AMD CPU speculation... and expert conjecture

[juanrga]

Why not see what you can find on cache latency for this new part ... thats always an interestign discussion because Intel do it so well ... and AMD have some clear ground to make up, as it obviously detracts from individual core IPC.

I believe that the cache behavior in AMD processors is due to Intel owning some crucial cache patents for the optimal design.

 

[Reynod]

Buddy your statement is complete rubbish ... there are a myriad of ways to optimise the cache and I know for a fact that Intels superior cache design has more to do with the superior prefetch design, along with optimised transistor design that AMD has just not managed to match. Intel's transistors in the cache basically switch quicker, and still function well thermally. AMD's cache is simpler and slower ... they have less engineers on the job wrt design and optimisation.

I am an AMD fan ... but also a realist.

I don't go around blaming the opposition for my own woes in design.

Whn you make fanboi comments like this you leave yourself open to getting slammed by others.

 

[Antaredja Kartasasmita]

I REALLY hope AMD makes steamroller on FX cpus

 

[juanrga]

And NVIDIA fires a major shot here:

http://www.anandtech.com/show/7622/nvidia-tegra-k1/3

NVIDIA's T1 GPU, the GPU being used in Project Denver, their ARM SoC, is basically just a cut down Kepler GPU. Farther, all their new designs, including GeForce, are going to start out as mobile designs.

This is significant. NVIDIA just got the most developer support for mobile, since they support the FULL OGL/OGL ES/DirectX/CUDA instruction sets. Makes it much more attractive for traditional PC gaming studios to make a low-cost mobile offering. Makes Mantle that much of a harder sell now, since it won't be a player in mobile.

I'm still waiting for AMD to make a mobile SoC. I'm stunned they haven't gone this route yet, especially since, via Trinity, they should be best positioned for it.

I am one of the guys impressed by that. The efficiency is incredible. Their marketing dept calling it super-chip is fine. Tegra K1 (6W) is able to offer almost one-half of the performance of the top 95W Kaveri (A10-7850) in a mobile TPD

It is interesting to compare the Denver core to Steamroller/Excavator. Kaveri CPU maxes out at 118 GFLOPs. The v1 version of Tegra K1 will use quad A15 cores at 2.3GHz. This implies a maximum of 74 GFLOPs. This is a mobile CPU with a 63% of the performance of a top desktop Steamroller

The v2 version will use dual Denver cores. This implies that each Denver core will be about 2x faster than A15 core. This implies that each Denver core can do up to 16 FLOPS/Clock, which is at the Sandy Bridge level. Steamroller only makes up to 8 FLOPS/Clock. If rumors are correct Excavator will do up to 16 FLOPS/Clock. This means that Nvidia first CPU cores are ahead of AMD ones.

i thought nvidia's demo was very weak and lacking. it rehashed the old stuff and made a big deal about project denver only to reveal it as a 64-bit dual core cpu and a 32bit version with a15 cores. wth, not even a53? did apple really node-block nvidia from 20nm?(rhetorical question, idc) the 64bit soc had a lot of cache but i don't think it'll help in 'wide' tasks.

i am a bit disappointed at only "45% better perf" information. i was expecting more on mantle (unless i've missed it). i blame call of .. i mean bf4.

afaik, amd's best shot at mobile soc will come around the time they launch seattle, prolly with a57+gcn 2.0 cores soc built on 20/28nm. jaguar/puma can't go lower power without sdp (intel's latest scam that amd joined), gcn 1.0, 1.1 don't seem too friendly towards low power. er.. assuming by mobile you mean smartphone soc.

The A53 core is optimized for throughput not latency, which makes it a bad option for a CPU. That is why AMD will be using A57 cores. Neither it makes sense for Nvidia use 64 bit cores when have ready their own 64 bit core. Their option to use a 32 bit core for legacy stuff and update to 64 bit latter makes all sense.

