If you read my statement I'm not the one declaring the imminent death of dGPU... However the iGPU is going to extend further and further up the performance scale in relation to dGPU moving forward. The point is that dGPU tech is going to become more and more niche. That *isn't* the same thing as saying it will die out. One thing I do predict, a combined APU *will become more efficient than separate components* once the correct balance of CPU and GPU resources is reached. In that context I think APU type chips will replace the 'dGPU as accelerator' cards for HPC workloads. Gaming / graphics rendering isn't really the same thing, as it is so specific that having a single ASIC with 100% transistors dedicated to that job is still going to be the better option overall.
The other thing to remember, modern GPU's are built in blocks. Having a large iGPU doesn't require you to turn it all on at once- each block will be individually gated and only used when it's needed (this is all current tech). The new consoles have already shown it's quite feasible to put a mid range dGPU on die with the CPU.
oh, you mean juan? he's parroting some s/a articles from before. and i've verified several times in this thread how he lies. his claims(which are actually blanket statements devoid of any specifics) are of no concern to me. this is why the reply was directed towards you.
as for dgfx - high end discreet gfx has always been niche. this is not something new. with the shrinking pc market, discreet parts are the things people stop buying if they think they have "good enough" performance from integrated parts. but they end up buying high performance discreet gfx when they need more than "good enough". it's not a technological concern, more of budget concern. in case of apus, the bottlenecks hit early due to it's integrated nature. right now for example, the igpu is faster than the memory it uses. so now the memory itself and it's usage will have to catch up. when the memory catches up, then the igpu will try to increase better utilization of faster memory. and so on.
an apu (e.g. 7850k) is efficient as any other device, for it's purpose and for it's price. it depends on the particular usage case.
in case of replacing multigpu accelerators... i don't know much, but i've always thought of maintaining memory coherency across processing nodes a major concern. with cpus you get 4-16 cores per socket, may be 64 max. per cluster with 4S. but with apu with 4+12cu or in the imaginary high perf case, e.g. an soc with 32-64+1024-2048CU, it may become a major problem. but i don't know much about hpc, i am speaking from basic p.o.v. additionally, in hpc you have the chip running on load full time, put that in perspective with my previous post.
actually, in load situations, you are using all of the cores in an igpu. gating only helps in lesser loads and when the power management logic in smart enough to determine the scenarios. the new consoles' socs are asics, not general purpose parts. and they have the huge benefit of close-to-metal programming - massively reducing software overheads. if you read trinity, richland and kaveri reviews where the stock cooler was used, you'll notice the mention of throttling on load. current apus have built in power virus detecting so that they throttle to protect the chip from thermal trip. so, it's happening already with current chips.
You are being a bit too excessive saying Haswell mops the floor vs 920....it is/does but in reality its not justification enough of a value/performance increase to go up to it for most purposes/application's in my opinion...
I think you underestimate like most how First Gen can still really perform well/do considering its age/dated...
You are being a bit too excessive saying Haswell mops the floor vs 920....it is/does but in reality its not justification enough of a value/performance increase to go up to it for most purposes/application's in my opinion...
I think you underestimate like most how First Gen can still really perform well/do considering its age/dated...
True, the first gen i7 was a really good chip. I must admit I still use an AMD Phenom II X6, another seriously underrated processor imo.
If we look back at processor design the amount of external components now on die: The FPU (maths co-processor for anyone as old as me), level 2 cache, level 3 cache (super socket 7 + K6-III) PCI controllers, USB controllers, PCIe controllers, the entire north and south bridges.... Why is graphics and main memory so sacred that they cannot ever be integrated? Eventually the number of transistors available on die reaches such a point that including these things becomes inconsequential.
And why were all those components eventually moved onto the CPU? Because of die shrinks, you got both the necessary transistor budget to fit it, and because of the inherent power savings. But as has been noted many times: We're nearing the limit of how much more you can physically shrink the CPU down. No more shrinkage, no more growth in transistor budget, no more space to put more things on the CPU die [without sacrificing yields and driving up costs, at least].
