AMD CPU speculation... and expert conjecture

[juanrga]

"At the time of the initial Mantle announcement, we were already investigating rendering overhead based on game developer feedback," an Intel spokesman said in an email. "Our hope was to build consensus on potential approaches to reduce overhead with additional data. We have publicly asked them to share the spec with us several times as part of examination of potential ways to improve APIs and increase efficiencies. At this point though we believe that DirectX 12 and ongoing work with other industry bodies and OS vendors will address the issues that game developers have noted."

Emphasis mine. Reading this literally, AMD hasn't given Intel the spec yet, given how Intel has asked for it "several times".

It is a bit difficult that AMD can give an inexistent spec... MANTLE is still in beta.

They basically told Intel to wait for MANTLE goes out of Closed beta, probably in 3 months or so.

Exactly. When the spec was ready everyone including Intel will access it.

As mentioned above AMD is going to release a new FX 8-core CPU for AM3+. One based in Piledriver.

They are? Do you have a source on that? Just curious!

The source for AMD releasing FX-something and my guess about what will be

https://twitter.com/juanrga/status/480315084255002624

After I wrote that, some sites have confirmed 5GHz+ PD

http://www.xbitlabs.com/news/cpu/display/20140621200650_AMD_Preps_a_New_FX_Series_Microprocessor_with_Bundled_Liquid_Cooling.html

 

[juanrga]

I recall some time ago some people here speculated about AMD releasing in future a GPU on socket as better alternative to APUs. Their belief was that APU is only a temporary product and that future products would be dual sockets mobos with CPU on one socket and GPU in other socket both connected by a fast link (a la HTT).

I already showed before why dcPU+dGPU cannot compete against a high-performance APU, but I found this old article where the own AMD admits that APUs are better

http://www.xbitlabs.com/news/other/display/20111211180811_AMD_GPGPU_Accelerators_in_CPU_Sockets_Make_No_Sense.html

You think differently than AMD...but what do they know...they have freedom fabric, and all the Terabit/sec goodness that comes with that...

In case you didn't notice the above link quotes what AMD thinks.

Freedom Fabric doesn't change a bit the problems with socketed co-processors.

 

[jdwii]

Devils canyon is more than marketing gimmick. It is pure lying from Intel.

Intel said that the chips crossed 5.5GHz on air during the open competition. This is false. I asked Intel in public for the proof/submission of the OC, and they gave me none, because air cooling was not used. All 'professional' tech sites did spread the over-5.5GHz-on-air lie, except one site, which (mentioning me) stated that all records used either water or LN2.

Since then we have seen the professional reviews of devils canyon and their inability to get 4.6GHz stable on air or 4.8GHz with water... but none of those sites mentioning the over-5.5GHz-on-air lie. And most forums silent.

But if AMD had made only a half of that, you would see the strong reaction in the press and in the forums. I still recall the ridiculous noise in some forums when AMD released the Centurion chip and advertised it as 5GHz chip, when that was the turbo frequency (4.7GHz base). I recall fanboys promoting a ban of the new AMD products because according to them AMD had violated certain European laws only in their imagination. Pathetic.

This is why we want Higher-end chips from Amd to give Intel more competition

 

[jdwii]

As mentioned above AMD is going to release a new FX 8-core CPU for AM3+. One based in Piledriver.

SMDH at these pointless 220w chips. For the cost of motherboard, CPU, and cooling, I could buy a 4970k, a Z97 extreme3 or 4, and less expensive cooler that would wipe the floor with it, for less. Water cooling is almost a must with these chips. If doing air, I wouldn't want less than a Noctua NH-D14, where as a Thermalright True Spirit 140 power would probably be sufficient. AMD is doing the same idiocy that Intel did with the P4. Even better, I could go with a Xeon E3, with an even less expensive cooler and board, giving me more for GPU.

