AMD CPU speculation... and expert conjecture

[juanrga]

I already announced before how Intel is following AMD strategy of breaking exclusivity with Windows and focusing in linux.

Now Intel has announced that Broadwell doesn't come to desktop, only to mobile/tablets...

http://www.fudzilla.com/home/item/32524-broadwell-won%E2%80%99t-make-it-to-desktop

which matches with AMD strategy of focusing on APUs

Here confirmation of AMD plans for the next gen of Hawai GPUs

http://www.fudzilla.com/home/item/32530-amd-talks-up-hawaii

Again agrees with my view AMD will be not releasing ultra-high cards for a 0.05% of the market. As a consequence, the arguments that I was said, about AMD needing an ultra super-high mega-powerful new CPU to match those future GPUs seems to vanish in the air.

Broadwell was never destined to be a DT part. Its going to be hot as fk trying to get to 3.5 ghz. Your being fed marketing BS to make up for the fact that 14nm is extremely dense and too much to keep cool. Shrinking from here on out will only be focused on mobile parts. This is why Intel wants to kill DT systems all together, so they can justify shrinking to 5nm ... lol.

Of course conspiracy theory at its best. Hard facts such as the desktop market is dying don't count a bit.

as for your hawaii bs, 30% smaller die size doesn't equate to 30% slower. Obviously you didn't even read the link you provided..

Nobody said or insinuated a 30% slower anything. Again you are answering to stuff only in your imagination.

 

[8350rocks]

The reason why people don't see AMD going APU + dGPU is that there's too much power to fit into a socket. BF4 recommends AMD 3m/6c CPU. If AMD goes APU only with maximum of 2m/4c, then they won't even have a suitable gaming chip for games like BF4.

Wait a moment, because BF4 recommends a six core Piledriver CPU. As I have 'shown' in my article about Kaveri, the four core Steamroller APU is essentially at the i5 / FX-6000 level of performance.

Moreover, BF4 is not still next gen. Next gen will offload the heavy computations to the GPU, which implies that CPU will perform still better for the remaining tasks: logic, gameplay...

Whereas I agree with most of the pcper article, I find highly unlikely that AMD will release a six core APU by the end of 2014.

Because 3M = 12 ALUs/3 FPUs in AMD and i5 = 12 ALUs/4 FPUs in Intel...

Even if you make fewer ALUs faster, you're still behind. Look at the i3 with it's 6 ALUs and HTT...it doesn't run BF3 multiplayer very well at all...much less something like Crysis 3 at frame rates worth bothering over (30 FPS with GTX 690 vs. 42 FPS with GTX 690 from FX 6300).

 

[8350rocks]

I committed a sin and got Intel stuff, got Battlefield microstutters, *scratches head* when I made this cash strapping investment it was made on the premise of how awesome Intel is supposed to be, I personally think as always, its a load of hot air and if you are not providing me random synthetics in the real world there actually is very little to tell between a 4770Krap and a APU. Anyways I am a little perplexed and disappointed, it may just be the shame of actually supporting what I universally detest or the fact that I burned a huge hole in the wallet.


PS: saw a lot of bitching on the Memory bandwidth article, why is there no HD5000's, well how is HD4000 and old intel craphics which have shown no scaling going to change with HD5000......*straws clutch grab em now*

This says everything I have been saying for a while about Intel vs. AMD. Synthetic benchmark king (INTEL) *wants* you to believe that it makes a difference with comparable hardware...

It doesn't...so the joke is on the Intel fanboys who shill out more money for the same performance.

 

[hcl123]

@juanrga:

I brought up QPI because Cray could have used it, but didn't. PCIe 3.0 is only a useful setup for Intel type systems, as HT *already* provides that much throughput. However, Intel doesn't do anything that interfaces with HT and so it would require a completely new hardware solution from the ground up costing lots of money and time resources.

Additionally, if IBM are going to PCIe 3.0 it's the first I've heard of it. They were using a HT based fabric in their large server clusters...and it's capable of 51.2 GB/s of throughput bandwidth...you'd have to look long and hard at another solution to yield that much and do some heavy tweaking. The *only* reason anyone would look at PCIe 3.0 over HT would be simply to take advantage of PCIe 3.0 GPGPUs; though, I cannot make a case for that at this point in time, as most GPGPUs cannot currently entirely saturate PCIe 2.0 x16 (note: I said most, as there may be one out there that can, even if I haven't heard of it).

As I already said Cray did look for a processor agnostic interconnect. This is the reason why are not using Intel QPI.

If IBM was using a "51.2 GB/s of throughput bandwidth" HT interconnect then the switch to PCIe 3 (96 GB/s of throughput) makes sense. In fact the numbers coincide with what IBM said at hot chips. IBM claimed that the new interconnect in the Power8 offers 2x the bandwidth of the interconnects used in Power7 and Power7+

Not only IBM is going PCIe 3 with the new Power8, but Oracle, Cray, AMD... as well

Only intel !... lol

If you noticed on the link you provide 96GB/s is for 96 lanes, and since its a serial "differential" interconnect, each link has 2 lanes ( 1 outbound, 1 inbound)... so a 16x link has ~32GB/s a 48x link is the one who has ~96GB/s, and this must be the max that a PCIe controller allows, a 48x link... and yes HTT is max 51.2 GB/s but is because its limited to 32x links... Yet PCIe v3 has still considerable more overhead, so from the theoretical max peak throughput to real realizable throughput there can be a notorious difference.

IBM has there own things that are clearly superior to anything PCIe... PCIe is used because if they want expansion to other "device" boards they have no other choice, unless they make themselves all the device expansion needed (not happening)... HTT was and is used because if they wanted "co-processor" boards HTT is the only one available... though now they invented a new business to sell special PCIe v3 hardware cache coherence for their Power 8 that others must license (royalties auch!) to use, sensible since their business is really to make money lol.