Already, cache memory is VERY space/cost inefficient compared to standalone DRAM, in that it draws more power and requires a larger transistor budget then a stand alone DRAM module does. And that's just for maybe a hundred MB. You simply do not have the space to slap down GB of the stuff.
What's going to happen, long term, is what you're seeing now, where most OEM level CPUs will have a capable GPU built in, that is capable of playing back high-def videos and playing whatever the newest gaming fad of the month is [Sims, Warcraft, etc]. But you're never going to see a built in GPU beat a dedicated GPU, simply due to the differences in transistor/power budgets available.
If we look back at processor design the amount of external components now on die: The FPU (maths co-processor for anyone as old as me), level 2 cache, level 3 cache (super socket 7 + K6-III) PCI controllers, USB controllers, PCIe controllers, the entire north and south bridges.... Why is graphics and main memory so sacred that they cannot ever be integrated? Eventually the number of transistors available on die reaches such a point that including these things becomes inconsequential.
And why were all those components eventually moved onto the CPU? Because of die shrinks, you got both the necessary transistor budget to fit it, and because of the inherent power savings. But as has been noted many times: We're nearing the limit of how much more you can physically shrink the CPU down. No more shrinkage, no more growth in transistor budget, no more space to put more things on the CPU die [without sacrificing yields and driving up costs, at least].
Already, cache memory is VERY space/cost inefficient compared to standalone DRAM, in that it draws more power and requires a larger transistor budget then a stand alone DRAM module does. And that's just for maybe a hundred MB. You simply do not have the space to slap down GB of the stuff.
What's going to happen, long term, is what you're seeing now, where most OEM level CPUs will have a capable GPU built in, that is capable of playing back high-def videos and playing whatever the newest gaming fad of the month is [Sims, Warcraft, etc]. But you're never going to see a built in GPU beat a dedicated GPU, simply due to the differences in transistor/power budgets available.
Well, that all depends on if we really are at the limit of die shrinks or not. We've got at least another 2 or 3 nodes yet, and people have been (and keep) predicting that no further shrinks are possible and "it's against the laws of physics" and such until the next node appears. I think if anything major process nodes are lasting longer these days, so the pace at which this changes (for everything) is going to be slower. We maybe even need an entirely different substrate other than silicon for real improvements (graphine has been shown to be workable for ICs for example).
I agree there isn't much more AMD can do at 28nm, my argument is based on allot of if's and does assume continued scaling one way or another. I do think though that the iGPUs are getting to the stage to be "good enough" for pretty much everything- albeit at lower resolutions and detail settings, which can only be a good thing. OEMs cutting corners on PCs for so long has been one factor that has really hurt PC gaming imo- the number of expensive machines sold to the public with 0 graphics capability was absurd. Thankfully by including a reasonable iGPU on every processor Intel and AMD are basically putting a stop to that.
ah. right now, materials are the biggest unknowns for me. if chipmakers find a way to use new materials to reduce power and heat while improving performance or keeping the same performance level, it'd be good.
Games yes but using the gpu for software acceleration is not as effective as an apu, heres a toms article from 2012, the q&as are interesting, think this was posted before but ot and worth a reread:
http://www.tomshardware.co.uk/photoshop-cs6-gimp-aftershot-pro,review-32461-13.html
I find it interesting NVIDA chips weren't compared, given their emphasis on gpGPU. Just saying, there's no real HW difference between a GPU/APU that would greatly affect performance. At the end of the day, they're executing the same functions on the same HW.
Yeah exactly Phenom II X6 was/is really underrated.
