AMD CPU speculation... and expert conjecture

[juanrga]

Oooops! I was expect Zen/K12 to bring a giant ~70% performance gain over Jaguar (~40% IPC plus higher frequencies). Some as gamerk said "this will not happen". Well, I was considering a best case possible scenario for AMD.

But ARM has just announced the new A72 with promises of more than 80% gains over the A57 (~50% IPC plus ~30% higher frequencies).

http://techreport.com/news/27767/arm-unveils-cortex-a72-cpu-mali-t880-graphics-and-more

Just note that the Cortex A57 is already faster than Jaguar

http://www.anandtech.com/show/7724/it-begins-amd-announces-its-first-arm-based-server-soc-64bit8core-opteron-a1100

My worst dream seems to convert on reality: AMD will be crushed by both sides ARM and x86.

 

[-Fran-]

I don't get your post, Juan.

If ARM Can get that 80%, why can't AMD? Did you get more inside information about the design differences or something?

Cheers!

 

[8350rocks]

I don't get your post, Juan.

If ARM Can get that 80%, why can't AMD? Did you get more inside information about the design differences or something?

Cheers!

Juan's expert source is really some guys in the mail rooms of the marketing departments at 2 or 3 different companies that send him the propaganda, which then becomes "Juan's Law".

 

[noob2222]

To Anyone

If you rewrite a program to reduce the number of instructions for a given task, did you increase the instructions that any given cpu is capable of doing?

what is soo hard to grasp here?

@ juan

stop with the AMD hate posts or go somewhere else. take your ARM crusade and make your own thread. seriously, remember how cortex a50 was going to be 3x faster?

http://www.engadget.com/2012/10/30/arm-cortex-a50/

 

[con635]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

mystery card is 35% faster while using 61% more power consumption compared to a 980. For some reason the equalized everything to a 970 not sure why they did that unless its going to be priced at that range.

Really wanting to see if they can make a 200 watt card to compete with a 980 in terms of performance if so i might make the switch.

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Its fake anyway, the 3dmark score is the same one proved to be 2x r9siirc

 

[Embra]

mystery card is 35% faster while using 61% more power consumption compared to a 980. For some reason the equalized everything to a 970 not sure why they did that unless its going to be priced at that range.

Really wanting to see if they can make a 200 watt card to compete with a 980 in terms of performance if so i might make the switch.

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

I recall this card is suppose to have a custom cooler (liquid), similar to what the 295x2 has. This would also comsume more power, just like adding a custom cooler to any card.

 

[gamerk316]

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Design sacrifice. pcper did a great analysis on the subject.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Looking-GTX-970-Memory-Performance
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-GTX-970-Memory-Issued-Tested-SLI

 

[gamerk316]

Transistors have now been made out of Silicene:

http://arstechnica.com/science/2015/02/first-transistor-built-using-two-dimensional-silicon/

Ways to go, but making progress.

 

[blackkstar]

If the new AMD cards are using HBM and something besides 28nm, Titan 2 will not be competitive. Nvidia might have something at 1080p if they aim for 144hz, but outside of that, it'll get destroyed, specially at 4K or surround/eyefinity. And if Bermuda and Fiji are on a smaller node, Nvidia may not be competitive with dollar per square mm per die, which means they won't be able to compete on price.

I'd expect 2015 to be a year of Nvidia losing in GPU performance. When they lose, they just talk about how great they think their software is and how great it is to have PhysX for the one good PhysX game that comes out every 3 years. 2015 will be the year of the Nvidia gimmicks. G-Sync, CUDA, PhysX, Gameworks, etc. They'll push these things so hard because that's all they have when they lose on hardware. However, it's going to be a difficult fight. AMD is basically going to go "look at all this cool stuff you can do with Mantle! Star Swarm can display so many more units on screen, you can use multiple GPUs far better, you can use your APU iGPU for other calculations, etc" and Nvidia will go "but Gameworks has some cool shader based anti aliasing and we have cool particles with PhsyX!"

Fottemberg at S|A is claiming Broadwell-K may be flat out cancelled. If that's the case, it'll be Haswell vs Piledriver on HEDT for all of 2015. If the problems are with 14nm and not the Broadwell architecture, AMD may be in a position to pit Zen against Haswell in HEDT. If Zen is comparable to Haswell in single core performance and there's an 8 core Zen, Haswell is going to have a lot of work cut out for itself.

