AMD Piledriver rumours ... and expert conjecture

[Reynod]

We have had several requests for a sticky on AMD's yet to be released Piledriver architecture ... so here it is.

I want to make a few things clear though.

Post a question relevant to the topic, or information about the topic, or it will be deleted.

Post any negative personal comments about another user ... and they will be deleted.

Post flame baiting comments about the blue, red and green team and they will be deleted.

Enjoy ...

 

[gamerk316]

Well you need space ~somewhere~ to put the GPU. If you look at the layouts for modern day CPU's you see that 30~50% of the die space will be used for L3 cache, that is an insane amount of room that rarely gets fully utilizes on desktop / mobile applications. L3's job is to act as a fall back to a L2 cache miss, it's really only effective when your predictor is sh!tty or when your application's memory space is so large that the predictor can't fit the hits inside L2. You can see this design in action when comparing the amounts of L2 cache on various uArchs. Llano / Sabine (Stars APUs) have 1MB per core of L2 Cache with the Phenom II X4's having 512KB of L2 and 6MB of shared L3 (1.5MB per core). Intel SB's have 256KB of L2 Cache but a very low latency L3 cache of 6~8MB (3MB of mobile platform). BD's have 2MB of shared L2 cache per module (1MB per ~AMD core~) and 4~8MB of shared L3 cache.

What you see is that you can reduce the amount of L2 cache per core to reduce latency but only if you have a good amount of L3. If you take away the L3 then you better have decent L2 cache performance or you'll stall our your CPU. The Phenom II's (Stars K10 core) had pretty good L2 cache performance, in both latency and hit rate. To further buffer the Llano / Sabines from cache issues they increased their cache from the standard 512KB to 1MB, this has a greater impact then introducing 2~3MB of L3.

Now BD has serious issues with L2 / L3 cache latency and from what I can tell branch prediction. If they haven't fixed that prior to Trinity being released, then the PD based Trinity APU without L3 is going to have serious performance issues CPU wise. If they have fixed it then we'll have a solidly performing product to look forward to.

Agreed. I suspect that, from a space usage standpoint, if RAM speeds continue to catch up to CPU speeds, we will see the L3 eliminated, and possibly a significant reduction in the size of the L2. Remember, the CPU cache is basically a speed enhancement to get around the fact reading RAM takes so bloody long in the first place.

The main question, is what ELSE could be used in the freed up space, if it isn't being used for the CPU cache? More cores isn't necessarily the right answer, but simply removing the cache to lower power probably isn't acceptable either [though might make sense for mobile...].

 

[g4114rd0]

Detailed AMD 2012 Desktop Roadmap Leaked

http://vr-zone.com/articles/detailed-amd-2012-desktop-roadmap-leaked/14831.html

 

[de5_Roy]

if trinity's igpu is 7660d - i think that's more in line with llano numbering.
......wait..
i don't understand why the 'roadmap' says a8 3820 will have 6630 igp while 3870k's igp is 6550d. stronger igp is always nice though.
a4 3420 will have 6510d (i think it'll perform close to 6550d?) igp. it'll make the new a4 a very good apu for entry level htpc.
and the zambezi cpus are listed with radeon hd 6450+ graphics. '2c 2012' has fx 8150 with radeon 6450+ discreet graphics...?! this last bit seems like it is from a pc builder company's roadmap.

 

[Chad Boga]

I realize your an Intel fanboi and therefor don't do much research.

I realise you are an AMD fanboy and therefore don't do much independent thinking.

A whole pile of gibberish.

Why are you saying it is an illegitimate question to ask about the performance of Trinity's IGP?

Triny is going around saying it is going to be the greatest thing since sliced bread, so I simply asked about its performance.

No need for you to get your panties twisted into a knot

 

[esrever]

even if AMD re launches llano with the supply to meet demand, they will own the lower end laptop market. Trinity's performance only has to be slightly better.

 

[g4114rd0]

 [strike]
-...you need Take Care only with of APUs slides."

http://www.donanimhaber.com/islemci/haberleri/AMDnin-2012-icin-planladigi-yeni-nesil-Fusion-platformlari-detaylandi.htm 31 Mayıs 2011[/strike]

 

[-Fran-]

I realise you are an AMD fanboy and therefore don't do much independent thinking.

A whole pile of gibberish.

Why are you saying it is an illegitimate question to ask about the performance of Trinity's IGP?

Triny is going around saying it is going to be the greatest thing since sliced bread, so I simply asked about its performance.

