AMD CPU speculation... and expert conjecture

[juanrga]

Adding two more memory channels definitely needs a new socket as you damn near double your interconnects. It's 200+ pins per memory channel (64-bit width) and then some.

Are not 240 pins?

 

[BeastLeeX]

@BeastLeeX, Thank you for your input good sir! AMD must be up to something big for them not to break their silence yet. The FX 6300 is a fantastic buy. I built my friend a rig and he uses that CPU. And man... it's a baller!

And look, another Rumor down the crapper!!!

http://www.maximumpc.com/driver_update_will_add_directx_112_support_amd_radeon_hd_7000_series_gpus

Sounds good to me, although im not really worried about 11.2. Maybe a little more stress on my 7970 ;p. Also correct if I'm wrong but this is Windows 8 exclusive, right?

 

[Ags1]

Also, I am curious. It sounds more and more like HSA is going to end up something where you just recompile instead of reprogram? Is this true? I thought that was a massive, massive problem to overcome. I'd find it humorous if I could buy a cheap APU laptop, slap gentoo on it, and then have it outperform my entire CPU only FX 8350 rendering rig.

Shouldn't even need a recompile. Due to the non-X86 nature of the HSA cores (or whatever you want to call them), they likely will be invisible to the OS, meaning its up to the drivers/hardware to handle thread assignment.

Granted, DirectX, OpenGL,OpenCL, DirectCompute, CUDA and the like should all automatically go to the GPU portion of the chip, but I would imagine the implementation by AMD will be smart enough to load threads to HSA cores if the resources are available. [If not, HSA is DOA]

Most old software, using only CPU, will have to be rewritten to use the GPU for compute. This is why AMD joined with LibreOffice devs. to enable HSA. Maybe in some special cases a smart HSA compiler can do some magic for old code.

However, some of the software already using the GPU for compute will work automatically with HSA software without any change:

Good news for all of us Java devs!

 

[Cazalan]

Adding two more memory channels definitely needs a new socket as you damn near double your interconnects. It's 200+ pins per memory channel (64-bit width) and then some.

Are not 240 pins?

Closer to 150 signals per DIMM. The rest are just powers and ground.

 

[8350rocks]

@BeastLeeX, Thank you for your input good sir! AMD must be up to something big for them not to break their silence yet. The FX 6300 is a fantastic buy. I built my friend a rig and he uses that CPU. And man... it's a baller!

And look, another Rumor down the crapper!!!

http://www.maximumpc.com/driver_update_will_add_directx_112_support_amd_radeon_hd_7000_series_gpus

Sounds good to me, although im not really worried about 11.2. Maybe a little more stress on my 7970 ;p. Also correct if I'm wrong but this is Windows 8 exclusive, right?

No, it should be in the Catalyst drivers when they release the newest version...

 

[palladin9479]

Adding two more memory channels definitely needs a new socket as you damn near double your interconnects. It's 200+ pins per memory channel (64-bit width) and then some.

Are not 240 pins?

That's for the DIMM slot, I'm talking about CPU pins for the memory bus. The Power lines and most of the ground lines come from the MB but each channel is 128 pins just for the interface then the control lines and any additional signal / ground lines. It all leads to an expensive motherboard as they have to map out all the traces.

 

[hcl123]

http://wccftech.com/amd-roadmap-20142015-products-updated-kaveri-apus-arrive-q1-2014-hawaii-gpu-late-september/

Looks like FX series is currently accounting for 30% of AMD's new CPU shipments...

With numbers like that I am pretty certain they wouldn't cancel the line. APUs aren't the 80-90% everyone thought they were. Though, they are clearly doing well.

Funny the article "wccftech" points "digitimes" talks about "entry-level", of what might constitute Kabini with new packages... it doesn't talk "Richland" as in the "performance" sector, it talks "high-end" for Kaveri (whatever... could be top of line)... but...