The 45% may disappoint you but I mentioned several times I expected between 30% and 50% for BF4. Future games will make more draw calls and the gap between MANTLE and DX will increase. AMD showed the next demo benchmarking

where Kaveri can by 3x faster under MANTLE. I would expect that kind of performance in future games designed with MANTLE in mind. BF4 is a 'patched' ordinary game.

The APU basically replaces the older SoC concept. Same deal though. The issue is, with a CPU+GPU on the same die, you have limited space to put other enhancements to improve performance. You'll never get top CPU or top GPU performance as a result. Dedicated CPUs and dedicated GPUs aren't going anywhere.

No. In fact the fastest supercomputers designed by both AMD and Nvidia use APUs/SoCs because that kind of ultra-high-performance (20 TFLOP)* cannot be achieved with a traditional dCPU+dGPU. The laws of physics are very clear. The evolution will be:

dCPU + dGPU ---> APU + dGPU ---> APU

Intel is also moving the next Xeon Phi generation from PCIe card to socket form.

* Note that a GTX Titan offers only 1.5 TFLOP.

 

[juanrga]

And jdwii, he is not far off. The APU was not designed with just cheap gaming in mind. They want HSA (what he is talking about with the CPU and GPU doing the work together) and as well are replacing entry and mainstream level builds with it. You can't find a mITX AM3 mobo or microATX AM3 mobo with more than 2 slots of RAM. Mainly FX is to be like Intels LGA2011. Top of the line. FM2 is to be everything else and I would not be surprised if they don't kill even top end and use only APUs.

Did you read what I read from Rory Reed? When he stated desktops have more than enough power? He isn't interested in giving you a better CPU than Intel. He is focusing on whatever will make them money instead which is sad as now Intel has no reason to keep their CPUs pushing as fast as they were. When Core 2 hit it was massive. Then Nehalem jumped again. Then SB was even better. But IB and Haswell are mostly not that great except in certain situations and now Broadwell is not supposed to be out until the 2H of 2014 when every step was supposed to be Q1 of the year.

As I said, AMD is focusing on money more than competing with Intel.

Rory Read words have been mentioned in this thread often. He has stated very well that he want transform AMD into a SoC company. Future plans about CPUs have been known also:

From a product standpoint, AMD is really focusing on its mainstream and entry level APUs. Rory didn't come out and say it here but no where in AMD's future direction is a focus on the high-end x86 CPU space.

This slide from AMD financial day was also given before

Where clearly says that AMD focus will be efficiency, SoCs, and HSA.

 

[de5_Roy]

it's time for newbie math...again...
kaveri 7850k die holds 2.41 billion transistors. afaik amd claimed that they allocated 47% or the die for igpu. that leaves 43% for cpu, imc, northbridge, cache etc. roughly 1.28 billion transistor for cpu et al. i don't know how much area the cpu cores take up, but from the dieshot, i assume less than 30% incl. L2 cache...? so that comes to 0.7B for 4 core+ cache. 1.4B for 8 core cpu and more for L3 cache... wait wth... this seems doable for amd while maintaining reasonable yields provided that kaveri yields are good. as for speculating performance of such cpu, it'll depend on how much internal bandwidth translates to performance.

Spoiler

wow, that almost meltted my brain! if the math is off, blame my monitor. but it made me realize that an 8C SR-FX with a 512 shader "big apu" made on 28nm shp node might not be possible to maintain under reasonable tdp and yields. replace sr with puma, and it looks doable - and tsmc's LPH (or is it HPM) nodes' yields look better than glofo's.

however, making a pure cpu while hyping all the hsa-related stuff is contradictory. amd doesn't have a very fast bus to communicate with a discreet gpu. i mean, it may have one but pcie is industry standard and who is going to ditch pcie for something else?

 

[juanrga]

Buddy your statement is complete rubbish ... there are a myriad of ways to optimise the cache and I know for a fact that Intels superior cache design has more to do with the superior prefetch design, along with optimised transistor design that AMD has just not managed to match. Intel's transistors in the cache basically switch quicker, and still function well thermally. AMD's cache is simpler and slower ... they have less engineers on the job wrt design and optimisation.