No need for you to get your panties twisted into a knot

It isn't what you ask; it's how you ask it... =/

The sarcasm and irony alarms start overheating from activity when you post, so well... It's hard (really hard) to tell apart what's a honest inquiry from you or just a flame-starting-war question.

And just for the sake of the thread... There's no way to answer your question in a way "we" know you'll be satisfied. All we can do is guesstimate some numbers and that's it. Which it's done already, plus the i'm-a-believer slides from AMD, hahaha.

And paladin, you know calling someone a "fanboi" is like insulting their mother around tech forums... 😛

Cheers!

 

[g4114rd0]

31 Mayıs 2011 (very old info). Sorry for any inconvenience caused.

 

[-Fran-]

 
-...you need Take Care only with of APUs slides."

http://www.donanimhaber.com/islemci/haberleri/AMDnin-2012-icin-planladigi-yeni-nesil-Fusion-platformlari-detaylandi.htm

Shouldn't come as a surprise that they're using the same chipsets, but I still feel cheap shot'd, lol.

I mean, they should've at least improved the mobile chipset somehow ~___~

Cheers!

 

[triny]

Shouldn't come as a surprise that they're using the same chipsets, but I still feel cheap shot'd, lol.

I mean, they should've at least improved the mobile chipset somehow ~___~

Cheers!

I heard of an A85 Chipset

http://thingsfinder.com/news/amd-trinity-apu-specs-revealed/
what would be the extra I wonder

 

[palladin9479]

It isn't what you ask; it's how you ask it... =/

The sarcasm and irony alarms start overheating from activity when you post, so well... It's hard (really hard) to tell apart what's a honest inquiry from you or just a flame-starting-war question.

And just for the sake of the thread... There's no way to answer your question in a way "we" know you'll be satisfied. All we can do is guesstimate some numbers and that's it. Which it's done already, plus the i'm-a-believer slides from AMD, hahaha.

And paladin, you know calling someone a "fanboi" is like insulting their mother around tech forums... 😛

Cheers!

Yeah and around here that are quite a few AMD / Intel fanboi's who act like acknowledging the accomplishments of the other is heresy and punishable by death. Similar to the Apple zealots.

Absolute funny part is, several times in the past I've criticized AMD's design with BD and praised Intel's amazing engineers. I'm about the last thing from an AMD / Intel fanboi. If I had to pick a HW company to attach myself to it would be Sun, I really like working on Sparc's and they had brilliant ideas for high end parallel processing. Much of what we think is new today was thought up back in the early 90's by Sun, even the basis for Apple's iOS (NextStep) was originally worked on by Sun. Sun's management just failed to realize the potential it had and squandered it's opportunity's.

Since Oracle bought them I've had mixed feelings. On one hand Oracle is the Evil Empire, their worse then MS and Apple combined. OS security updates are offered for free, but only to those with have a support contract with them, and those contracts can be ridiculously expensive. HW drivers / BIOS updates are offered in a similar way, freely available on My Oracle Support, but you can only get access if you have a valid support contract number. On the other hand, Oracle has invested large amounts of money into new Sparc chips and platforms. Since purchase Oracle has finished the T3 series and released the T4. Their nice, really nice, we've decided to go with a bunch of T4-2 servers @256GB memory each for our next package upgrade (its a destructive load so we get the chance to upgrade HW).

A single T4 chip has,
2.85~3.0 Ghz
Eight cores with eight threads each CMT (think a souped up version of HT). 64 simultaneous threads can be processed per chip.
16kb/16kb for L1, 128kb for L2 and 4MB shared L3.
On-die HW crypto acceleration unit
On-die 10Gbps (10Ge) Ethernet interface.
On-die memory controller (upgraded from T3)
http://www.oracle.com/us/corporate/features/sparc-t4-announcement-494846.html

The thing's a beast. They took the T3 and juiced it up so that it performs both parallel tasks and single thread tasks well. The problem we had with the previous T2/T3's was their wide flat design allowed them to perform lots of threads simultaneously but single threaded performance situations were slow. Made it a great work horse for web-services / databases. This recent update improved those capabilities but also allows it to focus most of it's resources on a few threads if its being under utilized.

Anyways, enough with the sidebar on Sun equipment. I'm most definitely not an AMD fanboi lol.

 

[-Fran-]

I heard of an A85 Chipset

http://thingsfinder.com/news/amd-trinity-apu-specs-revealed/
what would be the extra I wonder

Well, that's from nov 2011, and in the new slides, there was no mention to the new chipset... Maybe they just canned it since it will add up unnecessary cost to OEMs and just a little extra kick (maybe).