PCPartPicker part list / Price breakdown by merchant

CPU: Intel Core i7-4790K 4.0GHz Quad-Core Processor ($339.99 @ Newegg)
CPU Cooler: Thermalright TRUE Spirit 140 POWER 73.6 CFM CPU Cooler ($49.98 @ SuperBiiz)
Motherboard: ASRock Z97 EXTREME4 ATX LGA1150 Motherboard ($143.99 @ Newegg)
Total: $533.96
Prices include shipping, taxes, and discounts when available




PCPartPicker part list / Price breakdown by merchant

CPU: AMD FX-9590 4.7GHz 8-Core Processor ($318.99 @ SuperBiiz)
CPU Cooler: Noctua NH-D14 65.0 CFM CPU Cooler ($64.99 @ NCIX US)
Motherboard: ASRock 990FX Extreme9 ATX AM3+ Motherboard ($169.99 @ Amazon)
Total: $553.97
Prices include shipping, taxes, and discounts when available


PCPartPicker part list / Price breakdown by merchant

CPU Cooler: Cooler Master Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler ($30.99 @ Amazon)
Motherboard: Gigabyte GA-Z97X-SLI ATX LGA1150 Motherboard ($118.99 @ Amazon)
Other: Xeon 1241v3 ($274.71)
Total: $424.69
Prices include shipping, taxes, and discounts when available

FX-9370 is $229 @ 4 different locations...the 9590 is still high, but then again...it is super high bin silicon too...

Yet it would still get beat by a 1231v3 with a cheaper board. I want AMD to compete, but this emulating P4 isn't the way to do it. After FX 8320, AMD's value drops significantly.

PCPartPicker part list / Price breakdown by merchant

CPU: AMD FX-9370 4.4GHz 8-Core Processor ($229.89 @ SuperBiiz)
CPU Cooler: Noctua NH-D14 65.0 CFM CPU Cooler ($64.99 @ NCIX US)
Motherboard: ASRock 990FX Extreme9 ATX AM3+ Motherboard ($169.99 @ Amazon)
Total: $464.87
Prices include shipping, taxes, and discounts when available


PCPartPicker part list / Price breakdown by merchant

CPU Cooler: Cooler Master Hyper 212 EVO 82.9 CFM Sleeve Bearing CPU Cooler ($30.99 @ B&H)
Motherboard: Gigabyte GA-Z97X-SLI ATX LGA1150 Motherboard ($118.99 @ Amazon)
Other: Xeon 1231v3 ($247.98)
Total: $397.96
Prices include shipping, taxes, and discounts when available

I'm pretty sure a lot of people would only buy that chip if they are either an Amd fanboy that can afford a little more or a person on a Amd rig who can already get one if they are stuck with anything less then a 8320. Even then i have to say its not for me i would rather OC my self.

 

[etayorius]

It`s basically another FX-9590, only difference is Water Cooler, not even less TDP or more Speed.

 

[esrever]

And it was nothing. $350 for a piledriver chip is terrible.

 

[de5_Roy]

i was hoping for an fx4xxx branded igpu-less kaveri for socket fm2+. oh well. there's always next time.

 

[gamerk316]

"At the time of the initial Mantle announcement, we were already investigating rendering overhead based on game developer feedback," an Intel spokesman said in an email. "Our hope was to build consensus on potential approaches to reduce overhead with additional data. We have publicly asked them to share the spec with us several times as part of examination of potential ways to improve APIs and increase efficiencies. At this point though we believe that DirectX 12 and ongoing work with other industry bodies and OS vendors will address the issues that game developers have noted."

Emphasis mine. Reading this literally, AMD hasn't given Intel the spec yet, given how Intel has asked for it "several times".

It is a bit difficult that AMD can give an inexistent spec... MANTLE is still in beta.

Regardless of the implementation, the specifications should be LONG stabilized, otherwise it would be impossible to develop for. Specification != implementation.

 

[cemerian]

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

 

[gamerk316]

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

PhysX IS open. Proprietary to NVIDIA, but open. There's no reason AMD can't implement PhysX via OpenCL if they wished to.

Mantel is pretty much the same way. NVIDIA is about as likely to adopt Mantle as AMD is to adopt PhysX. Open standards without industry support die. Its that simple.

