AMD CPU speculation... and expert conjecture

[juanrga]

GPUs also have cache. GDDR5 has greater latencies than DDR3. Maybe HBM would solve all those problems, because right now they are better than GDDR5 and DDR3/4 in every aspect except price.

Technical specifications of DDR3 and GDDR5 modules show that the latencies are about the same. What happens is that the memory controllers of dGPUs are optimized for throughput, not latency, because a GPU is a TCU not a LCU. Thus when you measure the latency of GPUs you are measuring the higher latency of the memory controller not of the modules.

 

[juanrga]

so you guys are back on the 2020: GPU extinction topic again, hilarious. i just bought a gtx970 2 days ago, and i doubt any APU can match this card in 6 years time.

it's not like GPUs will stay @ GDDR5 forever even, so can anyone explain how stacked DDR4 APUS will catch up against stacked (volta onwards) GDDR6 (or 7)?

The APU recently announced by AMD for the extreme-scale supercomputer targets 10 TFLOPS of performance (DP) and 1 TB of memory. Your GTX-970 has only 109 GFLOPS and 4 GB of memory, just saying it...

Also "stacked DDR4", GDDR6, and GDDR7 don't exist. The JEDEC standard replacement for GDDR5 is HBM.

AMD future APUs will use HBM plus DDR4

 

[juanrga]

Maybe with a die shrink they could fit 4-6 sr/ex cores onto a 290, a bit of hbm and bam your 970 is matched within a year?

The 14nm node (Glofo) will bring about 2.15x more density than the current 28nm node. On 14nm one can fit 8 SR cores and 16 GCN cores on a ~240 mm2 die. On 480 mm2 one can fit 8 SR cores and 48 GCN cores; that is more graphics performance than R9-290X. It is technically possible to develop an enthusiast APU, but AMD will not cannibalize its own dGPU line so abruptly; the transition will be smooth...

 

[8350rocks]

Anyone see the dragon age inquisition techspot review/benches?

7970 ghz does better than a gtx 770, and r9 290 matches 780ti, while gtx 980 is 3 fps over 290x.

Also, all dual core processors, from any manufacturer, were unable to play the game. They experienced a locked system with 100% core load on both cores at launch screen.

But you still just need a dual core right? Also FX8350 matches 4770k from 3.5 to 4.5 ghz. Game is clearly gpu bound with 290x and 980 only hitting 56 and 59 fps @ 1200p on ultra.

 

[de5_Roy]

^^ this?:
http://www.techspot.com/review/921-dragon-age-inquisition-benchmarks/page6.html
y'know what this game needs? mantle or dx12 for the dual core cpus.

 

[noob2222]

The benefit to those customers, the two companies assert, will be that they'll have access to the same process and the same design tools to make use of it in multiple locations around the globe. Those customers can "save literally hundreds of millions of dollars in design costs," says Hunter.

This pretty much sais to me that AMD will be using whichever fab is up and running when they place the order. We will see if your prediction of GF being on time will turn out.

with AMD having the option to move away from gf for 14nm, GF had little option but to ask Samsung for help. Take it how you want, GF doesnt have the track record.

 

[esrever]

^^ this?:
http://www.techspot.com/review/921-dragon-age-inquisition-benchmarks/page6.html
y'know what this game needs? mantle or dx12 for the dual core cpus.

And people say the Phenom II x6 1100T was not worth upgrading. It gets beaten out by the x4 bulldozer.

 

[jdwii]

I still find it silly to state dgpu will be gone when they can still make a dgpu with stacked memory and have more transistors because they don't have to worry about the CPU taking up space. With ray tracing and multiple monitor 4k 120fps gaming they will need all the space they can get

 

[jdwii]

Edit post toms has an issue with mobile devices

Comment on the bulldozer taking phenom down its quite obvious its over optimization, I also notice amd 6 core fx is really close to that 8 core meaning scaling could still be better i5 still takes the lead

 

[juanrga]

I still find it silly to state dgpu will be gone when they can still make a dgpu with stacked memory and have more transistors because they don't have to worry about the CPU taking up space. With ray tracing and multiple monitor 4k 120fps gaming they will need all the space they can get

No. At 28nm about one half of the Kaveri APU is GPU and making a separate dGPU taking all the die space makes sense. But further die shrinks reduce the percentage of the die used by the CPU.