Since the APUs are not meant to be "high end laptops", I guess a feature rich chipset is kind of a moot point. Well, not that the A75 is "weak", but I'd love to have eSata and USB3 instead of a bunch of USB2s; for that we need more PCIe lanes IIRC =/

Got anything past the analyst conference about the chipsets?

Cheers!

 

[Cazalan]

On-die memory wouldn't be a good idea, too limited a space that you could put other things.

Better idea is next big socket revision, add a back side bus for 128-bit GDDR5 that can be used for dedicated GPU memory. On a board you could introduce a small socket for a small 512MB ~ 2GB GDDR5 memory add on. If no add on is detected it use's UMA, if add on is detected then it use's the much faster add on.

Not on-die memory, Die-on-top-of-Die memory, aka 3D stacked.
This is the future of CPU/GPU/APU/SoC. The first generation ones are being made today but primarily for the mobile/smart tv market.

Jedec WIDEIO is a 512bit memory interface.

http://www.cadence.com/Community/blogs/ii/archive/2011/12/15/an-update-on-the-jedec-wide-i-o-standard-for-3d-ics.aspx

Then there is Micron/IBM Hypercube memory.

http://www.theregister.co.uk/2011/12/01/ibm_micron_hybrid_cube_memory/

"IBM and Micron say that HMC devices can take up about a tenth the space of traditional 2D DRAM memory sticks and require 70 per cent less energy to transfer a bit of data from the memory chip to the CPU."

That's a huge power savings meaning less heat and higher clock speeds.

It's rumored that Intel's is adding new plants specifically for this. To break further into the mobile and low end market by stacking 2-8GB of RAM on the CPU.

I've heard Micron talking of stacking 20 or more die as the process matures.
Moore's law is running out in the 2D space around the 10nm mark, that's why stacking is only logical.

 

[Cazalan]

Anyways, enough with the sidebar on Sun equipment. I'm most definitely not an AMD fanboi lol.

Yeah Sun had a good run. They just weren't prepared for how quickly AMD/Intel ramped with their 64bit architectures combined with Linux adoption for cheap compute farms.

It is good to see them shipping new Sparc designs. 64 threads/CPU is beastly for database work which made a good fit for Oracle. It helps keep Intel honest in the high end server market for sure.

 

[triny]

Well, that's from nov 2011, and in the new slides, there was no mention to the new chipset... Maybe they just canned it since it will add up unnecessary cost to OEMs and just a little extra kick (maybe).

Since the APUs are not meant to be "high end laptops", I guess a feature rich chipset is kind of a moot point. Well, not that the A75 is "weak", but I'd love to have eSata and USB3 instead of a bunch of USB2s; for that we need more PCIe lanes IIRC =/

Got anything past the analyst conference about the chipsets?

Cheers!

http://news.softpedia.com/news/Unofficial-AMD-Catalyst-12-1-quot-Cape-Verde-quot-Drivers-Get-Modded-Support-More-Cards-252003.shtml

This is newer and I found out A85 is xfire supprt which A55 and A75 lack
which might just mean Trinity at least on some models will be able to add more vcards hopefully with hd7770 for discrete
between 6790 and 6850 on 128 bit only? could be a noisy little runt
http://www.tomshardware.co.uk/AMD-Radeon-HD7770-HD7750-Cape_Verde,news-37728.html

 

[triny]

I realise you are an AMD fanboy and therefore don't do much independent thinking.

A whole pile of gibberish.

Why are you saying it is an illegitimate question to ask about the performance of Trinity's IGP?

Triny is going around saying it is going to be the greatest thing since sliced bread, so I simply asked about its performance.

No need for you to get your panties twisted into a knot

No,Trinity in all likelihood won't surpass my 2500K or whatever your running ,But I do think it will nix Intel's mobile market.
I do think they are marshaling new tech that will rattle Intel's foundation.
I weigh and measure what I spend my money on ,frankly Intel mobile has been weighed and measured and was found to be lacking.
I don't understand fanboyism it seems to be like all the other isms a.k.a communism fascism etc..