 

[de5_Roy]

AMD exec doubles down on assertion that Nvidia Gameworks presents a threat to PC gamers
http://www.maximumpc.com/no_bs_podcast_226_-depth_interview_amd_graphics_guru_richard_huddy
everybody loves p.r. fu fighting

AMD A6-7400K up for pre-order in US
http://www.fudzilla.com/home/item/35052-amd-a6-7400k-up-for-pre-order-in-us
just check newegg(only), no sign of the new apus. zambezi cpus are gone. fx8320 is back to $160. i was hoping it'd come down to $140. fx9370 has limited duration $30 off with a promo code while 9590 sans cooler is $330. both kaveri apus are at $170 and $160. why there's a $10 difference boggles my ignorant mind.

apparently the cm seidon 120 aio lcs will be bundled with the new "fx refresh"
http://www.hardwarecanucks.com/forum/hardware-canucks-reviews/66761-exclusive-amd-launch-new-fx-9590-cpu.html

bjorn3d's a10 7850k review
http://www.bjorn3d.com/2014/06/amd-a10-7850k-kaveri/

 

[iron8orn]

http://www.tomshardware.com/forum/id-2193486/heterogeneous-architecture-parallel-proccessing-extented-instrution.html#13574037

come speculate on the latest architecture of AMD and the HSA foundation.

 

[etayorius]

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

PhysX IS open. Proprietary to NVIDIA, but open. There's no reason AMD can't implement PhysX via OpenCL if they wished to.

Mantel is pretty much the same way. NVIDIA is about as likely to adopt Mantle as AMD is to adopt PhysX. Open standards without industry support die. Its that simple.

I don`t think it`s open, it may be free to use but i have not ever heard or read anywhere about PhysX being open.

The only Open Physics engine is Bullet.

 

[8350rocks]

Except if Intel adopts Mantle, then ~82% of the GPU market would offer it and NVidia would be forced to do so or fight uphill against it eternally.

 

[de5_Roy]

Except if Intel adopts Mantle, then ~82% of the GPU market would offer it and NVidia would be forced to do so or fight uphill against it eternally.

in terms of marketshare (and marketing hype, promo slides etc.), definitely. but in terms of "gaming relevance" (windows), not so much, lol. however, intel supporting mantle-for-linux in linux distros will be much more impactful. amd simply has to be not lazy and port mantle to linux so that intel can take advantage of it. i know how it sounds, but that's how it'll turn out to be.

 

[cemerian]

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

PhysX IS open. Proprietary to NVIDIA, but open. There's no reason AMD can't implement PhysX via OpenCL if they wished to.

Mantel is pretty much the same way. NVIDIA is about as likely to adopt Mantle as AMD is to adopt PhysX. Open standards without industry support die. Its that simple.

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

PhysX IS open. Proprietary to NVIDIA, but open. There's no reason AMD can't implement PhysX via OpenCL if they wished to.

Mantel is pretty much the same way. NVIDIA is about as likely to adopt Mantle as AMD is to adopt PhysX. Open standards without industry support die. Its that simple.

I don`t think it`s open, it may be free to use but i have not ever heard or read anywhere about PhysX being open.

The only Open Physics engine is Bullet.

Actually physx is open, but it uses opencl and runs on cpu's if you don't have nvidia gpu, the promise at the beginning was to make it available for amd to run on theire gpu's aswell, but that didn't pan out

 

[gamerk316]

ROFL Mantle is basicly amd version of ''physx'' in the sense that when physx first launched it was intended to make it open in a year or two look how that ended, I am pretty sure we wont see mantle on any other hardware(theire own gaming scientist said its design precisely for GCN and will run better on GCN, so how far can they go with this open bs its a proprietary standard untill its open)

PhysX IS open. Proprietary to NVIDIA, but open. There's no reason AMD can't implement PhysX via OpenCL if they wished to.

Mantel is pretty much the same way. NVIDIA is about as likely to adopt Mantle as AMD is to adopt PhysX. Open standards without industry support die. Its that simple.

I don`t think it`s open, it may be free to use but i have not ever heard or read anywhere about PhysX being open.

The only Open Physics engine is Bullet.

PhysX has been an open standard since NVIDIA brought Ageia back in 2008. Not to be confused with Open Source.

http://www.sutor.com/c/essays/osvsos/

So I'll say it again and again: AMD is free to implement PhysX through OpenCL whenever they want to.

 

[iron8orn]

i was hoping for an fx4xxx branded igpu-less kaveri for socket fm2+. oh well. there's always next time.