Yeah i sold my 1100T for 100$ got the money in 3 days and i sold it on craigslist and got a 8350FX which to be honest is faster in general apps(handbrake is maybe 10%) but not by a lot it is however a lot faster in newer games by a decent amount at least 20%. Its worth noting that i was able to get my 1100T at 3.9Ghz on lower than stock voltage where this darn 8350fx can't get above 4.3Ghz without excessive voltage and heat, i guess its ok i usally get lucky with Amd my first CPU i bought from them was a Athlon II x4 and i got it OC to 3.2Ghz from the standard 2.6GHz on stock voltage and that was with the stock heat-sink even. Anyways you are right your 920 was probably the legend of all CPU's for the last 6 years its still at the top of the charts for gaming you can not say that for any Amd CPU back then for gaming maybe for general apps but not gaming.(and to think they only asked 320$ for it 1100T was only 35$ cheaper i think)
Games yes but using the gpu for software acceleration is not as effective as an apu, heres a toms article from 2012, the q&as are interesting, think this was posted before but ot and worth a reread:
http://www.tomshardware.co.uk/photoshop-cs6-gimp-aftershot-pro,review-32461-13.html
I find it interesting NVIDA chips weren't compared, given their emphasis on gpGPU. Just saying, there's no real HW difference between a GPU/APU that would greatly affect performance. At the end of the day, they're executing the same functions on the same HW.
What did you think of the q&a with the adobe dev?
Im sure I seen other benches in this thread where an apu destroyed a 780+cpu in a gpu accelerated task purely because there was no transfer over pcie, I think discrete only works if the task is big enough to warrant copying across the pci, the sandwich analogy is a good one.
Games yes but using the gpu for software acceleration is not as effective as an apu, heres a toms article from 2012, the q&as are interesting, think this was posted before but ot and worth a reread:
http://www.tomshardware.co.uk/photoshop-cs6-gimp-aftershot-pro,review-32461-13.html
I find it interesting NVIDA chips weren't compared, given their emphasis on gpGPU. Just saying, there's no real HW difference between a GPU/APU that would greatly affect performance. At the end of the day, they're executing the same functions on the same HW.
What did you think of the q&a with the adobe dev?
Im sure I seen other benches in this thread where an apu destroyed a 780+cpu in a gpu accelerated task purely because there was no transfer over pcie, I think discrete only works if the task is big enough to warrant copying across the pci, the sandwich analogy is a good one.
The issue is to utilize a dGPU properly for this type of task, you need to send large chunks of data over the bus a handful of times, rather then a little data over the bus all the time. The latency kills you if you take the second approach. Simply send the data you need processed ONCE, and the dGPU would pull ahead again. The APU wins these cases due to sub-optimal coding. Of course, this would result in reducing the APU performance due to the system RAM bottleneck. So now we have a situation where we can't optimize for both pieces of HW at the same time; one suffers, the other gains.
Yeah the 8350 is no slouch at all either...yes heat is big issue....My 920 I'm limited I can get to about 3.6-3.8ghz highest I can do with stock cooler, they are pretty wimpy....I was wanting to get the CM hyper 212 EVO for but its too tall/wide to fit in my Mid case....kinda sucks because I like this case and want to use it haha, and the 212 probably really is like one of the best deals/performance for that price....
Guess I might as well get a water unit, probally better anyways... one of the Corsair units look awful tempting to get...And would be really nice getting up to/seeing 4ghz+ hell even at 3.6-3.8ghz I'm getting into 70-75c load temps....so yeah better cooler would do wonders...
Unless there is some other decent air unit that will fit and be waaay better than stock heatsink....
AMD will produce another dCPU for HEDT, it is already being engineered, and will be part of the next uarch. So clearly they do not think everything is going to APUs.
Second, when the article about GPU app acceleration was written, openCL had not seen much interest outside a few niches, or as a marketing ploy. Since then, HSA has stirred massive interest and can be done over PCIE bus.
Third, APUs will always be a compromised compute solution in the consumer sector. They will either be serial compute heavy, parallel compute heavy, or a jack of all trades master of none solution. In some scenarios, you could build specialized APUs, a la the consoles. However, these would be niche solutions to niche markets with a specific function in mind. It honestly takes two 290X cards or one 295X2 card to run a 4k screen reliably with good fps. That is 2 pieces of silicon at ~7 bil transistors each. Now, even accouting for die shrinks and new substrates, at 40% improvement generation over generation (asking a lot by the way) in 2 generations you still need an iGPU with 6 bil transistors. Let us make more assumptions and assume by 2020 density doubles over kaveri. That will be about 2 bil transistors (not likely but hey...), so by then we will have had ~4 generations of GPUs. Which would put us needing a transistor budget of 2.2 bil transistors on an iGPU to do 4k resolution only as well as a 6 year old discrete card.