I am an AMD fan ... but also a realist.

I don't go around blaming the opposition for my own woes in design.

Whn you make fanboi comments like this you leave yourself open to getting slammed by others.

I wrote "I believe", I didn't affirm it as fact. If I am wrong about caches then I am glad to know, but the rest of your post is a gratuitous personal attack. Or do I want me to call you "fanboi" each instance you are wrong about something?

 

[qurious69ss]

Dude I see you on every forum spewing numbers that keep getting refuted over and over but you still keep posting the same thing over and over, I mean come on, when even Amd fans on even SA tell you that your wrong you should probably it a rest.

 

[juanrga]

What numbers?

 

[Reynod]

Well I know Charlie and Copper pretty well ... good people ... bacon is popular there.

 

[jdwii]

Buddy your statement is complete rubbish ... there are a myriad of ways to optimise the cache and I know for a fact that Intels superior cache design has more to do with the superior prefetch design, along with optimised transistor design that AMD has just not managed to match. Intel's transistors in the cache basically switch quicker, and still function well thermally. AMD's cache is simpler and slower ... they have less engineers on the job wrt design and optimisation.

I am an AMD fan ... but also a realist.

I don't go around blaming the opposition for my own woes in design.

Whn you make fanboi comments like this you leave yourself open to getting slammed by others.

"cache patent" either way the slower cache is probably the main reason why the memory controller is weaker compared to Intel's by around 20%.
I suspect the cache problem has to do with the design letting the cores share cache within a module that is the main issue in my eyes. The second issue has to do with the longer pipeline in Bulldozer-steamroller which yes causes more cache issues. And then we come to your reason which is right as well and this also has to do with the longer pipeline when they have a longer pipeline the prefetcher has to be superior to overcome issues.

 

[juggernautxtr]

Buddy your statement is complete rubbish ... there are a myriad of ways to optimise the cache and I know for a fact that Intels superior cache design has more to do with the superior prefetch design, along with optimised transistor design that AMD has just not managed to match. Intel's transistors in the cache basically switch quicker, and still function well thermally. AMD's cache is simpler and slower ... they have less engineers on the job wrt design and optimisation.

I am an AMD fan ... but also a realist.

I don't go around blaming the opposition for my own woes in design.

Whn you make fanboi comments like this you leave yourself open to getting slammed by others.

"cache patent" either way the slower cache is probably the main reason why the memory controller is weaker compared to Intel's by around 20%.
I suspect the cache problem has to do with the design letting the cores share cache within a module that is the main issue in my eyes. The second issue has to do with the longer pipeline in Bulldozer-steamroller which yes causes more cache issues. And then we come to your reason which is right as well and this also has to do with the longer pipeline when they have a longer pipeline the prefetcher has to be superior to overcome issues.

I thought they fixed some of the lengths when they moved the "cores" in piledriver/vishera. they were on the inside and moved to the outer edge's of the module,which also gave it better heat control which also affected turbo ability?

 

[jdwii]

Dude I see you on every forum spewing numbers that keep getting refuted over and over but you still keep posting the same thing over and over, I mean come on, when even Amd fans on even SA tell you that your wrong you should probably it a rest.

Buddy your statement is complete rubbish ... there are a myriad of ways to optimise the cache and I know for a fact that Intels superior cache design has more to do with the superior prefetch design, along with optimised transistor design that AMD has just not managed to match. Intel's transistors in the cache basically switch quicker, and still function well thermally. AMD's cache is simpler and slower ... they have less engineers on the job wrt design and optimisation.

I am an AMD fan ... but also a realist.

I don't go around blaming the opposition for my own woes in design.

Whn you make fanboi comments like this you leave yourself open to getting slammed by others.

"cache patent" either way the slower cache is probably the main reason why the memory controller is weaker compared to Intel's by around 20%.
I suspect the cache problem has to do with the design letting the cores share cache within a module that is the main issue in my eyes. The second issue has to do with the longer pipeline in Bulldozer-steamroller which yes causes more cache issues. And then we come to your reason which is right as well and this also has to do with the longer pipeline when they have a longer pipeline the prefetcher has to be superior to overcome issues.

