I agree. But AMD seems to be planning a HPC APU for 2017, which could be really, really fast. Imagine 8 Zen cores + SMT and a Fiji-like GPU in 14nm, together with 16GB+ HBM. But I doubt that we will see that thing on the PC early, maybe never.
Oh you would not have HBM system memory, bad idea due to costs, but a hybrid design with 2GB of HBM plus dual channel regular DDR4 sodimms would still be very compact and cheap.
I doubt that's a huge problem in the market in which those APU will be going. Also in 2017 HBM should be cheaper than it is now. And an unified memory pool should also be better (HSA).
HBM is not for system memory. It is for graphic memory only. And it is not so expensive as you make it sound: Fury is ~$560 and it is a huge silicon die, not much more expensive of a large R9 290.
Embedded HBM is much more expensive than normal Dram. So it can (and should) be used as video-memory, to avoid the bottleneck of the bussing (even the DDR4 cannot come close to the bandwidth of GDDR5 which, in turn, does not come close to the bandwidth of HBM). So yes, of course it could, but it wouldn't make any sense.
HBM on an APU would almost certainly just use a 1024-bit bus, which would provide 128-256GB/s of bandwidth - which is astonishing. However, it would be very difficult to use it as system RAM. I'd expect a 1 or 2GB HBM capacity, with possibly some acting as an L4 cache - but not likely in the first iteration.
I really hope that you’re wrong. Yes, the interposer is expensive, and no, I really don’t think we can get an APU with HBP for ~$110, but, hopefully, the technology for Fury can scale a bit and give us an APU with 1GB of HBM, is going to push the performance up one notch without needing any architectural improvement. Just look at how the Iris Pro on the Intel’s parts: without on-chip Dram they get creamed. Add the eDram and they fly (granted that the GPU is not identical, but the “jump” is huge). The issue with Intel’s solution is that it is insanely pricey ($300+), and make little sense, so it’s key to have HBM at ~$200 price point: that should blow the entry level discrete GPU out of the market.
If Zen is good, and competitive with Intel, then AMD can start selling $300+ APUs again, using HBM2 to accelerate the integrated graphics (which should be 1024 shaders at least on 14nm).
There will be a SKU that solely uses HBM (for ultra portables), probably with 8GB on board. There may be a SKU that also has DDR4 for expansion, and that uses a portion of the HBM as a massive cache.
The interposer for an APU would be a lot smaller than the one for Fiji - Fiji is 600mm^2! They are also quite cheap compared to logic - they use an old proven process with few layers.
AMD will move to using an interposer when it releases a Zen based APU, maybe late 2016 or 2017. Their current APUs don't have HBM memory controllers on board. I do wonder what their mid-2016 plan is in the APU area.
The interposer cost single digits $. It's made on an old process (65nm for Fiji GPU) and you can omit the most costly steps that need to be done when creating a real processor.
AMD could have gone for a Quad-Channel-FM Plattform (they have quad-channel opterons), or switched to DDR4 early. They don't seem very motivated to significantly improve the APU-Performance right now, rather just looking to collect their money from people with small budgets.
Switching to DDR 4 early would have been about the stupidest thing they could have done. Right now they are competing as a budget alternative to Intel. Having to buy memory that costs twice as much, or more, would not have helped them at all.
What is the point of reducing the GPU memory bottleneck, when the CPU computing portion is lacking in the first place?
Also, quad channel memory is going to increase the cost severely and you are very likely to end up with a massive 2000pin+ socket and, by association, massive mobos as well.
HBM is just a better design that can be applied to both GPU and APU.
Is there a comparison between AMD A10xxxx iGPU performance vs Nvidia 540M laptop gpu from 2010? I still use the laptop and it would give me a good point of comparison vs AMD A10 graphics performance. Thanks!
The A10 should be roughly 4-5 times as fast as the 540M, depending on the game. The 540M is 10.8 GT/s, 2.4 GP/s, and has 96 shaders. The A10 is 23.4 GT/s, 11.5 GP/s, and has 512 shaders. They both use DDR3. The 540M has dedicated DDR3, but the A10 supports up to 2133 on the motherboard, so the memory performance should be roughly the same. The A10 supports newer features and APIs, which can also bring performance increases.