I second that, but a 8,6 core piledriver on fm2+ would have been acceptable. Maybe there is still hope.

i think it is likely to happen within the next 6 months

 

[gamerk316]

Actually physx is open, but it uses opencl and runs on cpu's if you don't have nvidia gpu, the promise at the beginning was to make it available for amd to run on theire gpu's aswell, but that didn't pan out

Not quite; when NVIDIA brought Ageia, they decided to implement the PhysX API using CUDA, so it could run on their GPU's. This is how PhysX runs to this day. For AMD to implement the PhysX specification, they'd have to re-write the implementation to use OpenCL instead.

You know, kind of like converting your graphical backend from DirectX to Mantle. You're doing the processing, but what you have to do to get there is different due to using a different API to get there.

NVIDIA made the decision to not support split AMD/NVIDIA configs for cost reasons; NVIDIA would be on the hook for potentially debugging problems on AMD's end of the chain.

 

[de5_Roy]

I second that, but a 8,6 core piledriver on fm2+ would have been acceptable. Maybe there is still hope.

i think it is likely to happen within the next 6 months

imo, a 32nm, 4M/8C cpu for fm2+ shouldn't be difficult to build, but amd might not do it for resource and economic feasibility issues, as they allocate resources to excavator apus/socs and arm cpus/socs. they'd have to rework the vishera design for the added integrated components in the uncore, likely change the pin-outs for socket fm2+ (i can't confirm this). since majority of 4M and high clocked 3M vishera cpus have 125w tdp, mobo validation could be problematic. i don't know how much power savings glofo's 28nm bulk node's gonna offer. there's the new 22nm SOI fab from IBM but i don't think amd can quickly port their cpu design to that process. in the end, a lot of re-design, testing, validating, economic, timing issues will get in the way of this.

edit2: a sub $100, moderately high clocked, igpu-disabled, unlocked, kaveri "cpu" would be a good counter to intel's pentium a.e. cpu. i didn't think amd would (attempt to) counter the high end d.c. cpus.

 

[iron8orn]

I second that, but a 8,6 core piledriver on fm2+ would have been acceptable. Maybe there is still hope.

i think it is likely to happen within the next 6 months

imo, a 32nm, 4M/8C cpu for fm2+ shouldn't be difficult to build, but amd might not do it for resource and economic feasibility issues, as they allocate resources to excavator apus/socs and arm cpus/socs. they'd have to rework the vishera design for the added integrated components in the uncore, likely change the pin-outs for socket fm2+ (i can't confirm this). since majority of 4M and high clocked 3M vishera cpus have 125w tdp, mobo validation could be problematic. i don't know how much power savings glofo's 28nm bulk node's gonna offer. there's the new 22nm SOI fab from IBM but i don't think amd can quickly port their cpu design to that process. in the end, a lot of re-design, testing, validating, economic, timing issues will get in the way of this.

edit2: a sub $100, moderately high clocked, igpu-disabled, unlocked, kaveri "cpu" would be a good counter to intel's pentium a.e. cpu. i didn't think amd would (attempt to) counter the high end d.c. cpus.

read tourist's comment wrong.. i think there will be a 8 thread steamroller Kaveri soon a 22nm process could be great but i think amd is working on something a little better. i speculate a 16nm cpu with 16nm gpu's

 

[truegenius]

I recall some time ago some people here speculated about AMD releasing in future a GPU on socket as better alternative to APUs. Their belief was that APU is only a temporary product and that future products would be dual sockets mobos with CPU on one socket and GPU in other socket both connected by a fast link (a la HTT).

I already showed before why dcPU+dGPU cannot compete against a high-performance APU, but I found this old article where the own AMD admits that APUs are better

http://www.xbitlabs.com/news/other/display/20111211180811_AMD_GPGPU_Accelerators_in_CPU_Sockets_Make_No_Sense.html

this is what they said

Advanced Micro Devices said that making special-purpose GPU-based accelerators compatible with CPU sockets makes no sense.