At 10nm the CPU takes a very small percentage of the whole die. Do you know how much of the Nvidia extreme-scale SoC is used by the CPU? Only the 5% of the die is used by the 8-core CPU. The remaining 95% of the die is GPU. This means that Nvidia making a separate dGPU would bring only 5% more transistors... but then those extra transistors would be useless due to an interconnect bottleneck...

 

[juanrga]

The benefit to those customers, the two companies assert, will be that they'll have access to the same process and the same design tools to make use of it in multiple locations around the globe. Those customers can "save literally hundreds of millions of dollars in design costs," says Hunter.

This is a consequence of licensing.

This pretty much sais to me that AMD will be using whichever fab is up and running when they place the order. We will see if your prediction of GF being on time will turn out.

with AMD having the option to move away from gf for 14nm, GF had little option but to ask Samsung for help. Take it how you want, GF doesnt have the track record.

A pair of years ago when I said that Kaveri was built on bulk instead SOI, I explained that one of the reason to move away from SOI was the ability to pick different foundries. Then people did call me names, but the strategy has shown to be sound. AMD follows the same strategy now. Why do you believe that AMD has been mentioning both 14nm (Glofo/Samsung) and 16nm (TSMC) in last conferences? Because they keep the foundry options open...

Said that, at the time of writing this, Samsung starts mass production of the 14nm node. This means two things: (i) that Samsung has virtually caught Intel and (ii) that GF will have 14nm node ready for Zen/K12 to debut on 2016...

 

[Reepca]

What I mean is that I've read that iGPUs in traditional rendering scenarios (where it's being used specifically as a graphics processor) are bottlenecked by system memory bandwidth (hence why faster RAM helps such a fair bit). Would this limitation apply to other compute cases, that aren't traditional rendering scenarios?

Anything that needs to transfer data to the GPU over the main memory bus is going to be affected. Why do you think we've added so much cache each CPU generation? It's a hack to work around slow memory access times. dGPUs get around this by putting GB's of super-fast RAM on die, and you only have the initial latency of the first memory transfer into the VRAM. That's the primary reason dGPUs will ALWAYS outperform iGPUs/APUs, unless we ever go to putting memory directly on the CPU (which is CRAZY expensive to do).

Could you (or someone else) please explain to me why system memory (the shared memory an APU would use) can't use the "super-fast RAM" dGPUs use? If it's plausible to have 8GB of GDDR5 memory on a dGPU that alleviates the data transfer bottleneck, why wouldn't it be plausible to use 8GB GDDR5 (or its equivalent) to alleviate the iGPU data transfer bottleneck? '

 

[juanrga]

What I mean is that I've read that iGPUs in traditional rendering scenarios (where it's being used specifically as a graphics processor) are bottlenecked by system memory bandwidth (hence why faster RAM helps such a fair bit). Would this limitation apply to other compute cases, that aren't traditional rendering scenarios?

Anything that needs to transfer data to the GPU over the main memory bus is going to be affected. Why do you think we've added so much cache each CPU generation? It's a hack to work around slow memory access times. dGPUs get around this by putting GB's of super-fast RAM on die, and you only have the initial latency of the first memory transfer into the VRAM. That's the primary reason dGPUs will ALWAYS outperform iGPUs/APUs, unless we ever go to putting memory directly on the CPU (which is CRAZY expensive to do).

Could you (or someone else) please explain to me why system memory (the shared memory an APU would use) can't use the "super-fast RAM" dGPUs use? If it's plausible to have 8GB of GDDR5 memory on a dGPU that alleviates the data transfer bottleneck, why wouldn't it be plausible to use 8GB GDDR5 (or its equivalent) to alleviate the iGPU data transfer bottleneck? '

It is possible to have GDDR5 as system RAM. Example? The PS4, which has 8GB of GDDR5. Which is then the problem for PCs? Basically that ordinary GDDR5 doesn't come as DIMM modules but has to be soldered to the board, which implies you couldn't upgrade memory size or repair by yourself if some module fails. Moreover it would increase costs because motherboard makers would do different versions of same motherboard: one with 2GB memory, one with 4GB, one with 6GB, one with 8GB...