I can't get excited about APUs. Even in laptops it seems like it would be better to have the GPU separate to spread out the heat more. DDR3 for GPUs is something that needs to be put to pasture.
It really depends on the use. If you use the laptop only, or mostly for gaming, then you’re probably better off with a “real” gaming laptop, which will cost you. I do casual gaming and I bought a Lenovo Z50 with the A10-7300. Could not be happier: for $450 I can play most modern games at decent settings, and not-so-new games (Need for Speed Underground 2) maxed out. It is a steal, really. And the 4 cores make any other non-gaming tasks seamless. I am really pleased actually.
Actually, I think having a powerful graphics chip on the CPU would make a lot of sense. Look at the monstrous heatsinks that we put on CPUs even though at this point they use only a fraction of the power of a GPU. Sure, the CPU will run hotter, but the GPU would be cooled a lot more effectively. Most of the time that is a tradeoff that's worth making.
I would really like to see an actually powerful graphics APU. But I guess right now there's no point with AMD's lackluster CPU performance.
I agree with Nagorak. Yes, you do concentrate GPU and CPU heat into the same area, but cooling isn't an insurmountable technical problem at this point. For instance, a Dell Precision m4400 (15 inch professional notebook from 2009-ish) can be outfitted with a 45 watt TDP CPU and a Quadro FX 770m which can account for another 35 watts. It had a pair of copper heat pipes that ended with a small blower fan and single set of aluminum fins. While modest, this was workstation class hardware and was cooled properly for longevity. Even mobile platforms could handle the combined thermal load of an APU with HBM, I think.
And yeah, enthusiasts are sticking positively huge 6+ heatpipe sinks with dual 80+mm fans onto CPUs that have a TDP of 65 to 95 watts and then fussing about running at 55C when the Tjunction is close to 100C. I don't see it as an insurmountable technical problem to cram all those transistors into the same space given modern lithography.
So, your example for why it's a good idea to put the CPU and GPU heat together is a laptop that separated the CPU and GPU?
As for the fact that large CPU coolers are being used... looking more carefully one would notice that, in order to have quiet performance, one needs all of the CPU cooling available just for the CPU. Toss in GPU power consumption and you're going to run into trouble.
CCL in the UK has it listed for pre-order at £100, however considering the 7600 is £66 and the 7650 is £75 on the same site, and the 7870K can be found for £100 (+/- £5) at various other sites, it's not exactly a good deal.
I love how the A8-7670K is faster and better than the A10-7700K in every regard. They really screwed up the naming consistency with that SKU. The A level should always be used to differentiate the number of funtional units like when the APUs were first launched.
We’ve updated our terms. By continuing to use the site and/or by logging into your account, you agree to the Site’s updated Terms of Use and Privacy Policy.
Facebook
Twitter
RSS
33 Comments
Back to Article
IlllI - Monday, July 20, 2015 - link
do you think we'll ever see one of these with hbm on the package? i wonder what kind of improvement it would make to the on-die gfx.BillyONeal - Monday, July 20, 2015 - link
Certainly not at this price point. The interposer necessary for HBM is expensive to produce because it has an extremely large die size.Dobson123 - Monday, July 20, 2015 - link
I agree. But AMD seems to be planning a HPC APU for 2017, which could be really, really fast. Imagine 8 Zen cores + SMT and a Fiji-like GPU in 14nm, together with 16GB+ HBM. But I doubt that we will see that thing on the PC early, maybe never.Meaker10 - Monday, July 20, 2015 - link
Oh you would not have HBM system memory, bad idea due to costs, but a hybrid design with 2GB of HBM plus dual channel regular DDR4 sodimms would still be very compact and cheap.Dobson123 - Monday, July 20, 2015 - link
I doubt that's a huge problem in the market in which those APU will be going. Also in 2017 HBM should be cheaper than it is now. And an unified memory pool should also be better (HSA).yankeeDDL - Tuesday, July 21, 2015 - link
HBM is not for system memory. It is for graphic memory only. And it is not so expensive as you make it sound: Fury is ~$560 and it is a huge silicon die, not much more expensive of a large R9 290.Dobson123 - Tuesday, July 21, 2015 - link
Of course HBM can be used as system memory, there's no reason why that shouldn't be possible.yankeeDDL - Tuesday, July 21, 2015 - link
Embedded HBM is much more expensive than normal Dram. So it can (and should) be used as video-memory, to avoid the bottleneck of the bussing (even the DDR4 cannot come close to the bandwidth of GDDR5 which, in turn, does not come close to the bandwidth of HBM).So yes, of course it could, but it wouldn't make any sense.