AMD having a platform with SKUs that range from $100 budget chips to $500 8 core Zen, they'd completely destroy Intel in HEDT. Intel segmenting HEDT into enthusiast and "HEDT" might be something they massively regret in 2016.

I realize some of you think HEDT is a complete waste, but Lisa Su has already stated there's demand for high performance x86 cores from semi-custom clients in the last quarter interview. HEDT is the dumping ground for reject parts.

If AMD could manage to get a Zen core to be about a Haswell core, and can deliver 8 core Zen for ~$500, it'll be a very bad year for Intel. I hope AMD has figured some way to share some core resources to significantly reduce die size without sacrificing as much as they did with the Bulldozer family. Seeing AMD match Intel performance with a smaller die would be K7 and K8 all over again.

I am really optimistic with AMD opportunities. It seems like Nvidia is fumbling with VRAM and now Intel might be fumbling with HEDT.

Also, in before the regular claims from the usual suspects that "Broadwell-K is cancelled because of lack of demand!"

 

[szatkus]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

mystery card is 35% faster while using 61% more power consumption compared to a 980. For some reason the equalized everything to a 970 not sure why they did that unless its going to be priced at that range.

Really wanting to see if they can make a 200 watt card to compete with a 980 in terms of performance if so i might make the switch.

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Its fake anyway, the 3dmark score is the same one proved to be 2x r9siirc

The score is different. http://img1.mydrivers.com/img/20150119/69c24d86a978426aa0ce276915f4f401.jpg

Fake or not, looks plausible.

 

[szatkus]

Fottemberg at S|A is claiming Broadwell-K may be flat out cancelled. If that's the case, it'll be Haswell vs Piledriver on HEDT for all of 2015. If the problems are with 14nm and not the Broadwell architecture, AMD may be in a position to pit Zen against Haswell in HEDT. If Zen is comparable to Haswell in single core performance and there's an 8 core Zen, Haswell is going to have a lot of work cut out for itself.

Vishera isn't even HEDT :>

 

[jdwii]

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Design sacrifice. pcper did a great analysis on the subject.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Looking-GTX-970-Memory-Performance
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-GTX-970-Memory-Issued-Tested-SLI

tbh they should have called it a 3.5GB card with a 512MB cache or something or just disable the memory. Would have been better then not telling a single person at all about the slow memory. Nvidia has no reason to lie same as Intel they are both leaders in what they do. Its like Intel lying about the I7 4790K OC abilities its not like they really have a good reason to lie.

 

[jdwii]

"I'd expect 2015 to be a year of Nvidia losing in GPU performance. When they lose, they just talk about how great they think their software is and how great it is to have PhysX for the one good PhysX game that comes out every 3 years."

Wait isn't that kinda like Amd for the last few years? HSA, cough mantle support in loads of games cough cough. Then lets through in the so called Opencl bit. They both troll us. I have to say however i used HBAO more then any Amd feature, and i seen it in many games.

 

[gamerk316]

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Design sacrifice. pcper did a great analysis on the subject.

http://www.pcper.com/reviews/Graphics-Cards/NVIDIA-Discloses-Full-Memory-Structure-and-Limitations-GTX-970
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-Looking-GTX-970-Memory-Performance
http://www.pcper.com/reviews/Graphics-Cards/Frame-Rating-GTX-970-Memory-Issued-Tested-SLI

tbh they should have called it a 3.5GB card with a 512MB cache or something or just disable the memory. Would have been better then not telling a single person at all about the slow memory. Nvidia has no reason to lie same as Intel they are both leaders in what they do. Its like Intel lying about the I7 4790K OC abilities its not like they really have a good reason to lie.

I could just as much call a FX-8350 a quad core with a form of Hyperthreading. Everything is in the eye of the beholder.

At the end of the day, the issue is minor compared to the cards price/performance. Not like AMD has anything to counter it with.

 

[juanrga]

If ARM Can get that 80%, why can't AMD?

By the same reason(s) Intel can extract more performance from x86 than AMD.

seriously, remember how cortex a50 was going to be 3x faster?

http://www.engadget.com/2012/10/30/arm-cortex-a50/

Your link did claim "ARM's Cortex-A50 chips promise 3x performance of current superphones". And that is right. The A57 is up to 2x faster than A15 clock per clock on certain applications. If you compare a high clocked A57 against a low clocked A7 or so you can get 3x. No sure what is your problem.

What ARM is claiming now is different. Check the link given. They are giving concrete performance gains for concrete core models at concrete frequencies.