I thought they fixed some of the lengths when they moved the "cores" in piledriver/vishera. they were on the inside and moved to the outer edge's of the module,which also gave it better heat control which also affected turbo ability?

Well i never heard of that could be true however, Amd did use resonant clock mesh, which allowed for higher clock rates and lower temps.

 

[blackkstar]

Understanding Juanrga is a lot more easier if you realize that he's basically the first real ARM fanboy to defend ARM so heavily.

The footnotes in the Nvidia slides where they're making all these claims are laughable at best. Lines like "estimates" and "20 percent margin of error" and "PS3 only using one core and no SPUs".

Nvidia always does this with Tegra. They hype it up like none-other. In fact, they even boasted Tegra 2 was on 360 levels and you'd have the power of a console in your pocket. And look how that turned out.

The truth is Tegra K1 final silicon is a long ways away, at least 10 months. We will be talking about Excavator at that point, and the other ARM companies wilil probably be with ARM64 then too.

Nvidia always starts hyping Tegra way before it's actually due. They're very good at marketing, and if Qualcomm or Samsung were out here hyping their ARM products that would ber coming out around the time of Tegra, people would be going crazy too.

 

[jdwii]

Its arm and like i continue to say the most they will have is 5% market share in 3 years and they cannot say that will be from current markets probably newer markets which X86 would of never been feasible from a cost perspective anyways. We will have another 20 years with X86 being the dominant choice. Windows 8 on Arm failed horribly, and like i said someone from Amd stated Arm was a side project. Which probably means it looks good for investors at the moment and may take off. There will be many other’s already producing Arm products and really what can Amd bring to the table that others can't? GCN architecture (GPU), Arm+X86 support(bigger one).
Will Arm ever get popular and control 50% of the market? Or will Arm only be used in small form factors and low-end servers?
Arm processors in a server could work quite well and even more so if we can get server applications to have more support for GPU processing. Why devote so much power to the CPU when the GPU could do most of the heavy lifting?

 

[Steve Faugue]

I found some info about the info on Kaveri packaging, specifically the A10-7850K:

It confirms a 1 GB DDR3 frame buffer in the previously quoted pic.

I also have some new info in regards to the upcoming beta drivers (14.1 BETA). It is looking to be the big Mantle-enabling driver which will also enable full TrueAudio support. It also included updated Linux support, temperature-based OverDrive, frame-pacing improvements, and DirectX 11.2 (for the HD 7000 series and beyond on Windows 8.1).

Source
Source

 

[juanrga]

Understanding Juanrga is a lot more easier if you realize that he's basically the first real ARM fanboy to defend ARM so heavily.

The footnotes in the Nvidia slides where they're making all these claims are laughable at best. Lines like "estimates" and "20 percent margin of error" and "PS3 only using one core and no SPUs".

Nvidia always does this with Tegra. They hype it up like none-other. In fact, they even boasted Tegra 2 was on 360 levels and you'd have the power of a console in your pocket. And look how that turned out.

The truth is Tegra K1 final silicon is a long ways away, at least 10 months. We will be talking about Excavator at that point, and the other ARM companies wilil probably be with ARM64 then too.

Nvidia always starts hyping Tegra way before it's actually due. They're very good at marketing, and if Qualcomm or Samsung were out here hyping their ARM products that would ber coming out around the time of Tegra, people would be going crazy too.

You have used my comments about Denver core made in another part to launch here an anti-Nvidia rant... and you call "fanboy" to others! LOL

But at least this time you are spelling my nickname correctly. Thanks!

 

[juanrga]

and like i said someone from Amd stated Arm was a side project.

Yes you said he said. The problem is that AMD head says otherwise.

 

[juanrga]

I found some info about the info on Kaveri packaging, specifically the A10-7850K:

It confirms a 1 GB DDR3 frame buffer in the previously quoted pic.