In the same logic Cray doesn't have QPI because they must license it ( royalties would be real painful) from Intel... and they are also in the dire need to be in business of making money not waste it. (agnostic is for religion lol)

 

[noob2222]

I already announced before how Intel is following AMD strategy of breaking exclusivity with Windows and focusing in linux.

Now Intel has announced that Broadwell doesn't come to desktop, only to mobile/tablets...

http://www.fudzilla.com/home/item/32524-broadwell-won%E2%80%99t-make-it-to-desktop

which matches with AMD strategy of focusing on APUs

Here confirmation of AMD plans for the next gen of Hawai GPUs

http://www.fudzilla.com/home/item/32530-amd-talks-up-hawaii

Again agrees with my view AMD will be not releasing ultra-high cards for a 0.05% of the market. As a consequence, the arguments that I was said, about AMD needing an ultra super-high mega-powerful new CPU to match those future GPUs seems to vanish in the air.

Broadwell was never destined to be a DT part. Its going to be hot as fk trying to get to 3.5 ghz. Your being fed marketing BS to make up for the fact that 14nm is extremely dense and too much to keep cool. Shrinking from here on out will only be focused on mobile parts. This is why Intel wants to kill DT systems all together, so they can justify shrinking to 5nm ... lol.

Of course conspiracy theory at its best. Hard facts such as the desktop market is dying don't count a bit.

I love when people have no real answer they throw this junk ... OMG CONSPIRACY . <-- translation: I don't have an answer to that so im going to claim CONSPIRACY.

I said it when IVY came out that 14nm on DT would be hard, 10nm nearly impossible. 32 to 22nm increased heat output by 20c The only thing conspriacy theorists could claim is "omg thermal paste not solder, heat density is a myth".

Here is a clue:

If you want a certain market to slow down, what do you do to make that happen?

Answer:

Spoiler

don't give people a reason to upgrade. When was the last time Intel made a major improvement in DT cpus? OMG CONSPIRACY!!!!!!

 

[hcl123]

Except Charlie is wrong.

The SeaMicro Freedom Fabric interconnect included in AMD new Seattle chip is "high-performance" and provides up to 160GB/s.

The Freedom Fabric bandwidth is far above the best than both PCIe Gen 3 and Hyperthreading can offer us. For the sake of comparison the fastest Hyperthreading version tops to about 50GB/s.

ummm... can't remember but for sure a Seamicro cluster can have up to 512 nodes, each with a CPU/APU in it, so 160GB/s is from the concentrator ASIC not for each node, but it could be for all... meaning if all nodes access network and or storage all at the same time, the bandwidth will be limited to 160/512 = 81.2MB/s for each...

Perhaps that is why they are limited to 256 nodes now, PCIe v2 doesn't help at all...and perhaps that is why each node board has provisions for a dedicated SSD for each (like a storage big cache/buffer).

512x 160GB/s = 81.9TeraB/s.... that probably would make the ASIC more power hungry than the all cabinet itself lol... can never be per node (its shared).

 

[hcl123]

as for your hawaii bs, 30% smaller die size doesn't equate to 30% slower. Obviously you didn't even read the link you provided..

Nobody said or insinuated a 30% slower anything. Again you are answering to stuff only in your imagination.

Now that is relevant, since the Hawaii is ~30% smaller than big kepler, it means it is around the 400mm² mark (a little more), meaning it is less than 2500sp for sure, making GCN clearly ahead of Nvidia in the ALU punch, if the performance is identical or superior to titan.

 

[hcl123]

I love when people have no real answer they throw this junk ... OMG CONSPIRACY . <-- translation: I don't have an answer to that so im going to claim CONSPIRACY.

I said it when IVY came out that 14nm on DT would be hard, 10nm nearly impossible. 32 to 22nm increased heat output by 20c The only thing conspriacy theorists could claim is "omg thermal paste not solder, heat density is a myth".

Here is a clue:

If you want a certain market to slow down, what do you do to make that happen?

Answer:

Spoiler

don't give people a reason to upgrade. When was the last time Intel made a major improvement in DT cpus? OMG CONSPIRACY!!!!!!

ummm... here is another BIG CONSPIRACY http://www.fudzilla.com/home/item/32507-intel-chief-engineer-talks-haswell%E2%80%99s-5-year-journey

Is it me or Fuad is relating what a chief intel engineer told ... Broadwell transistors will be 30% smaller ???

How come !?... from 22 to 14nm according to past standards should be something in the region of 70%... 30% ?, Broadwell is not even real 16nm, is more like a good shrink on 20nm node... as if 19 or 18nm...

ummm... perhaps intel is leaving the traditional DT only for the 6 core chips, not a mad race to smaller and low power... now that they are REAL 6 core chips, perhaps Nvidia will not be dead soon, still a chance, otherwise disgusting i really regretted if Nvidia would be in big trouble (users need more competitors not less), again like in chipsets due to the same suspect (no console biss at all, discrete market shrinking abruptly with powerful iGPUs... no iGPU to show besides handheld... sounds pretty horeful to them).

 

[noob2222]

Intel would rather convince people they can save $5 a year on their electric bill than to support making a faster chip.

http://www.pcmag.com/article2/0,2817,2409738,00.asp
http://cc.bingj.com/cache.aspx?q=intel+wants+desktop+to+decline&d=4651089954541003&mkt=en-US&setlang=en-US&w=QvBBCP-VhepDEkl0MNiNIxerGWf1RKyz

and as you stated, one way Intel can get rid of Nvidia and AMD is to try and remove discrete gpus all together.

persoally I see this as a huge opportunity for AMD, If Intel leaves the DT to become just leftovers, that leaves AMD to scoop up those all to themselves.

 

[juanrga]

The reason why people don't see AMD going APU + dGPU is that there's too much power to fit into a socket. BF4 recommends AMD 3m/6c CPU. If AMD goes APU only with maximum of 2m/4c, then they won't even have a suitable gaming chip for games like BF4.