Anna Filatova, AMD's leading software expert Neal Robison said that Fusion-architecture - which integrates general-purpose [x86] processing cores with highly-parallel stream processors of Radeon GPUs - is a better solution for high-performance computing than to install special-purpose accelerators into CPU sockets.

that means they are not saying that igpus will be better in gaming than dgpus
in accelerating tasks using gpus, igpus can easily beat dgpus ( for example quicksync vs amd/nvidia hw acceleration for video conversion ) because interconnect is holding dgpus here and not the gpus them self ( you can try an experiment t home, clock whichever cpu have to minimum speed and turn off all cres except 1, and then run video conversion task using hw acceleration ( you can use atixcoder for amd's dgpus ) and see how much percent dgpu is utilized and does it even runs at its full speed, this way you will find the amount of dgpu utilized which will show you that its interconnect holding back )
but gaming is different, pcie is not bottleneck here, gpus are bottleneck, run any game at highest settings and dgpu reaches to 100% at full speed, thus gpu is bottleneck

so treat these statement while keeping in mind that gaming and general purpose are different things, rest you can understand by yourself

i always said and will say it again that igpus can't beat dgpus in gaming

 

[logainofhades]

Sure they can, if the dgpu is quite old.

 

[juanrga]

This is why we want Higher-end chips from Amd to give Intel more competition

Companies lie with independence of competition. In any case, have you been living under a rock (PUN intended) the last months? AMD has already announced its plans for higher-end: K12 and companion x86 sister.

i was hoping for an fx4xxx branded igpu-less kaveri for socket fm2+. oh well. there's always next time.

The fact there is none in the market, proves that igpu yields are very very good, which is not a surprise because if there was yields problems they would be detected first in the CPU. :-D

Regardless of the implementation, the specifications should be LONG stabilized, otherwise it would be impossible to develop for. Specification != implementation.

It is not the implementation which is in beta. It is the own API which is being still developed in collaboration with game developers.

PhysX IS open. Proprietary to NVIDIA, but open.

Please check next link Propietary vs Open Standards

Except if Intel adopts Mantle, then ~82% of the GPU market would offer it and NVidia would be forced to do so or fight uphill against it eternally.

The problem is this is a big IF.

i think there will be a 8 thread steamroller Kaveri soon a 22nm process could be great but i think amd is working on something a little better. i speculate a 16nm cpu with 16nm gpu's

There is no 8 thread steamroller Kaveri in the menu. The IBM 22nm SOI process couldn't provide the characteristics needed for GCN. Moreover, people here is assuming, as proven, that IBM will license its process to anyone including AMD, when it is likely that GF will use IBM technology only for IBM products.

In any case, the route that AMD will follow is rather clear: 28m bulk planar --> 20nm bulk planar --> 14nm bulk FF, except if there are unexpected problems/delays, of course.

that means they are not saying that igpus will be better in gaming than dgpus
in accelerating tasks using gpus, igpus can easily beat dgpus ( for example quicksync vs amd/nvidia hw acceleration for video conversion ) because interconnect is holding dgpus here and not the gpus them self ( you can try an experiment t home, clock whichever cpu have to minimum speed and turn off all cres except 1, and then run video conversion task using hw acceleration ( you can use atixcoder for amd's dgpus ) and see how much percent dgpu is utilized and does it even runs at its full speed, this way you will find the amount of dgpu utilized which will show you that its interconnect holding back )
but gaming is different, pcie is not bottleneck here, gpus are bottleneck, run any game at highest settings and dgpu reaches to 100% at full speed, thus gpu is bottleneck

so treat these statement while keeping in mind that gaming and general purpose are different things, rest you can understand by yourself

i always said and will say it again that igpus can't beat dgpus in gaming

They are saying that Torrenza initiative is killed. Sometime ago someone here pretended that it was alive...

In any case regarding gaming, I already explained before how gaming is evolving to blur the distinction between compute GPU and rendering GPU, how killing GPGPUs will do gaming GPUs economically inviable, how nonlinear silicon scaling will affect future gaming dGPUs...

Feel free to ignore the arguments once again.

 

[gamerk316]

So....NASA apparently has a 460GHz transistor prototype. Way to go for government funded projects! And yay Vacuum tubes!

http://www.extremetech.com/extreme/185027-the-vacuum-tube-strikes-back-nasas-tiny-460ghz-vacuum-transistor-that-could-one-day-replace-silicon-fets