The new A72 implies that several AMD products based in the A57 (Seattle, Hierofalcon...) will be seriously outdated next year.

ARM claims that one of the "Key target markets" for the new A72 are "servers". This means that AMD's K12 will have to fight another competitor.

 

[con635]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

mystery card is 35% faster while using 61% more power consumption compared to a 980. For some reason the equalized everything to a 970 not sure why they did that unless its going to be priced at that range.

Really wanting to see if they can make a 200 watt card to compete with a 980 in terms of performance if so i might make the switch.

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Its fake anyway, the 3dmark score is the same one proved to be 2x r9siirc

The score is different. http://img1.mydrivers.com/img/20150119/69c24d86a978426aa0ce276915f4f401.jpg

Fake or not, looks plausible.

There was another 'leak', cant find it now, I think its more an educated guess than a leak, I actually think its plausible that the 380x could be better. The extra cores should net around 40% over 290x alone without architectural improvements them or hbm.

 

[juanrga]

If the new AMD cards are using HBM and something besides 28nm, Titan 2 will not be competitive. Nvidia might have something at 1080p if they aim for 144hz, but outside of that, it'll get destroyed, specially at 4K or surround/eyefinity. And if Bermuda and Fiji are on a smaller node, Nvidia may not be competitive with dollar per square mm per die, which means they won't be able to compete on price.

Known numbers show that Titan 2 is faster than 390X at 1440p and neck-to-neck at 4K.

The smaller node, 20nm, is more expensive than 28nm, because requires multiple patterning. Add that HBM is expensive memory and AMD will have two options (i) to sell the new cards at high prices or (ii) to sell them at a discount, loosing the money that don't have.

Fottemberg at S|A is claiming Broadwell-K may be flat out cancelled. If that's the case, it'll be Haswell vs Piledriver on HEDT for all of 2015. If the problems are with 14nm and not the Broadwell architecture, AMD may be in a position to pit Zen against Haswell in HEDT. If Zen is comparable to Haswell in single core performance and there's an 8 core Zen, Haswell is going to have a lot of work cut out for itself.

AMD having a platform with SKUs that range from $100 budget chips to $500 8 core Zen, they'd completely destroy Intel in HEDT. Intel segmenting HEDT into enthusiast and "HEDT" might be something they massively regret in 2016.

Updated roadmaps show Skylake-K 95W coming in Q3 this year. This means that the competition will be Piledriver (32nm) vs Skylake (14nm) for last 2015.

Zen will be not ready before the end of 2016 or first 2017. In my best scenario for AMD I was expecting Zen to fight Skylake, but Intel has accelerated its roadmaps and want to launch 10nm by early 2017.

Thus Zen will fight Canonlake mostly.

I realize some of you think HEDT is a complete waste, but Lisa Su has already stated there's demand for high performance x86 cores from semi-custom clients in the last quarter interview. HEDT is the dumping ground for reject parts.

During last call, Lisa Su named "high-performance CPUs" to 35W CPUs for semicustom clients. She uses the term "high-performance" even when refers to AMD products based in phone-like cores. She doesn't use the term as synonym for HEDT.

If AMD could manage to get a Zen core to be about a Haswell core, and can deliver 8 core Zen for ~$500, it'll be a very bad year for Intel. I hope AMD has figured some way to share some core resources to significantly reduce die size without sacrificing as much as they did with the Bulldozer family. Seeing AMD match Intel performance with a smaller die would be K7 and K8 all over again.

I am really optimistic with AMD opportunities. It seems like Nvidia is fumbling with VRAM and now Intel might be fumbling with HEDT.

Also, in before the regular claims from the usual suspects that "Broadwell-K is cancelled because of lack of demand!"

People said me how Bulldozer was going to crush Intel on HEDT.
People said me how Piledriver was going to fix any problem on Bulldozer.
People said me how Steamroller was going to crush Intel on HEDT.
People said me how Excavator was going to crush Intel on HEDT.
Now people says me how Zen is going to crush Intel on HEDT.
People will say me how post-Zen will crush Intel on HEDT.
...

 

[szatkus]

I can't tell the difference between a FX8320 and a A6-3620 either without benchmarks. The difference between CPUs is so small now I doubt anyone can really tell an i7 from a pentium without running benchmarks. Even in most games, if you have a powerful enough GPU, there really isn't that much difference.

mystery card is 35% faster while using 61% more power consumption compared to a 980. For some reason the equalized everything to a 970 not sure why they did that unless its going to be priced at that range.