Wait a moment, because BF4 recommends a six core Piledriver CPU. As I have 'shown' in my article about Kaveri, the four core Steamroller APU is essentially at the i5 / FX-6000 level of performance.

Moreover, BF4 is not still next gen. Next gen will offload the heavy computations to the GPU, which implies that CPU will perform still better for the remaining tasks: logic, gameplay...

Whereas I agree with most of the pcper article, I find highly unlikely that AMD will release a six core APU by the end of 2014.

Because 3M = 12 ALUs/3 FPUs in AMD and i5 = 12 ALUs/4 FPUs in Intel...

Even if you make fewer ALUs faster, you're still behind. Look at the i3 with it's 6 ALUs and HTT...it doesn't run BF3 multiplayer very well at all...much less something like Crysis 3 at frame rates worth bothering over (30 FPS with GTX 690 vs. 42 FPS with GTX 690 from FX 6300).

But I was discussing final performance of the CPUs. An i3 will be behind both the i5 and the 4C kaveri in multi-threaded.

Moreover, next gen games will offload FP computations to the GPU doing irrelevant the CMT nature of the Steamroller modules.

@juanrga:

I brought up QPI because Cray could have used it, but didn't. PCIe 3.0 is only a useful setup for Intel type systems, as HT *already* provides that much throughput. However, Intel doesn't do anything that interfaces with HT and so it would require a completely new hardware solution from the ground up costing lots of money and time resources.

Additionally, if IBM are going to PCIe 3.0 it's the first I've heard of it. They were using a HT based fabric in their large server clusters...and it's capable of 51.2 GB/s of throughput bandwidth...you'd have to look long and hard at another solution to yield that much and do some heavy tweaking. The *only* reason anyone would look at PCIe 3.0 over HT would be simply to take advantage of PCIe 3.0 GPGPUs; though, I cannot make a case for that at this point in time, as most GPGPUs cannot currently entirely saturate PCIe 2.0 x16 (note: I said most, as there may be one out there that can, even if I haven't heard of it).

As I already said Cray did look for a processor agnostic interconnect. This is the reason why are not using Intel QPI.

If IBM was using a "51.2 GB/s of throughput bandwidth" HT interconnect then the switch to PCIe 3 (96 GB/s of throughput) makes sense. In fact the numbers coincide with what IBM said at hot chips. IBM claimed that the new interconnect in the Power8 offers 2x the bandwidth of the interconnects used in Power7 and Power7+

Not only IBM is going PCIe 3 with the new Power8, but Oracle, Cray, AMD... as well

Only intel !... lol

If you noticed on the link you provide 96GB/s is for 96 lanes, and since its a serial "differential" interconnect, each link has 2 lanes ( 1 outbound, 1 inbound)... so a 16x link has ~32GB/s a 48x link is the one who has ~96GB/s, and this must be the max that a PCIe controller allows, a 48x link... and yes HTT is max 51.2 GB/s but is because its limited to 32x links... Yet PCIe v3 has still considerable more overhead, so from the theoretical max peak throughput to real realizable throughput there can be a notorious difference.

IBM has there own things that are clearly superior to anything PCIe... PCIe is used because if they want expansion to other "device" boards they have no other choice, unless they make themselves all the device expansion needed (not happening)... HTT was and is used because if they wanted "co-processor" boards HTT is the only one available... though now they invented a new business to sell special PCIe v3 hardware cache coherence for their Power 8 that others must license (royalties auch!) to use, sensible since their business is really to make money lol.

In the same logic Cray doesn't have QPI because they must license it ( royalties would be real painful) from Intel... and they are also in the dire need to be in business of making money not waste it. (agnostic is for religion lol)

Cray, AMD, IBM all them claim something different.

Except Charlie is wrong.

The SeaMicro Freedom Fabric interconnect included in AMD new Seattle chip is "high-performance" and provides up to 160GB/s.

The Freedom Fabric bandwidth is far above the best than both PCIe Gen 3 and Hyperthreading can offer us. For the sake of comparison the fastest Hyperthreading version tops to about 50GB/s.

ummm... can't remember but for sure a Seamicro cluster can have up to 512 nodes, each with a CPU/APU in it, so 160GB/s is from the concentrator ASIC not for each node, but it could be for all... meaning if all nodes access network and or storage all at the same time, the bandwidth will be limited to 160/512 = 81.2MB/s for each...

Ah! "160/512 = 81.2MB/s" makes sense. Also a bandwidth interconnect of [strike]160/512 = 0.3 GB/s[/strike] "160/512 = 81.2MB/s" per CPU makes more sense, specially when noted that would be 50x slower than an old PCIe 1.0 can provide. LOL.

At the same time, the maximum 51.2 GB/s that you quote for HT is not for CPU-CPU, because that maximum has to be shared with chipsets and other resources. Usually the split is that processor, chipset, and memory link uses 16 lanes while the rest of system uses the other 16 lanes. This means that the effective bandwidth CPU-CPU is reduced to 25.6 GB/s.

 

[juanrga]

persoally I see this as a huge opportunity for AMD, If Intel leaves the DT to become just leftovers, that leaves AMD to scoop up those all to themselves.

AMD is moving away from the falling desktop market. Kaveri APU will be rated at top 100W. However its successor will be rated to a maximum of 65W (according to leaked docs). This implies the next excavator-based APUs will be still more oriented to mobile.

 

[hcl123]

Except Charlie is wrong.

The SeaMicro Freedom Fabric interconnect included in AMD new Seattle chip is "high-performance" and provides up to 160GB/s.

The Freedom Fabric bandwidth is far above the best than both PCIe Gen 3 and Hyperthreading can offer us. For the sake of comparison the fastest Hyperthreading version tops to about 50GB/s.

ummm... can't remember but for sure a Seamicro cluster can have up to 512 nodes, each with a CPU/APU in it, so 160GB/s is from the concentrator ASIC not for each node, but it could be for all... meaning if all nodes access network and or storage all at the same time, the bandwidth will be limited to 160/512 = 81.2MB/s for each...