Get the point yet? Now, ask yourself, if ~2.2 bil transistors can power 4k in 2020, what can a 6 bil transistor dGPU do with the same tech?
APUs as anything more than a mainstream solution are farther out than 2020.
Games yes but using the gpu for software acceleration is not as effective as an apu, heres a toms article from 2012, the q&as are interesting, think this was posted before but ot and worth a reread:
http://www.tomshardware.co.uk/photoshop-cs6-gimp-aftershot-pro,review-32461-13.html
The APU only won 1 out of 5 of those benchmarks. The 8150 w/dGPU won the rest.
Games yes but using the gpu for software acceleration is not as effective as an apu, heres a toms article from 2012, the q&as are interesting, think this was posted before but ot and worth a reread:
http://www.tomshardware.co.uk/photoshop-cs6-gimp-aftershot-pro,review-32461-13.html
I find it interesting NVIDA chips weren't compared, given their emphasis on gpGPU. Just saying, there's no real HW difference between a GPU/APU that would greatly affect performance. At the end of the day, they're executing the same functions on the same HW.
What did you think of the q&a with the adobe dev?
Im sure I seen other benches in this thread where an apu destroyed a 780+cpu in a gpu accelerated task purely because there was no transfer over pcie, I think discrete only works if the task is big enough to warrant copying across the pci, the sandwich analogy is a good one.
The issue is to utilize a dGPU properly for this type of task, you need to send large chunks of data over the bus a handful of times, rather then a little data over the bus all the time. The latency kills you if you take the second approach. Simply send the data you need processed ONCE, and the dGPU would pull ahead again. The APU wins these cases due to sub-optimal coding. Of course, this would result in reducing the APU performance due to the system RAM bottleneck. So now we have a situation where we can't optimize for both pieces of HW at the same time; one suffers, the other gains.
I can't believe I've been agreeing with you this much. People who are making an issue out of PCIe or other bus latency for GPGPU are seriously deluded about what types of tasks you're doing on the GPU.
I don't know how many times I have to explain it. If you're doing GPGPU workloads, you're doing something that will take minutes or hours or even days and weeks. It's not something where even a second of latency is relevant.
And you are absolutely right. You can program it such that you reduce latency. Think of Open Office spreadsheet acceleration. If you do GPGPU operations on a bunch of cells and you have latency between GPU and CPU and memory, you aren't going to update each cell at a time and send them between CPU and GPU and memory. You're going to ship the whole thing to GPU, use VRAM as a cache, work on it, then update main memory as the operation completes. So while operation is still finishing, you do the latency intensive operation of updating main memory while it's still computing to hide the latency under the fact it is still computing.
I just came up with that type of solution and I'm just a lowly computer science guy with a bachelor's degree. There's no way that's the best way, but it would work.
The whining about latency with dGPU reminds me of when people were whining about dual cores and how we'll never use more than one core, it'll be bad to sync data between two cores, etc.
The problem is that some of you look at HSA and the benefits and you decide that the only way to do GPGPU in the future will be with HSA and everything sharing memory 100% of the time. Instead you will see GPU used as traditional GPGPU and HSA with shared memory used as needed. They are two different tools.
Get the point yet? Now, ask yourself, if ~2.2 bil transistors can power 4k in 2020, what can a 6 bil transistor dGPU do with the same tech?
You'll want to rethink the math there. GPUs have only gotten faster by throwing MORE transistors on them. Sure they can do some tweaks with optimizations but largely its by adding transistors.
The problem now is the transistor cost curve is going sideways and slightly up. When GPUs went from 40nm to 28nm the transistors got cheaper and we got faster graphics cards for nearly the same price. Now they're not. With some clever packaging like 2.5D they can continue making larger parts but we'll be paying more for them.