Dobson123 - Tuesday, July 21, 2015 - link
It would in case of a HPC APU.menting - Tuesday, July 21, 2015 - link
If it's soldered down on the motherboard along with a soldered down CPU maybe. The short reach of the pin drivers make it unfeasible otherwise.looncraz - Tuesday, July 21, 2015 - link
HBM on an APU would almost certainly just use a 1024-bit bus, which would provide 128-256GB/s of bandwidth - which is astonishing. However, it would be very difficult to use it as system RAM. I'd expect a 1 or 2GB HBM capacity, with possibly some acting as an L4 cache - but not likely in the first iteration.yankeeDDL - Tuesday, July 21, 2015 - link
I really hope that you’re wrong. Yes, the interposer is expensive, and no, I really don’t think we can get an APU with HBP for ~$110, but, hopefully, the technology for Fury can scale a bit and give us an APU with 1GB of HBM, is going to push the performance up one notch without needing any architectural improvement.Just look at how the Iris Pro on the Intel’s parts: without on-chip Dram they get creamed. Add the eDram and they fly (granted that the GPU is not identical, but the “jump” is huge).
The issue with Intel’s solution is that it is insanely pricey ($300+), and make little sense, so it’s key to have HBM at ~$200 price point: that should blow the entry level discrete GPU out of the market.
psychobriggsy - Tuesday, July 21, 2015 - link
If Zen is good, and competitive with Intel, then AMD can start selling $300+ APUs again, using HBM2 to accelerate the integrated graphics (which should be 1024 shaders at least on 14nm).There will be a SKU that solely uses HBM (for ultra portables), probably with 8GB on board. There may be a SKU that also has DDR4 for expansion, and that uses a portion of the HBM as a massive cache.
psychobriggsy - Tuesday, July 21, 2015 - link
The interposer for an APU would be a lot smaller than the one for Fiji - Fiji is 600mm^2! They are also quite cheap compared to logic - they use an old proven process with few layers.AMD will move to using an interposer when it releases a Zen based APU, maybe late 2016 or 2017. Their current APUs don't have HBM memory controllers on board. I do wonder what their mid-2016 plan is in the APU area.
beginner99 - Tuesday, July 21, 2015 - link
The interposer cost single digits $. It's made on an old process (65nm for Fiji GPU) and you can omit the most costly steps that need to be done when creating a real processor.Shadow7037932 - Monday, July 20, 2015 - link
Not any time soon. Probably a generation or so out before the midrange and up GPU lineup gets the HBM treatment, followed by the APUs. However, it'szodiacfml - Tuesday, July 21, 2015 - link
Too expensive. Imagine, it is more expensive than GDDR5. Maybe we will see HBM on APUs once their GPUs uses HBM2.ShieTar - Tuesday, July 21, 2015 - link
AMD could have gone for a Quad-Channel-FM Plattform (they have quad-channel opterons), or switched to DDR4 early. They don't seem very motivated to significantly improve the APU-Performance right now, rather just looking to collect their money from people with small budgets.Nagorak - Tuesday, July 21, 2015 - link
Switching to DDR 4 early would have been about the stupidest thing they could have done. Right now they are competing as a budget alternative to Intel. Having to buy memory that costs twice as much, or more, would not have helped them at all.meacupla - Monday, August 3, 2015 - link
What is the point of reducing the GPU memory bottleneck, when the CPU computing portion is lacking in the first place?Also, quad channel memory is going to increase the cost severely and you are very likely to end up with a massive 2000pin+ socket and, by association, massive mobos as well.