Ah! "160/512 = 81.2MB/s" makes sense. Also a bandwidth interconnect of [strike]160/512 = 0.3 GB/s[/strike] "160/512 = 81.2MB/s" per CPU makes more sense, specially when noted that would be 50x slower than an old PCIe 1.0 can provide. LOL.

At the same time, the maximum 51.2 GB/s that you quote for HT is not for CPU-CPU, because that maximum has to be shared with chipsets and other resources. Usually the split is that processor, chipset, and memory link uses 16 lanes while the rest of system uses the other 16 lanes. This means that the effective bandwidth CPU-CPU is reduced to 25.6 GB/s.

Don't talk nonsense man!..

Now it seems you don't know what is the difference between lanes, links and max aggregated bandwidth that a controller can provide... and you make a mess of topologies...

But yes you guessed right with AMD CPUS the "links" are 16x so CPU to CPU is at max 25,6GB/s and is not needed more.

The same doesn't apply to Seamicro because there isn't direct CPU to CPU whatsoever... its a cluster... its like a collection of individual systems glued by a network, that in Seamicro has the advantage of having the storage and external LAN virtualized, and only by these you should had guessed that whatever bandwidth the FF ASIC provides, its shared to the bone.

HTT is point-to-point, PCIe is not really, FF is not by any phantom shred of imagination you could apply... and only has 160GB/s max aggregated!?... don't really remember now, but not going to digg up only to see what points you missed.

( but i think its 160GB/s for a 256 node aggregation, a full 512 node should have 2 FF ASICs chained together, so those numbers should be double (162MB/s) for theoretical, and since in average not even 20% access at exact the same time, you could multiply those numbers several times ... )

 

[juanrga]

Except Charlie is wrong.

The SeaMicro Freedom Fabric interconnect included in AMD new Seattle chip is "high-performance" and provides up to 160GB/s.

The Freedom Fabric bandwidth is far above the best than both PCIe Gen 3 and Hyperthreading can offer us. For the sake of comparison the fastest Hyperthreading version tops to about 50GB/s.

ummm... can't remember but for sure a Seamicro cluster can have up to 512 nodes, each with a CPU/APU in it, so 160GB/s is from the concentrator ASIC not for each node, but it could be for all... meaning if all nodes access network and or storage all at the same time, the bandwidth will be limited to 160/512 = 81.2MB/s for each...

Ah! "160/512 = 81.2MB/s" makes sense. Also a bandwidth interconnect of [strike]160/512 = 0.3 GB/s[/strike] "160/512 = 81.2MB/s" per CPU makes more sense, specially when noted that would be 50x slower than an old PCIe 1.0 can provide. LOL.

At the same time, the maximum 51.2 GB/s that you quote for HT is not for CPU-CPU, because that maximum has to be shared with chipsets and other resources. Usually the split is that processor, chipset, and memory link uses 16 lanes while the rest of system uses the other 16 lanes. This means that the effective bandwidth CPU-CPU is reduced to 25.6 GB/s.

Don't talk nonsense man!..

You are the one who writes "160/512 = 81.2" and the one who believes that FF will provide 50x less bandwidth than outdated PCIe 1.0. LOL

 

[hcl123]

Intel would rather convince people they can save $5 a year on their electric bill than to support making a faster chip.

http://www.pcmag.com/article2/0,2817,2409738,00.asp
http://cc.bingj.com/cache.aspx?q=intel+wants+desktop+to+decline&d=4651089954541003&mkt=en-US&setlang=en-US&w=QvBBCP-VhepDEkl0MNiNIxerGWf1RKyz

and as you stated, one way Intel can get rid of Nvidia and AMD is to try and remove discrete gpus all together.

persoally I see this as a huge opportunity for AMD, If Intel leaves the DT to become just leftovers, that leaves AMD to scoop up those all to themselves.

Intel has arguably been the volume king of GPUs since long time... but a fake king, only for fake statistics, cause even if they counted the iGPU of intel, APU like or on the chipset, no one used/uses intel iGPUs not even at entry level LOL ( one thing they will try hard to correct in future).

I think they will fail again on this, they can only defeat Nvidia, and then address AMD if they close the platform. For this their GPUs must be quite better than now... they might get close with Broadwell Iris which is double of Hasfail, but they will never ditch the desktop, that is a clear WRONG assumption, its their main cash caw... what they will try is close the platform, meaning there will not be there PCIe SLOTS for the competition boards which are the cash caw of them (special nvidia)

Broadwell BGA is a first approach, yet there could be there PCIe slots... its up to mobo an other integrators... but in any case it makes more sense for soldered discrete parts also, which perhaps should launch Intel in the business of kind of mobile discrete GPUs (depends on reference designs, but seems Intel is in the approach of making every *mainstream* system like a laptop... want upgrades !? you are out of luck, you must buy an entire new system ... just give some more money to papa.. you cheap low live! LOL)... a clever VERY STUPID scheme in order to sell more LOL

Next phase could be Intel form factors, like the old BX push... with mostly their interconnects, their peripherals etc

Then close the platform... so much for PCIe... now only on reference designs, and since no one is making chipsets for intel now that Nvidia is out of the way, PCIe only in x1, x4 or x8 links... want big graph cards !?... you got to buy the much more expensive Extreme CPUS... and more money for papa LOL...

After this, depends on how things evolve, but seems like a "requiem" to Nvidia... perhaps intel thinks it can take Nvidia place, perhaps with PhyX as real GPGPU cards big.LOL ...

The fly in the ointment, the sand in the gear is HSA... if i were Nvidia i would join HSA tomorrow... they will eventually, but not now, since they will defend their proprietary CUDA business to the last breath (hope not too late, its attacked by 2 sides, x86/AVX Phy X and HSA).