 

[de5_Roy]

i don't know about a55.. some a 75 motherboards support crossfirex (from x16+x4 to x8+x8 modes) along with (but not together e.g. 'triple graphics', afaik) dual graphics.
trinity's support for multi gpu cfx/sli (tri-fire with apu) won't be enough.
it will have to be capable of letting the gpus perform to their max without getting bottlenecked by the cpu(part of the apu).
amd will have to provide working, bug-free drivers with proper cfx support.
let's assume, for speculation's sake, that amd satisfies all these conditions (they're not obligated to, in reality). we're looking at a lot of graphics power at entry level, for example - a8/10 + a85 mobo + 2x radeon hd 7670 1 gb ddr3 (i think gddr5 versions won't be able to tri-fire with the apu).

 

[triny]

google translate:


Brazos and weighted with the code name as netbooks be to input level laptop computers, all oneMeanwhile, computers and Mini-ITX motherboards aimed at explaining the Fusion processors and more than 5 million sold production AMD announced that the demand on the capacity of the Sabine platform will introduce to the market.
Expected to be made official today promoted Sabine platform, code named LIano new Fusion processors with withA series of models includes the new chipset. Bobcat architecture platform that uses the Brazosdifferent as K10.5 architecture power at the same time the platform so more powerfu

AMD next year for Comal and Deccan, code-named new Fusion is preparing platforms.
AMD high performance class laptop and performance-class desktop computers computers for prepared under the Fusion platform, A4, A6 and A8 series processorsto be marketed. Mobile versions of the platform will be dual and quad core processors, integrated graphics processor units 400x parallel processing unit according to the modelfuture.
32nm production technology prepared and the above-mentioned as Used in the core design code-named K10.5 architecture, the Husky Winterpark, code-named processors and integrated graphics units are featured in Beavercreek. Soon Sabine platform market as we shall see future Comal place during the year to leave the platform.
Trinity, with code names in Weatherford, and Richland newPiledriver, code-named Fusion processors will use the core design. Bulldozer architecture, the processor will takeso more strong given in units of integrated graphics is being expressed, rather than FS1 FS1r2 socket structure will be passed butapparently remain the same chipsets.

Comal platform new Trinity, Weatherford and Richland will be code-named Fusion processors.
Platform for Comal one other important detailwill be accompanied on the external graphics cards.According to AMD's documentation, with Comal platfomrunalaptop external computers screen card asLondon carries the code name new I wonder who will be included in a series for This series is a new generation of 28nm GPU designs also specify that.
AMD future year with with input update the level of the Fusion platform. JanuaryThe Brazos platform launched in 2012, the location of the platform to leave Deccan. Bearing the names of Krishna and Wichita codenew Fusion processors, the current models different as Enhanced architecture was developed by Bobcat Bobcat to use instead.
2012'Will be presented in the Comal and Deccan platforms, will bear up the bar in terms of graphics capabilities. Technical about the properties net one information platforms will include, but notnewclassification of graphics processors into sharp focus. 2012for prepared high-end mobile processors, Trinity Fusion, the most high GPU performance will have the APU.

A60M and A70M chipsets will use a platform which will be released in 2012.
Performance and medium Weatherford and Richland is focused on the segment's Fusion processors with integrated graphics units will be included. More parallel processor, graphics processor designs are supposed to contain a large number of more efficient Turbo and thus are capable of dynamic frequency control implementations are expected to be more aggressive.
AMD 2012 fordetails prepared by the Comal and Deccan platforms were getting clear on the chipset.Visuals can be seento new platforms in the coming days that will be available at this point that can be used çipsetleriyle A60M and A70M correct also discussed compliance as looks.

No word on which discrete hd####

 

[triny]

i don't know about a55.. some a 75 motherboards support crossfirex (from x16+x4 to x8+x8 modes) along with (but not together e.g. 'triple graphics', afaik) dual graphics.
trinity's support for multi gpu cfx/sli (tri-fire with apu) won't be enough.
it will have to be capable of letting the gpus perform to their max without getting bottlenecked by the cpu(part of the apu).
amd will have to provide working, bug-free drivers with proper cfx support.
let's assume, for speculation's sake, that amd satisfies all these conditions (they're not obligated to, in reality). we're looking at a lot of graphics power at entry level, for example - a8/10 + a85 mobo + 2x radeon hd 7670 1 gb ddr3 (i think gddr5 versions won't be able to tri-fire with the apu).

I think crossfire means more than one discrete which wasn't supported on Lano
I don't see any information that 7670 is the only discrete ,I think 7770, 7670 ,7550 would be the discrete for the desktop models .I could well be wrong as information on the matter is at a premium .