"AMD has declined to comment on unannounced products."( so there must be some) LOL

Seems completely fabricated (and old)... all announced since long... there isn't such official chart... or does it ?
http://cdn.wccftech.com/wp-content/uploads/2013/08/AMD-Roadmap-2014-2015.png

If they can do that on less relevant kind of pointless simple roadmap chart, what they cannot do with "benchmark charts" LOL .. (yes edited by hand for lack of optimization time lol... do you hear hafijur ?.. for the battle of propaganda and rumors anything goes, thats the intrinsic nature of the beast... )

[UPDATE: and funny wccftech itself talks as if "kabini" will be new in 2014, when they posted an article stating the "official launch" of temash and kabini: "AMD recently introduced their Jaguar based Temash and Kabini products feature improved power levels and performance enhancements." ... more than 1 month ago...

http://wccftech.com/amd-mullins-beema-apus-based-puma-core-architecture/

and now !!!??? ... can we say very confused journalists ... or worst ? ]

 

[hcl123]

Adding two more memory channels definitely needs a new socket as you damn near double your interconnects. It's 200+ pins per memory channel (64-bit width) and then some.

Are not 240 pins?

That's for the DIMM slot, I'm talking about CPU pins for the memory bus. The Power lines and most of the ground lines come from the MB but each channel is 128 pins just for the interface then the control lines and any additional signal / ground lines. It all leads to an expensive motherboard as they have to map out all the traces.

According to my calculations only those cascaded bulky transistorizes for a 128bit DDR3 interface occupy ~25mm² at 32nm... and this without counting the specific controller part, and all the wire routing... and those cascaded things, necessary for a steep voltage-domain crossing, are the worst pigs of them all to scale. So at 28nm we might be talking with luck of 20mm², 4 channels 40mm²... ballooned by controller and much more wire routing...

I think 32MB of ESRAM at 28nm could be notoriously smaller than that(just scales pretty well, perhaps the best scaler of all structures) (edt)... and provide almost the double of bandwidth, triple better latency, and cherry on top of cake consume quite less power (its on tiny Wuii).

 

[hcl123]

Yes it is...

In other news...Steve Ballmer has announced his impending retirement within the next year:

http://www.tomshardware.com/news/retirement-bill-gates-steve-ballmer-surface-rt-valueact,24009.html

Will it be too little, too late for M$? I don't know if they can recover from the fiasco that is Winblows 8/RT/Blue/Surface/Phone/XBone.

If anyone @ canonical is watching...launch a massive ad campaign now and get Ubuntu Linux out into the world while you have the best opportunity. If M$ gets their act together in time to right the ship...Linux may be destined for a small % user base indefinitely.

Has already picked some steam, Ubuntu is installed from origin in what with some vendors can be ~25% of all new laptops... and is installed in some desktop systems to.

To me is the only thing that makes sense on a HEDT systems of intel... of course games are scarce (but is getting there).

The fiasco of " Winblows 8/RT/Blue/Surface/Phone/XBone" is due to MSFT alienating all its partners with incredibly restrictive issues... as if they might be honored to support Windows 8 derivatives... and "PAY" an awesome homage for the privilege... fingers on the air was to be expected lol..

But MSFT is "backpedaling" fast and furious, the backpedaling with XBone is notorious lol...

 

[de5_Roy]

xbone(R) details slides:
http://semiaccurate.com/2013/08/26/xbox-one-details-in-pictures/

 

[Cazalan]

I think 32MB of ESRAM at 28nm could be notoriously smaller than that... and provide almost the double of bandwidth, triple better latency, and cherry on top of cake consume quite less power (its on tiny Wuii).

Assuming that's enough capacity. Intel has 128MB and they hit a wall at higher resolutions. I guess we will see once XBone ships just how good DDR3 w/32MB can be.

Another benefit of these chip contracts. They become excellent test cases for how they should evolve the APU.

 

[de5_Roy]

^^ i think crystal well's shaders are weaker than amd and nvidia's.

 

[Cazalan]

Has already picked some steam, Ubuntu is installed from origin in what with some vendors can be ~25% of all new laptops... and is installed in some desktop systems to.

To me is the only thing that makes sense on a HEDT systems of intel... of course games are scarce (but is getting there).