AMD tried to fabricate an 'enthusiast' version of Kaveri. This was planned to have three Steamroller modules, a bigger GPU and would use GDDR5M memory as system RAM. GDDR5M is a special version that comes in SODIMM format

http://www.brightsideofnews.com/2013/03/05/amd-kaveri-unveiled-pc-architecture-gets-gddr5/

http://www.thingsmicro.com/2013/09/amd-kaveri-gddr5m/

But this had to be finally canceled because the company fabricating the SODIMM modules was out of business.

 

[Reepca]

Thanks for the explanation Juan. I guess that part of the issue is that I remember reading somewhere that DDR4 and GDDR5 were the same thing (which, apparently, they were not) - and DDR4 sticks are available now. Out of curiosity, though, how big of an improvement in memory bandwidth would DDR4 bring, if we start seeing any FM2+ (or whatever Carrizo is on) boards with support for it? Would it un-bottleneck the a10-7850k?

 

[8350rocks]

If you had something like supported ddr4 3200 or similar, it would help quite a bit. Though other factors begin to come into play limiting the effectiveness of simply throwing more bandwidth at the issue. For example HTX would have to be upgraded to the 3200 spec, which could be done fairly easily since IBM already used it quite a bit with POWER7

 

[8350rocks]

@jdwii:

Actually, Intel only wins @720p, and nothing was ever unplayable in a quad core or better from both sides, some were just less playable than others. They commented that anything over 30 fps was sufficient with the engine.

 

[8350rocks]

@juanrga:

You are commenting on a design meant only to babysit large gpu portions. Look at intel at 14nm, a larger portion of their die is still cpu because that is what is required for the designed workload. A very small serial processor designed to power large arrays of parallel processing nodes has a different workload than your average server, or consumer core. When you talk about fluid dynamics and such, there is virtually no serial computation to such calculations. Hence the sucess of things like folding@home.

 

[palladin9479]

dGPU's aren't going anything in the next 10~15 years, physics is what it is. We've had this discussion already.

SOC / APU type designs are best when your power / space / cost envelope is such that a discrete GPU isn't feasible. The onboard graphics co-processor is sufficient to get the job done. The vast majority of work the consumer does won't be utilizing the iGPU as a vector coprocessor, it's an entirely different problem set that it caters to. AMD seems to be focusing on the value segment, which makes sense for them. Rumors are that another high performance desktop chip will be out, we'll see how that goes if / when it does happen.

On a side note, recently upgraded my rig to a Asus Maximus Formula VII with an i7-4790K. Time to put the fx8350 out to pasture.

 

[palladin9479]

What I mean is that I've read that iGPUs in traditional rendering scenarios (where it's being used specifically as a graphics processor) are bottlenecked by system memory bandwidth (hence why faster RAM helps such a fair bit). Would this limitation apply to other compute cases, that aren't traditional rendering scenarios?

Anything that needs to transfer data to the GPU over the main memory bus is going to be affected. Why do you think we've added so much cache each CPU generation? It's a hack to work around slow memory access times. dGPUs get around this by putting GB's of super-fast RAM on die, and you only have the initial latency of the first memory transfer into the VRAM. That's the primary reason dGPUs will ALWAYS outperform iGPUs/APUs, unless we ever go to putting memory directly on the CPU (which is CRAZY expensive to do).

Could you (or someone else) please explain to me why system memory (the shared memory an APU would use) can't use the "super-fast RAM" dGPUs use? If it's plausible to have 8GB of GDDR5 memory on a dGPU that alleviates the data transfer bottleneck, why wouldn't it be plausible to use 8GB GDDR5 (or its equivalent) to alleviate the iGPU data transfer bottleneck? '

There is nothing technically wrong with using GDDR5 memory as system memory, but there are some practical implications. Firstly it's significantly more expensive then DDR3, it has greater latency due to how it's command buffer works, and while a DIMM specification exists for it, I have never seen anyone manufacture it. The memory tends to be soldiered onto the devices that use it which limits your memory expansion capabilities. So it could work in an appliance but problematic to work inside a general purpose desktop.