Like I've mentioned before you can buy a 20 Billion transistor part from Xilinx but it will cost you 2 arms and 2 legs, making the GTX Titan Z look like child's play. The GPU makers can learn a lot from Xilinx.
Why is graphics and main memory so sacred that they cannot ever be integrated? Eventually the number of transistors available on die reaches such a point that including these things becomes inconsequential.
I think that the idea of losing one's ability to mix and match these major components separately scares most enthusiasts. They'll make excuses as to the relevancy of dedicated graphics down to the very end.
I'm not sure I buy the line of reasoning that APUs and HEDT are somehow mutually exclusive.
So i7-4770K (it's also APU) isn't HEDT?
Exactly my point.
I don't believe that AMD will specifically target a niche market (HEDT) except as a by product of one of their main markets, exactly as in the case with Intel. I also believe that to devote resource to a non-APU line would be a step back for them and a dangerous division of resources.
I'm not sure I buy the line of reasoning that APUs and HEDT are somehow mutually exclusive.
So i7-4770K (it's also APU) isn't HEDT?
Exactly my point.
I don't believe that AMD will specifically target a niche market (HEDT) except as a by product of one of their main markets, exactly as in the case with Intel. I also believe that to devote resource to a non-APU line would be a step back for them and a dangerous division of resources.
It's a matter of perceived performance and how much brute force your CPU/GPU combo can muster.
In other words, if you take a look at games only, you could say the i7 4770K (or even the old 2700K) is HEDT (or whatever you wanna call it) since the performance is on par or even better than the i7 3970X which is a cheaper Xeon (server class and what not) for consumer market, but we all know the *real* heavy hitter will be the i7 3970X for all intended "heavy" CPU work. For a lot of "pro" workloads, the i7 4770K is a kids tool.
Point is, HEDT will be defined by how much grunt work the APU/CPU in combination with the GPU (if it needs one) can muster. I am perfectly fine with an APU being "HEDT" as long as it delivers within my expectations of being a tough grunt, haha. AMD's HSA is part of that, but I haven't seen much as of late. Hope they're still hard at work
8350rocks, the dGPU will have less than 10% more of transistors, but less performance due to interconnect bottlenecks. What are you arguing here?
Please. dGPUs have a while to grow yet, especially with at least 2-3 more node shrinks to follow. ((That being said, EVERYONE is going to stall at ~6nm or so.))
And again: Interconnects are not a bottleneck. They're actually an advantage, since you can just stuff lots of faster-then-system-RAM directly on the GPU die to create a very large high speed cache. The only downside to the interconnect is the latency transferring the data over into the cache, and even that is minimal. Meanwhile, APUs are limited by the CPU IMC and slow system RAM.
The way of thinking is quite archaic, anything you can do with memory on a gpu you can do on an apu. Interconnects are pretty much always bottleneck, which is why everything is being integrated on die. Anyways, by the time we have stuff like this, the whole chip would be different. We would be able to stack high speed ram on die by then I would hope.
No.
For the same reason that dGPU's will never go away, transistor budget. Whatever you can fit onto or near a CPU, you can cheaply fit 16~32x the amount on a DIMM or other moulder ram technology. You put 1GB of memory onto the CPU, you can fit 16~32GB of memory on a system buss. You put 8GB of memory on /near the CPU using super sayen technology, then you can fit 128~256GB of it on a local system bus. Dedicated peripheral cards also can fit even faster memory, typically 8~16x of whatever you can fit onto or directly attached to the CPU. So if you put that 1GB of shared memory on die, then you can put 8~16GB of high speed dedicated memory on the GPU. This quickly becomes a game of "whatever you can do, I can do better!" with the discrete solution always being at least an order of magnitude ahead of the integrated solution.
