suddenly throw our powerful gpus/cpus in the bin for a ddr3 igp and weak cpu, hes saying a powerful gpu will share a space with a powerful cpu in the future and there will no one making discrete parts in the future.
Maybe if everyone just keeps posting 7850k vs r9 290x benchmarks integration will not happen....
Seriously, once memory bandwidth is sorted I can see everything under r9 270 becoming 'legacy'

 

[colinp]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

 

[cdrkf]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

In the kinda time frame were talking, we went from gaming at 800 x 600 on high end dedicated hardware, to gaming at 1920 x 1080p on an integrated chip.

We went from MB of ram to the same in GB, we went from 1 core at around 1ghz to 8 cores @ 4+ ghz

This isn't going to happen next year, but within another decade computing could look very very different Also bear in mind, a single SOC in a phone can now drive 1920 x 1080p in basic 3-D games, and Nvidia Shield shows that even some 'proper' PC titles can be run on this circa 2w hardware (e.g. Halflife 2 and Portal).

 

[juanrga]

Spoiler

@blackkstar, I mentioned time ago when and how FX desktop will do a comeback. I said the process node used, the number of cores, caches, memories, and other stuff. E.g. I said that minimum number of cores will be 8 (according to my sources).

JPR is one of the analysts that showed that discrete GPU market is dropping. His last 2014 reports show a new decrease in shipments of dGPUs

http://www.fudzilla.com/home/item/34800-discrete-gpu-shipments-down-in-q1
http://www.fudzilla.com/home/item/34764-intel-gains-gpu-market-share-again

Shipments of discrete GPUs were down 11.5 sequentially. Part of the drop can be attributed to seasonal trends, but shipments were also down 14.3 from last year. JPR says the long-term trend for discrete GPUs is not good. The firm forecasts a CAGR of 0.1 percent from 2014 to 2017.

Can you read the bold part? long-term trend for discrete GPUs is not good.

John predicted this about 4 years ago and the cause are guess what, integrated graphics:

The impact in the total PC and related market on discrete GPUs due to the combination of devices being offered with integrated graphics (IGPs, EPGs, and HPUs) will break the historical rise of discrete GPU sales and put the category in decline.

Can you read the bold part? integrated graphics [...] will break the historical rise of discrete GPU sales and put the category in decline.

He predicted four years ago that discrete GPUs would be reduced to a tiny 10% market share by 2015. This is a niche market. This year is more or less the year when Intel also predicts the kill of dGPUs. I could agree 2015 is when things will start to look very interesting. 2015 is also the year when Intel will start killing Nvidia and AMD GPGUs with the new Knight Landing

However, I maintain my 2020 prediction. Everyone knows that dGPUs will be killed.

I can not wait for the second half of 2015. Seriously, do you even read the propaganda you are spouting out, or do you honestly believe what you are saying?

Whats so hard to believe? They are going to be integrated soon like everything,just a matter of time, juans not saying we'll all suddenly throw our powerful gpus/cpus in the bin for a ddr3 igp and weak cpu, hes saying a powerful gpu will share a space with a powerful cpu in the future and there will no one making discrete parts in the future.
Maybe if everyone just keeps posting 7850k vs r9 290x benchmarks integration will not happen....
Seriously, once memory bandwidth is sorted I can see everything under r9 270 becoming 'legacy'

Note how I mentioned I maintain my prediction for the year 2020, but he he decided to comment about the year 2015. And when the year 2015 was here, my words will be twisted, falsified, misquoted, ignored...

 

[juanrga]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

It is rather irrelevant if the answer is 2018, 2025, or 2030. The question is that we know that by 2020 APUs will be about 10x faster than GPUs. You can purchase the slow and power hungry system, I prefer the fastest and efficient system, thanks!

Next year, Intel will launch a 'CPU' with a performance between 1.5x and 4x (depending on application) above the faster discrete card by Nvidia today: the $5000 Tesla K40

'CPU' KL = 1.5--4x dGPU K40

 

[Fidgetmaster]

I will believe when I see it...haha

I don't know man you have made some wild prediction's haha

 

[8350rocks]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

It is rather irrelevant if the answer is 2018, 2025, or 2030. The question is that we know that by 2020 APUs will be about 10x faster than GPUs. You can purchase the slow and power hungry system, I prefer the fastest and efficient system, thanks!

Next year, Intel will launch a 'CPU' with a performance between 1.5x and 4x (depending on application) above the faster discrete card by Nvidia today: the $5000 Tesla K40

'CPU' KL = 1.5--4x dGPU K40

Your bolded section will never happen. A dGPU will always be faster than iGPU...why? Transistor budget.

You can make a massive die with 7-8 bil transistors or more for a dedicated dGPU (a la the Maxwell cards coming and volcanic/pirate islands flagship SKUs) that will do 10 times the work that a SoC can do. It will always be that way because there are these laws of physics that they have to play within.

Sorry, your prediction will be wrong. Because dGPUs will always be more powerful than iGPUs. You cannot defy physics. The same die space dedicated to full on GPU will be more powerful.