HBM is just a better design that can be applied to both GPU and APU.
txjr88 - Monday, July 20, 2015 - link
Is there a comparison between AMD A10xxxx iGPU performance vs Nvidia 540M laptop gpu from 2010? I still use the laptop and it would give me a good point of comparison vs AMD A10 graphics performance. Thanks!barleyguy - Tuesday, July 21, 2015 - link
I looked at the comparison charts on TechArp.The A10 should be roughly 4-5 times as fast as the 540M, depending on the game. The 540M is 10.8 GT/s, 2.4 GP/s, and has 96 shaders. The A10 is 23.4 GT/s, 11.5 GP/s, and has 512 shaders. They both use DDR3. The 540M has dedicated DDR3, but the A10 supports up to 2133 on the motherboard, so the memory performance should be roughly the same. The A10 supports newer features and APIs, which can also bring performance increases.
txjr88 - Tuesday, July 21, 2015 - link
Thank you very much!Oxford Guy - Tuesday, July 21, 2015 - link
I can't get excited about APUs. Even in laptops it seems like it would be better to have the GPU separate to spread out the heat more. DDR3 for GPUs is something that needs to be put to pasture.yankeeDDL - Tuesday, July 21, 2015 - link
It really depends on the use.If you use the laptop only, or mostly for gaming, then you’re probably better off with a “real” gaming laptop, which will cost you.
I do casual gaming and I bought a Lenovo Z50 with the A10-7300.
Could not be happier: for $450 I can play most modern games at decent settings, and not-so-new games (Need for Speed Underground 2) maxed out. It is a steal, really.
And the 4 cores make any other non-gaming tasks seamless. I am really pleased actually.
Nagorak - Tuesday, July 21, 2015 - link
Actually, I think having a powerful graphics chip on the CPU would make a lot of sense. Look at the monstrous heatsinks that we put on CPUs even though at this point they use only a fraction of the power of a GPU. Sure, the CPU will run hotter, but the GPU would be cooled a lot more effectively. Most of the time that is a tradeoff that's worth making.I would really like to see an actually powerful graphics APU. But I guess right now there's no point with AMD's lackluster CPU performance.
BrokenCrayons - Tuesday, July 21, 2015 - link
I agree with Nagorak. Yes, you do concentrate GPU and CPU heat into the same area, but cooling isn't an insurmountable technical problem at this point. For instance, a Dell Precision m4400 (15 inch professional notebook from 2009-ish) can be outfitted with a 45 watt TDP CPU and a Quadro FX 770m which can account for another 35 watts. It had a pair of copper heat pipes that ended with a small blower fan and single set of aluminum fins. While modest, this was workstation class hardware and was cooled properly for longevity. Even mobile platforms could handle the combined thermal load of an APU with HBM, I think.And yeah, enthusiasts are sticking positively huge 6+ heatpipe sinks with dual 80+mm fans onto CPUs that have a TDP of 65 to 95 watts and then fussing about running at 55C when the Tjunction is close to 100C. I don't see it as an insurmountable technical problem to cram all those transistors into the same space given modern lithography.
Oxford Guy - Thursday, July 23, 2015 - link
So, your example for why it's a good idea to put the CPU and GPU heat together is a laptop that separated the CPU and GPU?As for the fact that large CPU coolers are being used... looking more carefully one would notice that, in order to have quiet performance, one needs all of the CPU cooling available just for the CPU. Toss in GPU power consumption and you're going to run into trouble.
meacupla - Tuesday, July 21, 2015 - link
WHAT? AMD actually having products in store on launch?It's the end of the world now.
silverblue - Tuesday, July 21, 2015 - link
CCL in the UK has it listed for pre-order at £100, however considering the 7600 is £66 and the 7650 is £75 on the same site, and the 7870K can be found for £100 (+/- £5) at various other sites, it's not exactly a good deal.[email protected] - Tuesday, July 21, 2015 - link
Umm does amd have cpu with ddr 4 on die for the gpu. I don't think a small frequencies change will really help much.Assimilator87 - Wednesday, July 22, 2015 - link
I love how the A8-7670K is faster and better than the A10-7700K in every regard. They really screwed up the naming consistency with that SKU. The A level should always be used to differentiate the number of funtional units like when the APUs were first launched.Suckacok - Wednesday, July 29, 2015 - link
SO DAMN WEAK!!! AMD IS CRAP SO SLOW COMPARED TO INTEL