What you will see is that integrated solutions will cannibalize the low end markets, things like the 820/830 are already nearly obsolete and the 840 is coming very close. Middle tier holds it's own because it has access to more expensive high speed dedicated memory while integrated solutions must deal with lower cost shared system bus's. Discrete sound cards died off because there is only so much you can do to create a 44100hz two channel audio signal going to a set of trashy walmart speakers. As others have pointed out, those who actually care about their audio quality go with more exotic solutions like external DSP's. Hell my entire living room is powered by a Yamaha RX-V3800 which was something like $1600 or $1700 when I first purchased it. It does all my signals processing via HDMI or SPDIF. Side note, I think Sony was seriously f*cked in the head when they abandoned the SPDIF interface for HDMI, many of us want to send uncompressed multi-channel audio to our DSP's from the sound sources.
So if people are willing to spend serious change on high end audio equipment for a more fuller experience, don't you think they would also spend equally if not more money on high end graphics equipment?
Thats like saying off die caches will never go away. One day you will be able to do TBs of memory close to the die. It doesn't matter. Sure you could probably get 128MB of cpu cache if you made a die just for it but nobody does that any more because it is not efficient. Who cares what you can do right now, if this is a discussion about technology in 10 years then things evolve. What do we need faster cpus anyways, the current ones are plenty fast. Who know what future APUs will be, maybe we put 90% of the transistor budget to the parallel compute engines and 10% to serial cores, what would the difference be then between a dedicated card and an APU? There really isn't any limit to what the future holds.
Anyways I would never trust anyone who would call themself an audiophile because audiophiles buy dumb stuff all the time, they buy $1000 cables just to hear placebo effects. Sure the built in soundcard isn't as good as a dedicated one and those aren't as good as a external amp but its still all the same, people don't need more and the products are a niche now.
Umm you do realize we still have off die cache's ... we call it system memory. After that is slow long term storage, HDD's. We have since created a memory hierarchy that is as follows.
The fastest memory possible is the L1 cache which is only one step away from physical CPU registers, it's also the most expensive in space utilization. After that is L2 and L3 cache. We then move off the CPU die and can utilize different transistor technology for modular memory which is slower but far bigger. Once we're done with memory we move onto the non-volatile storage mechanisms which have the greatest storage density.
Just to illustrate how obscene your comment is I will use a common power setup.
i7-4790K
Total L1 cache: 64KB per core: 256KB total per chip
Total L2 cache: 256KB per core: 1MB total per chip
Total L3 cache: 8MB
Notice the insane leap of sizes from 8MB of L3 cache to 16GB of system memory and 3~8GB of dGPU memory. So even assuming you could get 256MB of on die memory using today's technology, your still 64 times lower then what the user had in system memory and 12~16 times lower then that same power use's graphics memory.
Congrats you can now load a small linux kernel with nothing else happening. Once you put 1GB of on-die super memory, your now talking a system with 64GB of memory and 12~16GB of fast graphics memory. This would all be with the same technology. You can never catch up.
Using your exaggerated example of DNA memory. 1TB of magical on-die memory would have 64TB of system memory at 12~16TB of fast graphics memory. So while "ONE TERABYTE" sounds big in today's terms, it's awfully small when contrasted to what else will be available.
They used to have both off die cache and ram, now that is much more simplified by having sufficient on die cache and fast enough ram. Memory hierarchy will evolve. There is a reason why ram is dynamic and that we don't load all 50 GB of a game into ram when you play it. Eventually your ram would be large enough and all the software would be smart enough that you would just be streaming in things from the HDD in real time for a personal computer no matter what the software requirements would be. There will be a point where you won't need more ram what so ever.
Now we stack that much ram next to the CPU and what would we need ram on the motherboard for? Mobile SoCs will be coming out in the next year or so without ram in their pcb. Sooner or later its going to happen to laptops and eventually desktops. Its simply going to become less and less useful to have the extra power you get from a desktop PC.
What we have today in desktop PC will go the way of the mainframe to make way to integrated solutions for most people in due time. Hardware is so far ahead of software at this point there isn't even a reason to upgrade for most people running i7-920s for a long time. When you integrate that into a soc with a decent pool of ram?
The moving target doesn't work when software is stagnant for the things the general public would use. Who needs a 400HP muscle car in a city when a hybrid is efficient and will get their job done? Sure there will be people who buy the most powerful stuff but that will become a smaller and smaller niche.
Facebook
Twitter
Instagram