I also have some new info in regards to the upcoming beta drivers (14.1 BETA). It is looking to be the big Mantle-enabling driver which will also enable full TrueAudio support. It also included updated Linux support, temperature-based OverDrive, frame-pacing improvements, and DirectX 11.2 (for the HD 7000 series and beyond on Windows 8.1).

Source
Source

The new catalyst drivers are being also included in SteamOS. Steam machines now support Nvidia, Intel and AMD hardware officially.

 

[Reynod]

Thanks Juan.

There is a 30W TDP difference between the 7850K and the 7800 with the same GPU cores and clocks and only a slight decrease in CPU clockspeed (200Mhz - 100Mhz on turbo).

 

[de5_Roy]

MSI's new Socket FM2+ boards are primed for Kaveri
http://techreport.com/news/25902/msi-new-socket-fm2-boards-are-primed-for-kaveri
Tiny Steam machine with AMD guts coming from Maingear
http://techreport.com/news/25900/tiny-steam-machine-with-amd-guts-coming-from-maingear
^^must release a mobile kaveri version of this.

Spoiler

I found some info about the info on Kaveri packaging, specifically the A10-7850K:

It confirms a 1 GB DDR3 frame buffer in the previously quoted pic.

I also have some new info in regards to the upcoming beta drivers (14.1 BETA). It is looking to be the big Mantle-enabling driver which will also enable full TrueAudio support. It also included updated Linux support, temperature-based OverDrive, frame-pacing improvements, and DirectX 11.2 (for the HD 7000 series and beyond on Windows 8.1).

Source
Source

amd benchmarking labs ought to pick right competitors. this is poor benchmarketing - picking on gt630 is... mean. in that bf4 (possibly offline single-player, scripted) benchmark, the fps bars are so amplified that without the numeric values, you'd never realize that hd4600 is just 1fps slower than gt630 (why even bother using it, then?) in that bench. at least intel had the audacity to aim at gt650m (and not succeed).
amd should compare a10 7850k against core i3 4xxx or core i5 4570R(iris pro), since there's no 'proper' performance-competitor for kaveri at it's max. price point. mobile comparions are going to be even more interesting since both amd and intel will be bound by thermal restrictions.

edit: amd should also start selling kaveri to emerging markets asap, and not pull another fx6300 this time. china has recently opened up it's markets for legally selling consoles, where amd sales have started to improve due to apus.

 

[juanrga]

A hand-made review of Kaveri

http://foro.noticias3d.com/vbulletin/showthread.php?t=421821

work in progress...

 

[palladin9479]

I'm still waiting for AMD to make a mobile SoC. I'm stunned they haven't gone this route yet, especially since, via Trinity, they should be best positioned for it.

I never really understood that from AMD. APU's will never compete with dGPU / dCPU's due to cost and space / heat limits, physics is a hard master. What they do well is putting a lot of power in a very small space / heat window which works amazingly well for mITX design's (and I mean real mini-ITX not those micro-ATX box's labeled as mini-ITX). Can scale it down to work on smaller devices, assuming you use Android and not Windows.

 

[etayorius]

A hand-made review of Kaveri

http://foro.noticias3d.com/vbulletin/showthread.php?t=421821

work in progress...

Hmm horrible CPU performance, ¡t looks like in the end Kaveri is not worth the upgrade for me anymore... oh well, ill just stick to my current CPU (980BE at 3.9Ghz) and get a R7 270X instead and a new HDD and more Ram instead of a new FM2+ system with Kaveri.

There are no Single Core benchmarks in there, but the Muti Threaded is horrible, ill be the first to declare i was wrong, i was expecting more, and kaveri just does not cut it for me... from what i understand in those benchmarks Kaveri barely beats Trinity, were not even talking Richland here.

 

[Baralis]

I have wondered recently why they do not use a double sized socket with a full CPU and GPU solution for APU. It looks like cooling with a new extra large style CPU cooler would work. Maybe I am missing something?