AMD could defeat this with exactly HSA... and since hUMA/HSA needs cache coherency, with an evolution of HTT tech, that could be rapidly at hand for every form factor or config, not depending on intel or anyone else... and then HSA GAMES in NSF... "compute" based, Ray-Traced, OpenCL kind of physics... extensive... LOL ... Intel says its not centered on MSFT Windows, in a way that seems more like "go to hell" (since they camped inside Linux Foundation, which is a bad augur, cause everything they touch inside dies)(edt), they may yet regret it, since MSFT could be another one joining HSA LOL

 

[hcl123]

You are the one who writes "160GB/s /512 = 81.2MB/s (fixed)" and the one who believes that FF will provide 50x less bandwidth than outdated PCIe 1.0. LOL

Yes that is right... wrong calculator copy... can't remember how that 81.2 MB/s (edt) came about (perhaps something i was doing at the same time) , can happen to anyone... but not wasting more time.

And who says FF will provide less bandwidth than PCIe 1 is you (the real nonsense)... do you have any friends ? doubt VERY much ... as i doubt you don't understand what is the difference between "shared" and "point-to-point". (it could be switched and every node see almost all of those 160GB/s (umm or is it 160Gbit/s = 20GB/s !?)... yet its "shared" nonetheless... its not 512x160)

 

[palladin9479]

People do realize you can just keep adding HT links if need be? Current 3.2 spec (which isn't even used yet) is 25.6GB/s per 32-bit lane or 12.8GB/s if you split it in half. You can just keep adding 32-bit HT lanes to get whatever bandwidth you need, it's a very expandable interconnect. Which is what HT really is, just an interconnect to link two chips together, those chips can be anything and it wouldn't matter. PCIe can be used in a similar manor but it's not quite as effective due to how it handles signaling. It's good as an expansion bus for adding devices to a controller chip, not so good for linking multiple CPU's together. It has significantly higher latency then HT by virtue of it's encoding and control mechanisms.

So for general purpose I/O (disk / video / network / ect..) PCIe works find, for synching cache or accessing remote memory, HT is vastly superior.

 

[juanrga]

You are the one who writes "160GB/s /512 = 81.2MB/s (fixed)" and the one who believes that FF will provide 50x less bandwidth than outdated PCIe 1.0. LOL

Yes that is right... wrong calculator copy... can't remember how that 81.2 MB/s (edt) came about (perhaps something i was doing at the same time) , can happen to anyone... but not wasting more time.

And who says FF will provide less bandwidth than PCIe 1 is you (the real nonsense)... do you have any friends ? doubt VERY much ... as i doubt you don't understand what is the difference between "shared" and "point-to-point". (it could be switched and every node see almost all of those 160GB/s (umm or is it 160Gbit/s = 20GB/s !?)... yet its "shared" nonetheless... its not 512x160)

Now are 20GB/s? LOL What about giving real numbers, such as 1.28 Tbit/s = 160GB/s, instead making your own numbers again?

 

[juanrga]

Those experts who still believe that ARM is a low-performance thing and that AMD moving to ARM servers is a mistake, would take a look to the new Apple A7. This is still a phone chip, but already destroy the best Intel could offer. This is what IntelUser2000 says:

Dude, even I will admit that it seems Intel needs to admit defeat to ARM. 1.3GHz dual core A7 with Smartphone class TDP is beating 4 core 1.4-2.4GHz Bay Trail with Tablet class TDP. And that's "28nm" vs 22nm.

The IPC of the A7 is on par with Ivy Bridge parts! Clock that sucker to 3GHz and it'll have a no worry replacement for the Macbook Air line and threaten even with x86 emulation.

And Intel Bayl Trail is a very good chip. It is evident that Intel worked very hard to improve its performance. Bayl Trail (Silvermont) is beyond AMD jaguar in both performance and performance per-watt.

AMD Seattle (using A57 ARM cores) will be at about the same level of performance than Intel E5 Xeons >> i7-3770k/4770k. It is understandable why AMD don't waste time releasing 8-core Steamroller server chips.

Nvidia has already promised that its Denver chip will include a custom ARM64 core that will be faster than the A57. And its Boulder chip (aimed to servers and supercomputers) will be still more powerful than its Denver chip.

I expect Apple to release a desktop/laptop ARM chip for 2015/2016

 

[de5_Roy]

since no one's gonna address this trollbait post, i'll try to reply. you're welcome.

Those experts who still believe that ARM is a low-performance thing and that AMD moving to ARM servers is a mistake, would take a look to the new Apple A7. This is still a phone chip, but already destroy the best Intel could offer. This is what IntelUser2000 says:
<blah blah>

i agree. those inteluser2000 and amduser8350 guys are so awesome that their word is teh truth when it comes to intel and amd. let's not forget nvidiatitan780 and armv8king as well. those are the guys. 😗
don't confuse amduser8350 with 8350rocks. amduser8350 just has way too much street cred. no offense.

And Intel Bayl Trail is a very good chip.

it's not. it's a decent 3rd try (or was it 4th? i forget). kabini and temash are better.

It is evident that Intel worked very hard to improve its performance. Bayl Trail (Silvermont) is beyond AMD jaguar in both performance and performance per-watt.

intel worked very hard and jaguar is still better, so is gcn. hint: no avx for silvermont.

AMD Seattle (using A57 ARM cores) will be at about the same level of performance than Intel E5 Xeons >> i7-3770k/4770k. It is understandable why AMD don't waste time releasing 8-core Steamroller server chips.

e5 and core i7 are targeted to different markets and price points. btw, do you have official or independently reviewed, reliable parformance data to corroborate your claim? no promo slides or fanboy linear mathematics, please.

Nvidia has already promised that its Denver chip will include a custom ARM64 core that will be faster than the A57. And its Boulder chip (aimed to servers and supercomputers) will be still more powerful than its Denver chip.
I expect Apple to release a desktop/laptop ARM chip for 2015/2016

riight.