They've picked up steam literally.

Been watching this site and there is a fair amount of activity.

http://steamforlinux.com/?q=en

 

[iafro1989]

Good to see AM3+ isn't EOL, now time to install this sabertooth board lol.

 

[hcl123]

I think 32MB of ESRAM at 28nm could be notoriously smaller than that... and provide almost the double of bandwidth, triple better latency, and cherry on top of cake consume quite less power (its on tiny Wuii).

Assuming that's enough capacity. Intel has 128MB and they hit a wall at higher resolutions. I guess we will see once XBone ships just how good DDR3 w/32MB can be.

Another benefit of these chip contracts. They become excellent test cases for how they should evolve the APU.

But Intel is a totally different can or worms. Their IRIS will have a L4 cache, those supposed up to 128MB, 64MB now, eDRAM... is off die. It needs an additional serial interconnect only for this, and by being as a normal cache it needs quite large "flag files" on CAM structures on die, for all that large exterior cache (cache usually is only a few MB).

ESRAM doesn't have separated flag files, its totally on die, nothing off die (wuii estimation is ~40mm² for 32MB at 32nm) and is a cache on the GPU side, but visible from the CPU... will be quite smaller and quite less power hungry comparatively. Its on Wuii, though no one will benchmark those small portable consoles, i think we can say functions pretty well (Wuii is already in the market some time ago).

 

[8350rocks]

^^ i think crystal well's shaders are weaker than amd and nvidia's.

Beyond the shadow of a doubt they are...GCN 2.0 is likely the most advanced architecture that will appear for a while. Especially since NVidia just got done cancelling Maxwell (their next gen), and laid off quite a few engineers...

 

[hcl123]

xbone(R) details slides:
http://semiaccurate.com/2013/08/26/xbox-one-details-in-pictures/

Now we know that the 32MB ESRAM in XBone is divided into 4 banks of 8MB each, its 109GB/s min and 204GB/s peak... is not what it says in SoC components ? (2 image)

So all in all, XBone has more "peak" bandwidth than PS4 for the GPU. Doesn't make "that difference" because its only 32MB, 8GB GDDR5 will still be better for games, but the difference might be shorter than many think.

 

[de5_Roy]

Now we know that the 32MB ESRAM in XBone is divided into 4 banks of 8MB each, its 109GB/s min and 204GB/s peak... is not what it says in SoC components ? (2 image)

So all in all, XBone has more "peak" bandwidth than PS4 for the GPU. Doesn't make "that difference" because its only 32MB, 8GB GDDR5 will still be better for games, but the difference might be shorter than many think.

the second slide shows that the DRAM is off the die. iirc a teardown shows that the memory module is seperate on the pcb. why do you keep calling it ESRAM? if it's static ram, won't be on the die?

 

[hcl123]

^^ i think crystal well's shaders are weaker than amd and nvidia's.

Beyond the shadow of a doubt they are...GCN 2.0 is likely the most advanced architecture that will appear for a while. Especially since NVidia just got done cancelling Maxwell (their next gen), and laid off quite a few engineers...

Do we have any details of GCN 2.0 yet ?

From a layman (all more or less layman from the outside) POV, analyzing GCN 1.0, i think it has the *potential* of being the most advanced uarch around. One striking improvement i see with that design of a "scaler co-processor", is give it one more core, put their the branch calculation, the L/S and TMU filtering, "double pumping" (double frequency) this new enlarged "scaler core"... and then with some tiny "reservation stations" give it 6 SIMD cores (96 sp), double all the cache including the scratchpad (local store), and for a smaller size than 2 GCN CUs you could have the same performance or more in a "access/execute" approach.

I think Nvidia should go back and re-start from Fermi... for quite less "sp" it could have the same performance. But i think it wont please intel, it will be much less CPU bound about everything, and how much intel has influence and how much they are entangled in deals.