 

[viridiancrystal]

I find it hard to believe that trinity is going to be clocked so high at base. In way to many cases that is not good for AMD. 1. It shows there are having large difficulty improving IPC on BD architecture. Which only further drives itself into the ground.
2. They either A. are not concerned about people mixing it up with higher level cpu's or B.are truly dropping out of the High end desktop market.
3. power draw. Unless this is all they worked on, i don't find it possible to clock 4 cores at 3.8 Ghz with a few hundred Radeon shaders and stay under 65 or 100w.
4. meeting demand. They had issues keeping up with Llano demand. We have also seen rumors of BD chips not able to reach the clock speeds and may be sold as a different chip. Along with their consisting issues with GF.

Thinking about it, i just think is doesnt make any sense to clock Trinity that high.

 

[fazers_on_stun]

IBM Fab Club to Reveal Details on 20nm, 14nm and Beyond Process Technologies Next Month.

http://www.xbitlabs.com/news/other/display/20120208232412_IBM_Fab_Club_to_Reveal_Details_on_20nm_14nm_and_Beyond_Process_Technologies_Next_Month.html

IIRC it was IBM who led their 'fan club' down the gate-first HKMG path, which apparently they are all regretting now 😛... I dunno of anybody who is keeping gate-first at 22nm and lower nodes.

IBM has a reputation of being long on process research and short on mass-production scaling issues.

 

[fazers_on_stun]

http://www.tested.com/news/researchers-develop-method-to-boost-cpugpu-hybrid-performance/3569/

up to 113% performance gain for APU
the future is APU
No wonder Intel is scrambling to improve igp

Hmm, I didn't see anything in that article that would prevent the same benefits to an Intel CPU with onboard GPU - seems like it would be a software issue, having the CPU do memory fetches for the CPU portion.

From what I have read, a far more massive improvement would be for the GPU to have a wide and low-latency pipe to its own dedicated memory instead of having to access it via an external bus such as PCIe. And coincidentally Intel appears to be far in advance of the techniques needed to accomplish just that - the silicon through-via interposer which permits low-power memory to be stacked directly on top of the GPU portion of the chip. IIRC Intel is already using that tech on other integrated circuits.

 

[esrever]

intel igps lack the raw compute power.

 

[triny]

I find it hard to believe that trinity is going to be clocked so high at base. In way to many cases that is not good for AMD. 1. It shows there are having large difficulty improving IPC on BD architecture. Which only further drives itself into the ground.
2. They either A. are not concerned about people mixing it up with higher level cpu's or B.are truly dropping out of the High end desktop market.
3. power draw. Unless this is all they worked on, i don't find it possible to clock 4 cores at 3.8 Ghz with a few hundred Radeon shaders and stay under 65 or 100w.
4. meeting demand. They had issues keeping up with Llano demand. We have also seen rumors of BD chips not able to reach the clock speeds and may be sold as a different chip. Along with their consisting issues with GF.

Thinking about it, i just think is doesnt make any sense to clock Trinity that high.

3.8 +turbo 3.0 ? xfire GCN ?

I wonder what excavator will be like 8840D igp 7990 xfire HSA

 

[jdwii]

I find it hard to believe that trinity is going to be clocked so high at base. In way to many cases that is not good for AMD. 1. It shows there are having large difficulty improving IPC on BD architecture. Which only further drives itself into the ground.
2. They either A. are not concerned about people mixing it up with higher level cpu's or B.are truly dropping out of the High end desktop market.
3. power draw. Unless this is all they worked on, i don't find it possible to clock 4 cores at 3.8 Ghz with a few hundred Radeon shaders and stay under 65 or 100w.
4. meeting demand. They had issues keeping up with Llano demand. We have also seen rumors of BD chips not able to reach the clock speeds and may be sold as a different chip. Along with their consisting issues with GF.

Thinking about it, i just think is doesnt make any sense to clock Trinity that high.

This is the way i feel also, I'm thinking A little worse though i'm thinking the CPU portion of trinity might even be slower then Llano at certain things. Like i said before the scaling of a BD core is 80% so times that by 4 and you have a 3.2 Core processor. Not to mention the IPC is easily 10-15% Lower on BD(note why the 6 core BD is about even with the 4 core Phenom in multithreading test's). Unless Amd can fix these things i don't see it being any faster then Llano's CPU. Again i do feel however Trinity will have at least 30% better graphics performance. I also feel Trinity will have better Performance/Per watt when compared to Llano. Also i'm strictly talking about the Laptop versions of Trinity and Llano, Because on these versions its harder to just Clock the CPU real high because of power consumption and heat.