 

[hcl123]

Now are 20GB/s? LOL What about giving real numbers, such as 1.28 Tbit/s = 160GB/s, instead making your own numbers again?

No ...smart arse ... just gives a link from where you picked that 160GB/s... an "official" link...

Do you know for sure ? (i don't and aint gonna digg it up) ... did you read it right ? ... somehow i suspect neither (that is what is about).

 

[sakria]

I love when people have no real answer they throw this junk ... OMG CONSPIRACY . <-- translation: I don't have an answer to that so im going to claim CONSPIRACY.

I said it when IVY came out that 14nm on DT would be hard, 10nm nearly impossible. 32 to 22nm increased heat output by 20c The only thing conspriacy theorists could claim is "omg thermal paste not solder, heat density is a myth".

Here is a clue:

If you want a certain market to slow down, what do you do to make that happen?

Answer:

Spoiler

don't give people a reason to upgrade. When was the last time Intel made a major improvement in DT cpus? OMG CONSPIRACY!!!!!!

ummm... here is another BIG CONSPIRACY http://www.fudzilla.com/home/item/32507-intel-chief-engineer-talks-haswell%E2%80%99s-5-year-journey

Is it me or Fuad is relating what a chief intel engineer told ... Broadwell transistors will be 30% smaller ???

How come !?... from 22 to 14nm according to past standards should be something in the region of 70%... 30% ?, Broadwell is not even real 16nm, is more like a good shrink on 20nm node... as if 19 or 18nm...

ummm... perhaps intel is leaving the traditional DT only for the 6 core chips, not a mad race to smaller and low power... now that they are REAL 6 core chips, perhaps Nvidia will not be dead soon, still a chance, otherwise disgusting i really regretted if Nvidia would be in big trouble (users need more competitors not less), again like in chipsets due to the same suspect (no console biss at all, discrete market shrinking abruptly with powerful iGPUs... no iGPU to show besides handheld... sounds pretty horeful to them).

Mmmm so interesting .. I agree

mariachis en bogota

 

[blackkstar]

I committed a sin and got Intel stuff, got Battlefield microstutters, *scratches head* when I made this cash strapping investment it was made on the premise of how awesome Intel is supposed to be, I personally think as always, its a load of hot air and if you are not providing me random synthetics in the real world there actually is very little to tell between a 4770Krap and a APU. Anyways I am a little perplexed and disappointed, it may just be the shame of actually supporting what I universally detest or the fact that I burned a huge hole in the wallet.


PS: saw a lot of bitching on the Memory bandwidth article, why is there no HD5000's, well how is HD4000 and old intel craphics which have shown no scaling going to change with HD5000......*straws clutch grab em now*

http://www.techspot.com/news/52070-pc-game-sales-hit-20-billion-last-year-no-signs-of-slowing.html

Traditional, non-gaming desktops that were used for browsing the web, checking emails, etc and not much more are dying. The desktops that folks play games on and do serious work on are growing or at least staying the same.

The market is not going to die completely. Desktops will never die.

As for going Intel, the one thing I've learned from building systems since the P4 days is that when you buy Intel, you're disappointed in real world performance because it gets over-hyped in benchmarks, and you're surprised by AMD because it usually doesn't look so good in a lot of reviews.

I committed a sin and got Intel stuff, got Battlefield microstutters, *scratches head* when I made this cash strapping investment it was made on the premise of how awesome Intel is supposed to be, I personally think as always, its a load of hot air and if you are not providing me random synthetics in the real world there actually is very little to tell between a 4770Krap and a APU. Anyways I am a little perplexed and disappointed, it may just be the shame of actually supporting what I universally detest or the fact that I burned a huge hole in the wallet.


PS: saw a lot of bitching on the Memory bandwidth article, why is there no HD5000's, well how is HD4000 and old intel craphics which have shown no scaling going to change with HD5000......*straws clutch grab em now*

persoally I see this as a huge opportunity for AMD, If Intel leaves the DT to become just leftovers, that leaves AMD to scoop up those all to themselves.

AMD is moving away from the falling desktop market. Kaveri APU will be rated at top 100W. However its successor will be rated to a maximum of 65W (according to leaked docs). This implies the next excavator-based APUs will be still more oriented to mobile.

AMD is not moving away from the desktop market. Intel is moving away because they are basically throwing away desktop market for mobile and they are pushing mobile chip architecture as hard as they can.

It is quite obvious Intel is doomed on desktop. Every tick and tock Intel loses overclocking headroom, which means Intel is losing potential maximum clock speed. If Intel continues on this path, it is just a matter of time until Intel has chips that can not hit the 3.5ghz base, 3.9ghz turbo established by the last few high end DT parts.

Common overclocks went like this (yes I know there are outliers, this is COMMON and basically guaranteed overclocks):

SB: high 4ghz, 5ghz
IB: 4.5ghz
HWL: 4.2ghz

Broadwell would probably have a hard time hitting 3.9ghz turbo speed required and Skylake, if this trend continues, would probably be around a 3.2ghz/3.5ghz turbo part if Intel continues focusing on mobile, power consumption, and all that stuff over actual performance (IPC AND clockspeed). The IPC gains per generation are not enough to make up for it.

Juaranga's estimations for SR performance are quite liberal as well. However if he's right it means 3m/6c part would compare to a theoretical Intel Hex without HT, 4m/8c part would wreck anything Intel has in multi-thread, and if anyone wants to believe seronx's claims of 5m/10c and even 6m/12c parts with HDL library, Intel would be completely demolished on high end DT.

The reason why I could see AMD doing this is because gaming and work DT market is still alive and well, and those are the types of folks who would buy a big rig like that. The desktops we're losing to tablets, smartphones, etc. are all for people who do casual gaming with low system specs at best. Think of your old celeron and sempron DT rigs. Those are the people going tablet/laptop/smartphone, not the people who had Core 2 Quads, high end Phenoms, etc.