 

[hcl123]

Now we know that the 32MB ESRAM in XBone is divided into 4 banks of 8MB each, its 109GB/s min and 204GB/s peak... is not what it says in SoC components ? (2 image)

So all in all, XBone has more "peak" bandwidth than PS4 for the GPU. Doesn't make "that difference" because its only 32MB, 8GB GDDR5 will still be better for games, but the difference might be shorter than many think.

the second slide shows that the DRAM is off the die. iirc a teardown shows that the memory module is seperate on the pcb. why do you keep calling it ESRAM? if it's static ram, won't be on the die?

No.. from http://semiaccurate.com/2013/08/26/xbox-one-details-in-pictures/ 2th image... on the bottom of that image, "inside" the rectangle that represents the SoC chip, there are represented 4 8MB smaller squares, that represent ESRAM.... 32MB (4x8), and its clearly on die.

Seems 100% on the "mouche"...
http://www.vgleaks.com/durango-gpu-2/

[ UPDATE: and S|A is late if "~correct" rumor is concerned. One thing that seems to detach is the effective issue power... Durango is a 1.2+TFLOPS GPU, that can address each SIMD core separately, it can issue 768 threads per second without too much entanglements with wavefronts i think, so wasting less resources and augmenting the scheduling efficiency. ( that seems what this implies )

http://www.vgleaks.com/world-exclusive-durango-unveiled-2/

Don't now how that compares with PS4 efficiency... which has more sp and should be better at games, more if hUMA/HSA centered... but having 1.2+TFLOPS out of 68GB/s DRAM, to be really efficient, then ESRAM must provide a real awesome boost ]

 

[de5_Roy]

No.. from http://semiaccurate.com/2013/08/26/xbox-one-details-in-pictures/ 2th image... on the bottom of that image, "inside" the rectangle that represents the SoC chip, there are represented 4 8MB smaller squares, that represent ESRAM.... 32MB (4x8), and its clearly on die.

Seems 100% on the "mouche"...
http://www.vgleaks.com/durango-gpu-2/

you mean.. the 3rd image, right? at first i skimmed over thinking 5MB of some kind of cache. thanks.

edit: in my defense, the guy's hair was distracting...

 

[hcl123]

No.. the 2th from the "top", the slide that says "SoC components" ... it represents the entire APU SoC... it has a small rectangle on the top that says "CPU", the ESRAM is represented at the bottom in 4 pieces (8MB) close tight together.

and it says related with that structure; 109GB/sec min (minimal) ... 204GB/sec peak BW ... 4x 256bit read & write...
Meaning the ESRAM has an interface of 1024bit.. 4x more than 4 channels DDR3. (edt)

 

[de5_Roy]

i still blame the hair. ^w^

edit: btw, the die size says 363mm2 consisting of the soc along with ESRAM, so if you take out the custom blocks, to me it looks like a pretty doable i.e. marketable apu (got it, amd?). i'd rather see one like that than a new am3+ cpu.

 

[hcl123]

@juanrga

http://mygaming.co.za/news/hardware/53750-amds-plan-for-the-future-huma-fully-detailed.html

pertinent slide
http://mygaming.co.za/news/wp-content/uploads/2013/05/AMD-Kaveri-hUMA-benefits-600x297.jpg

Fully hardware coherent on GPU/APU ( pay attention to the distinction GPU and or APU)

* Probe filters and directories (coherency directories) will maintain power efficiency

so it has the same attributes of ccNUMA and some more, we can say hUMA is like Heterogeneous ccNUMA or HccNUMA LOL

[UPDATE: from 1th link (read first link article)

For bonus points, this works with discrete AMD graphics cards from the HD7000 family as well, although there will be some latency issues when accessing GDDR5 memory. Its possible to have the CPU, integrated GPU and the discrete GPU all working on the same thing.

Do i deserve "booos" if i say kind of seamless xfire is at hand ? ... even double or maybe triple xfire of discrete with an APU lol

Now what is really missing are HSA games

 

[8350rocks]

Now what is really missing are HSA games

Precisely this...it will be quite some time before that happens. Because OGL is certainly not getting rewritten anytime soon, and DX11.2 is not HSA compatible.

[NOTE: M$ does not have HSA in it's XBone console...this should be an indication of how much they care about this.]