Really wanting to see if they can make a 200 watt card to compete with a 980 in terms of performance if so i might make the switch.

We need Amd more then ever right now with Nvidia abusing the market just a bit with this 3.5GB card nonsense.

Its fake anyway, the 3dmark score is the same one proved to be 2x r9siirc

The score is different. http://img1.mydrivers.com/img/20150119/69c24d86a978426aa0ce276915f4f401.jpg

Fake or not, looks plausible.

There was another 'leak', cant find it now, I think its more an educated guess than a leak, I actually think its plausible that the 380x could be better. The extra cores should net around 40% over 290x alone without architectural improvements them or hbm.

"50% better, 5% more power consumption", right?

 

[con635]

http://www.tomshardware.com/news/nvidia-samsung-14nm-apple-qualcomm,28493.html
Sorry if repost, interesting 'fight' next year!

 

[szatkus]

People said me how Bulldozer was going to crush Intel on HEDT.

That was the plan

People said me how Piledriver was going to fix any problem on Bulldozer.

It fixed a lot. FXs went from "bad" to "meh".

People said me how Steamroller was going to crush Intel on HEDT.

I remember you saying that. My opinion were closer to reality...

People said me how Excavator was going to crush Intel on HEDT.

I don't remember anyone...

Now people says me how Zen is going to crush Intel on HEDT.

I think it will be rather as competetive as Phenom II rather than K7-level crush.

People will say me how post-Zen will crush Intel on HEDT.

Hell, yeah!

 

[logainofhades]

The fact that AMD has yet to truly crush its own Phenom II processors, I do not see them crushing Intel anytime soon.

PhII X4 and X6 is faster clock for clock than Faildozer 4xxx and 6xxx. It is even, at best with Piledriver versions. Steamroller it might be about even to slightly faster. The lack of L3 cache can be quite the handicap. I have 0 hope for Excavator. The only reason FX can win against PhII, with the same amount of cores, is the higher clock speeds that the FX chips can reach.

Zen I truly hope does well. AMD has hopefully learned from their mistakes, like Intel did with crapburst, and will produce a truly competitive chip.

 

[etayorius]

Zen I truly hope does well. AMD has hopefully learned from their mistakes, like Intel did with crapburst, and will produce a truly competitive chip.

I have a feeling they have not learned their lesson...

Hey juan, can you post the link to the quote by AMD Engineer saying the goal was to achieve 100% more ipc over Bulldozer?

In my opinion, the best case sceneario for ZEN will be close to Haswell IPC level, i don`t even expect them to beat Haswell... and actually they don`t have to, just get as close as they can and sell it cheap, i would buy one but i will not wait till Zen arrives.

 

[juanrga]

I have 0 hope for Excavator.

Excavator will bring about 10--15% IPC gains over Steamroller. The lowest Carrizo models will receive an frequency upgrade and could perform up to 30% faster than Steamroller-based Kaveri of similar TDP.

However, Excavator hits a frequency wall of about 3GHz. This means that the top 35W Carrizo models will see the IPC gains compensated by a reduction on clocks. You are advised guys, the top 35W Carrizo could perform poor than the 35W Kaveri FX-7600P on certain benchmarks.

Excavator is a smaller and efficient core than Steamroller. The new core is about 30% smaller and has less L2 cache.

 

[logainofhades]

Unless you really need the better IGP, sounds like a 4130T would be a better buy, if talking desktop.

 

[juanrga]

People said me how Bulldozer was going to crush Intel on HEDT.

That was the plan

And we know its zero resemblance to reality.

People said me how Piledriver was going to fix any problem on Bulldozer.

It fixed a lot. FXs went from "bad" to "meh".

"A lot" != "any"

People said me how Steamroller was going to crush Intel on HEDT.

I remember you saying that. My opinion were closer to reality...

I remember saying "nope" to the person what said it.

People said me how Excavator was going to crush Intel on HEDT.

I don't remember anyone...

I do.

Now people says me how Zen is going to crush Intel on HEDT.

I think it will be rather as competetive as Phenom II rather than K7-level crush.

I think will be good enough for their embedded/semicustom business and we will see consoles and some dense servers based in them.

People will say me how post-Zen will crush Intel on HEDT.

Hell, yeah!

And if AMD is still open by then, people will say me how post-post-Zen will crush Intel on HEDT.