Also, people love to throw around the bad 8 core numbers on Steam Hardware Survey. Just a reminder there's 54 million accounts on Steam. .28% of 54 million (which is theoretical max) is around 151,000 chips sold. That's no drop in the bucket, and I'm completely ignoring the 6 core CPU count in SHS because there's probably Intel hex mixed in there. but the number of Intel Hexs has to be a lot smaller than AMD hex phenom and FX 6000 series if you make the safe assumption cheaper things sell more.

Do not forget, AMD may not move a lot of big systems compared to APUs, but the margins on, say, an Opteron workstation with FirePro are much, much higher than the margins on a mid-range APU.

ARM for micro-servers with big, scary, steamroller cores would really put AMD in a good place. I am also assuming that AMD was trying to show us that with FX 9000 series, they don't care about power consumption or heat in a desktop system and they are not afraid to release 200w+ monster CPUs. I took that as AMD basically saying that they wouldn't be afraid to release a 6m/12c 4ghz part at 200w+ if it meant good performance.

I don't know if it'll happen, but the fact that AMD released FX 9590 and touted it as being a part of a gaming platform and all sorts of "ultimate BF3 rigs!" leads me to believe that AMD is not going to back out of this market anytime soon. It is just not the primary focus because now is a horrible time to introduce new traditional CPUs without HSA with AMD getting massive HSA design wins by getting Xbone and PS4. AMD wants gaming to stay. Intel is scared of gaming and trying to kill off dGPU. Nvidia still trying to make big, impressive dGPUs.

All signs point to the gaming DT not going anywhere for a long, long time yet there's still people who insist that it's going to disappear. If it was going to disappear naturally, Intel wouldn't care about trying to off dGPUs because the market would die out on its own. They did the same with competing against AMD and shifting their focus to ARM, not sure why Intel wouldn't just abandon a fight it didn't see worth fighting again.

EDIT: I do think AMD is staying hush because they have something massive coming. I did the math with Seronx's claims of 32nm HDL 6m/12c parts to see how big the die was and it ended up around 330mm^2, while Vishera 4m/8c is 315mm^2. If AMD did this, they would have a high end part that was smaller than 4930k and 3930k, mauled it in performance, and all of that could trickle down.

AMD could open up Intel slaughter by releasing 6m/12c chip at 330mm^2 for $499, which would wreck Intel hex in multi-thread, and then trickle that down. 5m part at 4770k price and 4m part at 4670k price, APUs fill in the rest. It would be Athlon days all over again, except AMD would have control of the software in gaming benchmarks as they have access to the source of most of the tools being used and they have the hardware for the two big consoles.

There is huge potential for AMD to come back stronger than they ever were in desktop and I simply don't see them throwing that away. They have the possibility of making a superior chip to Intel hex EE and even Haswell-E 8 core and I don't see them throwing that away, not after AMD shows us they care about gaming market by releasing FX 9590 as an extreme gaming chip.

 

[juanrga]

And Intel Bayl Trail is a very good chip.

it's not. it's a decent 3rd try (or was it 4th? i forget). kabini and temash are better.

It is evident that Intel worked very hard to improve its performance. Bayl Trail (Silvermont) is beyond AMD jaguar in both performance and performance per-watt.

intel worked very hard and jaguar is still better, so is gcn. hint: no avx for silvermont.

AMD Seattle (using A57 ARM cores) will be at about the same level of performance than Intel E5 Xeons >> i7-3770k/4770k. It is understandable why AMD don't waste time releasing 8-core Steamroller server chips.

e5 and core i7 are targeted to different markets and price points. btw, do you have official or independently reviewed, reliable parformance data to corroborate your claim? no promo slides or fanboy linear mathematics, please.

Nvidia has already promised that its Denver chip will include a custom ARM64 core that will be faster than the A57. And its Boulder chip (aimed to servers and supercomputers) will be still more powerful than its Denver chip.
I expect Apple to release a desktop/laptop ARM chip for 2015/2016

riight.

Bayl Trail is a so good chip (I refer to CPU alone) as Kabini. Silvermont offers about the same performance than jaguar but with less power consumption. AMD will be releasing enhanced jaguar the next year, which will provide AMD advantage again. Moreover, Kabini Temash substitutes will be fully HSA enabled, giving AMD extra advantage. But that does not dismiss that Intel has done a very good work with Silvermont (I must confess that I expected a poor chip).

Sure e5 and i7 are targetted to different price points and workloads, but it is irrelevant for comparing raw performance (MT).

Sorry, but it is a bit difficult that Seattle had been "independently reviewed" because is not released still, but its performance will be that +- margin of error.

Now are 20GB/s? LOL What about giving real numbers, such as 1.28 Tbit/s = 160GB/s, instead making your own numbers again?

No ...smart arse ... just gives a link from where you picked that 160GB/s... an "official" link...

Do you know for sure ? (i don't and aint gonna digg it up) ... did you read it right ? ... somehow i suspect neither (that is what is about).

Yes I know for sure that you are making up numbers again. The correct bandwidth is 1.28 Tbit/s = 160GB/s.

 

[hcl123]

where is the link ? ... i'm making up nothing, and still not going to digg up, i'm waiting for you (your word worths po,,, your authoritative mode less)

 

[juanrga]

AMD is not moving away from the desktop market. Intel is moving away because they are basically throwing away desktop market for mobile and they are pushing mobile chip architecture as hard as they can.

It is quite obvious Intel is doomed on desktop. Every tick and tock Intel loses overclocking headroom, which means Intel is losing potential maximum clock speed. If Intel continues on this path, it is just a matter of time until Intel has chips that can not hit the 3.5ghz base, 3.9ghz turbo established by the last few high end DT parts.