As for the whole "stacked ram!!!" folks, there are even bigger problems when you try to put that onto a die. There is only so much physical space, so no matter what you can put on a CPU you can put 10~16X that amount on a cheap DIMM. And while memory doesn't get particularly hot when compared to CPU/GPU's, it does use power and thus generate thermal energy that needs to be removed. Like all things in the industry it becomes a giant balancing act. Ultimately, outside of dedicated appliances, and amount of stacked memory would be treated as a L4 cache / buffer between the CPU and the giant pool of system memory on nearby DIMMs.

 

[gunrave]

@gunrave: amd eventually reduces cpu/apu prices which is why kaveri apus are as cheap as, if not cheaper and better than haswell/refresh core i3 cpus.

Thx for reply,
Yeah the 7850k is definitely Haswell refresh core i3 killer, but in the beginning the price are really not make sense, unless AMD know the their high end kaveri are outperform Intel counterpart and adjust the price to add the margin.

The last price-cut is, i think, the most attracting for kaveri to sale.

When the price get in to my country, i will buy it..haha

When did i3 became able to play anything at higher than 15 FPS with the IGP? Oh yeah.
A10-7700 and A10-7850K IGP give you 30 FPS at worst and 45~50~55~60 FPS on quite modern games on medium/high 720p... Some people don't need a dGPU, I think the price is justified.

Agree with you,
I think i3 will never have higher FPS than that before the use of dGPU.
But such a shame that attractive price cut of 7850K did not spread as fast as the news.

Let's hope AMD will punch back..

 

[noob2222]

Apus sacrifice cpu for the igp, look at the performance deficit between a10 6800k and the 4350 fx much less the 8350.

Im seeing stacked memory bringing its own set of problems, primarily being cost and longevity of the parts from the factory and being passed down as a $1000+ solution. Overclocking wont be viable and making custom designs will go away. Oems will love the latter.

 

[szatkus]

so you guys are back on the 2020: GPU extinction topic again, hilarious. i just bought a gtx970 2 days ago, and i doubt any APU can match this card in 6 years time.

it's not like GPUs will stay @ GDDR5 forever even, so can anyone explain how stacked DDR4 APUS will catch up against stacked (volta onwards) GDDR6 (or 7)?

Maybe with a die shrink they could fit 4-6 sr/ex cores onto a 290, a bit of hbm and bam your 970 is matched within a year?

Anyone see the dragon age inquisition techspot review/benches?

7970 ghz does better than a gtx 770, and r9 290 matches 780ti, while gtx 980 is 3 fps over 290x.

Also, all dual core processors, from any manufacturer, were unable to play the game. They experienced a locked system with 100% core load on both cores at launch screen.

But you still just need a dual core right? Also FX8350 matches 4770k from 3.5 to 4.5 ghz. Game is clearly gpu bound with 290x and 980 only hitting 56 and 59 fps @ 1200p on ultra.

Reminds me Alan Wake. When I disabled two cores game ran at ~30FPS, when I left only 1 active core frame rate dropped to less than 5FPS.

Looks like DAI works fine even on FX-4100, I don't think that locking it on 4 threads is ok.

 

[wh3resmycar]

alright, i think i've said this a year or more ago, wake me up when an apu can do 60fps with contemporary games (max settings). i have a hunch portable nuclear reactors will come first before an apu replaces a dGPU.

 

[gamerk316]

Anyone see the dragon age inquisition techspot review/benches?

7970 ghz does better than a gtx 770, and r9 290 matches 780ti, while gtx 980 is 3 fps over 290x.

Also, all dual core processors, from any manufacturer, were unable to play the game. They experienced a locked system with 100% core load on both cores at launch screen.

But you still just need a dual core right? Also FX8350 matches 4770k from 3.5 to 4.5 ghz. Game is clearly gpu bound with 290x and 980 only hitting 56 and 59 fps @ 1200p on ultra.

GPUs are VRAM limited, which I suspected based on my 770 results on Ultra settings. I also think MSAA is KILLING performance, since I go from silky smooth on high defaults to a suffer fest with MSAA 2x enabled.

The dual core CPU results are likely due to a bug, since even PII can handle the game at >60 FPS on average. I'm suspecting that will be fixed fairly shortly.

 

[8350rocks]

Phenom II only averaged 32 fps with 980 on ultra at 1200p. What makes you think a dual core can run that?