 

[blackkstar]

Juan, I think you are confusing performance per watt for raw performance. If dGPU and dCPU die off, I better be able to put 2 or 4 APUs in a system and have them work as one unified device (like 2m/4c CPU * 4 for 8m/16c CPU + 10CU APU * 4 for 40 CU GPU). There are a lot of other content creators like me who need a lot of power who feel the same.

That's the only way I see APUs winning entirely. They are going to have to offer something vastly superior to dCPU + dGPU and because of transistor budget, I don't see that happening. Yeah you can have a system with 4 APUs under the situation I gave you, but you can also have 4 big dGPUs if you need the power. So I don't see it unless you're doing something like running HSA. And then you run into the problem of having a system that is slower than comparable systems in traditional workloads and much faster with HSA. So if something in your toolchain isn't HSA enabled, it's probably going to be a waste unless you can afford multiple systems.

All of this compiling, rendering, transcoding, etc all form a massive part of the internet.

 

[gamerk316]

Next year, Intel will launch a 'CPU' with a performance between 1.5x and 4x (depending on application) above the faster discrete card by Nvidia today: the $5000 Tesla K40

'CPU' KL = 1.5--4x dGPU K40

No. Just no.

 

[Cazalan]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

Real time ray tracing with 4 Knights Ferry cards.

http://www.tomshardware.com/news/wolfenstein-ray-tracing-knights-ferry,11282.html

Going to be a while for what you're asking for. So many factors need to line up.

450mm wafers are key to getting costs down but that's been delayed several times.
The next market "crash" is due to happen at any time, which may put more companies out of business.

 

[gamerk316]

When do we think that an APU will be able to deliver photo-realistic rendering with fully modelled physics and AI at 4k resolution in 3d at 60fps, all at less than 100w TDP?

Real time ray tracing with 4 Knights Ferry cards.

http://www.tomshardware.com/news/wolfenstein-ray-tracing-knights-ferry,11282.html

Going to be a while for what you're asking for. So many factors need to line up.

450mm wafers are key to getting costs down but that's been delayed several times.
The next market "crash" is due to happen at any time, which may put more companies out of business.

That being said, Ray Tracing (or in the short term, Ray Casting) is the obvious next step, since there really aren't many more cheap ways to improve image quality anymore. Notice all the focus on new AA modes, physics, multi-display, and resolution the past few years? That's because there really isn't much that you can do to improve IQ anymore.

Look at GTA:SA, and compare it to Saints Row or GTA IV. Obvious and significant IQ improvements. Compare GTA V to Watch Dogs. Not so much. Heck, compare GTA IV to watch dogs if you want. Graphically, we're tapped out, focusing more now on sharpening the image and making the more subtle effects more accurate. But there isn't anything huge in the works right now.

 

[juanrga]

8350rocks, the dGPU will have less than 10% more of transistors, but less performance due to interconnect bottlenecks. What are you arguing here?

blackkstar, am I discussing TFLOPS or TFLOPS/Watt? Hint: I are mentioning "40 TFLOPS APUs"?

 

[gamerk316]

8350rocks, the dGPU will have less than 10% more of transistors, but less performance due to interconnect bottlenecks. What are you arguing here?

Please. dGPUs have a while to grow yet, especially with at least 2-3 more node shrinks to follow. ((That being said, EVERYONE is going to stall at ~6nm or so.))

And again: Interconnects are not a bottleneck. They're actually an advantage, since you can just stuff lots of faster-then-system-RAM directly on the GPU die to create a very large high speed cache. The only downside to the interconnect is the latency transferring the data over into the cache, and even that is minimal. Meanwhile, APUs are limited by the CPU IMC and slow system RAM.

 

[esrever]

8350rocks, the dGPU will have less than 10% more of transistors, but less performance due to interconnect bottlenecks. What are you arguing here?

Please. dGPUs have a while to grow yet, especially with at least 2-3 more node shrinks to follow. ((That being said, EVERYONE is going to stall at ~6nm or so.))

And again: Interconnects are not a bottleneck. They're actually an advantage, since you can just stuff lots of faster-then-system-RAM directly on the GPU die to create a very large high speed cache. The only downside to the interconnect is the latency transferring the data over into the cache, and even that is minimal. Meanwhile, APUs are limited by the CPU IMC and slow system RAM.

The way of thinking is quite archaic, anything you can do with memory on a gpu you can do on an apu. Interconnects are pretty much always bottleneck, which is why everything is being integrated on die. Anyways, by the time we have stuff like this, the whole chip would be different. We would be able to stack high speed ram on die by then I would hope.

 

[cdrkf]

8350rocks, the dGPU will have less than 10% more of transistors, but less performance due to interconnect bottlenecks. What are you arguing here?

Please. dGPUs have a while to grow yet, especially with at least 2-3 more node shrinks to follow. ((That being said, EVERYONE is going to stall at ~6nm or so.))

And again: Interconnects are not a bottleneck. They're actually an advantage, since you can just stuff lots of faster-then-system-RAM directly on the GPU die to create a very large high speed cache. The only downside to the interconnect is the latency transferring the data over into the cache, and even that is minimal. Meanwhile, APUs are limited by the CPU IMC and slow system RAM.

For APU's to grow they will needed dedicated memory on die, no doubt. I think that is likely to happen sooner rather than later, after all Kaveri would match a 7750 for graphics oomph if it just had enough memory bandwidth!