Common overclocks went like this (yes I know there are outliers, this is COMMON and basically guaranteed overclocks):

SB: high 4ghz, 5ghz
IB: 4.5ghz
HWL: 4.2ghz

Broadwell would probably have a hard time hitting 3.9ghz turbo speed required and Skylake, if this trend continues, would probably be around a 3.2ghz/3.5ghz turbo part if Intel continues focusing on mobile, power consumption, and all that stuff over actual performance (IPC AND clockspeed). The IPC gains per generation are not enough to make up for it.

Juaranga's estimations for SR performance are quite liberal as well. However if he's right it means 3m/6c part would compare to a theoretical Intel Hex without HT, 4m/8c part would wreck anything Intel has in multi-thread, and if anyone wants to believe seronx's claims of 5m/10c and even 6m/12c parts with HDL library, Intel would be completely demolished on high end DT.

The reason why I could see AMD doing this is because gaming and work DT market is still alive and well, and those are the types of folks who would buy a big rig like that. The desktops we're losing to tablets, smartphones, etc. are all for people who do casual gaming with low system specs at best. Think of your old celeron and sempron DT rigs. Those are the people going tablet/laptop/smartphone, not the people who had Core 2 Quads, high end Phenoms, etc.

Also, people love to throw around the bad 8 core numbers on Steam Hardware Survey. Just a reminder there's 54 million accounts on Steam. .28% of 54 million (which is theoretical max) is around 151,000 chips sold. That's no drop in the bucket, and I'm completely ignoring the 6 core CPU count in SHS because there's probably Intel hex mixed in there. but the number of Intel Hexs has to be a lot smaller than AMD hex phenom and FX 6000 series if you make the safe assumption cheaper things sell more.

Do not forget, AMD may not move a lot of big systems compared to APUs, but the margins on, say, an Opteron workstation with FirePro are much, much higher than the margins on a mid-range APU.

ARM for micro-servers with big, scary, steamroller cores would really put AMD in a good place. I am also assuming that AMD was trying to show us that with FX 9000 series, they don't care about power consumption or heat in a desktop system and they are not afraid to release 200w+ monster CPUs. I took that as AMD basically saying that they wouldn't be afraid to release a 6m/12c 4ghz part at 200w+ if it meant good performance.

I don't know if it'll happen, but the fact that AMD released FX 9590 and touted it as being a part of a gaming platform and all sorts of "ultimate BF3 rigs!" leads me to believe that AMD is not going to back out of this market anytime soon. It is just not the primary focus because now is a horrible time to introduce new traditional CPUs without HSA with AMD getting massive HSA design wins by getting Xbone and PS4. AMD wants gaming to stay. Intel is scared of gaming and trying to kill off dGPU. Nvidia still trying to make big, impressive dGPUs.

All signs point to the gaming DT not going anywhere for a long, long time yet there's still people who insist that it's going to disappear. If it was going to disappear naturally, Intel wouldn't care about trying to off dGPUs because the market would die out on its own. They did the same with competing against AMD and shifting their focus to ARM, not sure why Intel wouldn't just abandon a fight it didn't see worth fighting again.

EDIT: I do think AMD is staying hush because they have something massive coming. I did the math with Seronx's claims of 32nm HDL 6m/12c parts to see how big the die was and it ended up around 330mm^2, while Vishera 4m/8c is 315mm^2. If AMD did this, they would have a high end part that was smaller than 4930k and 3930k, mauled it in performance, and all of that could trickle down.

AMD could open up Intel slaughter by releasing 6m/12c chip at 330mm^2 for $499, which would wreck Intel hex in multi-thread, and then trickle that down. 5m part at 4770k price and 4m part at 4670k price, APUs fill in the rest. It would be Athlon days all over again, except AMD would have control of the software in gaming benchmarks as they have access to the source of most of the tools being used and they have the hardware for the two big consoles.

There is huge potential for AMD to come back stronger than they ever were in desktop and I simply don't see them throwing that away. They have the possibility of making a superior chip to Intel hex EE and even Haswell-E 8 core and I don't see them throwing that away, not after AMD shows us they care about gaming market by releasing FX 9590 as an extreme gaming chip.

Both AMD and Intel are moving away from the desktop market, because this is dying. Three years ago I could see my favourite computer stores full of big (tower) desktops. Today all them have been replaced by laptops, and the place previously used by laptops is now full of tablets. If I find one desktop computer for sell, it is some old model under rebate. Of course people still purchase desktops, but sales dropped at least by one order of magnitude.

Intel is moving away to the ultra-mobile market (phone, tablet...). AMD is moving to the custom and embedded market. AMD said that they except at least 50% of future revenue to come from the new markets.

My estimations of SR performance are based in simulations that use all the information that AMD has disclosed about Steamroller, plus some few guesses (such as exact freq. of the CPU). I estimate the error of the simulation must be less than 5%.

I didn't simulate a Steamroller 4m/8c because none is coming, but I can say you that it would be faster (even ~20% faster) than an i7-4770k in tests like

but would perform similar to the i7-4770k in tests like

which exploits Haswell improved FPUs (th. 2x the Ivy Bridge FPUs performance).

Seronx's claims of 5m/10c and even 6m/12c parts are invalid. There is no market for such chips, specially at gaming. The percentage of gamers using a 4m/8c is about 0.3% and 12/16 cores accounts to 0%. Moreover, the tendency in gaming is towards offloading the CPU, which implies that 'weak' CPUs will be enough for next gen games. The heavy workloads will be running in the GPU. That is not anything new but the tendency since decades ago.

Steamroller is not competitive for ARM. A hypothetical 16-core Steamroller Opteron would offer slightly better performance than Seattle, but would consume much more power, doing it highly inefficient for servers and, thus, unable to be sold. Moreover, the x86 server market is also dying and it makes no sense that AMD release an uncompetitive product for a dying market. Their move to ARM is the correct move and Seattle is an excellent chip. Warsaw, using Piledriver, is aimed to small percentage of legacy users; so small that a Steamroller version was not considered by economic reasons.