[Mynavi.jp] AMD reveals plans to 2020.

[RussianSensation]

I have no problem with AMD's dgpus. In fact, I am running an AMD dgpu now. It is their cpu/APU lineup that has no appeal at all to me. AMD had better get on board with efficiency in dgpus though, or they are going to have it come back to bite them just like it has in the cpu space already. If the 300 lineup is one super high power expensive HBM flagship and the rest is a bunch of rebrands (for the second time), they are in serious trouble in the dgpu market as well. As for the "next chip syndrome", we have been hearing for years from AMD and now from Intel how great the next generations of igps will be, and how HSA will dominate, but we still are far from a powerful solution that rivals even low end cards like HD7750 in desktop and GT 730 or so in mobile, and HSA is still a very niche market at best.

Gotcha. You are right and I agree with you that their APUs have been way underperforming on the CPU/GPU side and too expensive. However, I think with access to 14nm node + new CPU architecture with Zen and HBM 1/2, it's possible there will be a huge revolution in AMD's APU performance. Right now AMD is stuck between a rock and a hard place:

1) Their partner's node manufacturing is way behind Intel's. That means even if AMD's CPU was exactly as efficient in IPC and as fast as Intel's chips, they would always lose in perf/watt simply because of the node.

2) AMD can't really increase GPU speed 100-200% because DDR3 and even DDR4 is way too slow. These memory technologies aren't good enough for R9 280X/290 level of APU graphics. For that reason, HBM/HBC is required.

The current GCN and CPU architectures in their APUs are outdated. On the dGPU side, AMD can use AIO CLC and enlarge the die size to get improvements in performance. For APUs, there isn't enough memory bandwidth so it would be pointless right now. AMD basically mistimed the ATI acquisition about 10 years too early. The right time to go after high-performance APUs is when HBM provides 500GB/sec+ of memory bandwidth.

We know that the lower node and advancements in GCN/post-GCN will take care of the GPU side and power usage of their APUs. The big unknown is just how good Zen CPU architecture is vs. Skylake. The gap in IPC on a per core basis compared to Haswell today is mind-boggling. The other major challenge for AMD will be pricing. If they start going with HBM and large sized APUs, I think APUs at $149-179 will be wishful thinking unless we are talking about low end performance parts.

 

[Silverforce11]

If they start going with HBM and large sized APUs, I think APUs at $149-179 will be wishful thinking unless we are talking about low end performance parts.

Ofcourse. If they want to get 300W APU, it has to have performance to back it up, thus, die size 500mm2+ is expected.

Cost can be very flexible for a premium product, to give them plenty of margins.

 

[swilli89]

I'm wondering about the people wanting a 300w APU in a sff. 300w is a lot of heat to dissipate. Isn't the ps4 like 150w total?

 

[.vodka]

Nothing a CLC AIO can't handle.

A SFF case capable of housing a 300w APU will have a 120mm exhaust where you can mount the AIO, at least.

 

[Silverforce11]

I'm wondering about the people wanting a 300w APU in a sff. 300w is a lot of heat to dissipate. Isn't the ps4 like 150w total?

You can put a i7 + Titan X into a SFF and it'll be fine. PS4 is limited by its cooling potential & the power brick. SFF doesn't have those limits.

The cool thing with 300W APU with high end GPU performance is the slim factor, with a well designed AIO 120mm rad, can get quite short & compact cases, basically as long as there's space for the radiator, it's fine.

But besides SFF, if you build a rig and the APU cost is not greater (or heck even cheaper) than a separate CPU + dGPU combo of equivalent graphics performance, it's a good deal as it simplifies the build.

XDMA Hybrid Crossfire allows upgrade-ability without making the APU useless. Currently we buy an i5 or i7 with a dGPU, that iGPU is worthless, its not doing squat for your money. With the APU + Hybrid CF, the iGPU gets put to work. It's potentially a MORE potent combo than i5/i7 + dGPU, when CF works (in almost everything except GameWorks heh).

If AMD does it right and consumers can get their hands on these MB with HBM on it.. there's no way Intel + dGPU combo is going to compete on gaming performance compared to APU + dGPU combo. Think about it. It's quite revolutionary.

 

[raghu78]

A 200-300W APU at 14-10nm FF means it will have a huge die with lots of HBM memory capacity, huge heat-sinks and extremely complicated and expensive motherboards making it very expensive for the consumer market.
Also AMD would not want to cannibalize its High-End CPU + dGPU products and they will most likely only produce such a chip for the HPC and Server market where they can sell at high margins.

Edit: BUT, a low power, low cost derivative with performance that could rival a 100mm2 dGPU could actually be made for the Mobile/Desktop consumer market.

HBM simplifies system architecture. With HBM you can eliminate DDR4. btw HBM is on the same package as the GPU. The GPU and HBM are built on the silicon interposer. HBM is not on the motherboard. Why is this very important detail lost on you ? With HBM there is no need for DDR4 memory slots. Power efficiency improves as you do not have to run memory I/O on motherboard traces. The power required to drive memory I/O on an interposer is much lesser than that required to drive on motherboard.

At 14nm FF with HBM2 you can get 32 GB HBM DRAM with 1 TB/s bandwidth for a 250W APU. Some really interesting HPC server designs are going to be possible with HBM.

 

[Silverforce11]

HBM simplifies system architecture. With HBM you can eliminate DDR4. btw HBM is on the same package as the GPU. The GPU and HBM are built on the silicon interposer. HBM is not on the motherboard. Why is this very important detail lost on you ? With HBM there is no need for DDR4 memory slots. Power efficiency improves as you do not have to run memory I/O on motherboard traces. The power required to drive memory I/O on an interposer is much lesser than that required to drive on motherboard.

At 14nm FF with HBM2 you can get 32 GB HBM DRAM with 1 TB/s bandwidth for a 250W APU. Some really interesting HPC server designs are going to be possible with HBM.

Ahh forgot about that. Incorporating HBM onto the APU package itself would surely simplify motherboard designs. But if its used together as system ram + vram combined, that complicates things for the OS, no? Would be safer with DDR4 system ram and HBM is pure vram.

 

[BFG10K]

It seems like AMD is banking on the slow death of the discrete GPU and that the future is APUs, which is why they are going for 300W APUs. This seems to corrobate with Linus Torvald's position that dGPUs are going to inevitably die because they are inefficient.

What a load of nonsense. Have you seen the typical cooler needed for a 250W+ TDP GPU?

Why don't you explain to us how they're going to integrate such a cooler onto a motherboard:

 

[antihelten]

What a load of nonsense. Have you seen the typical cooler needed for a 250W+ TDP GPU?

Why don't you explain to us how they're going to integrate such a cooler onto a motherboard:

Erm, your average high end CPU cooler already has far more fin area and fan air flow* than your average high end GPU cooler, so at least that part of the equation shouldn't really be a problem.

*CPU cooler fans are generally run at a lower rpm though.

 

[dacostafilipe]

Erm, your average high end CPU cooler already has far more fin area and fan air flow* than your average high end GPU cooler, so at least that part of the equation shouldn't really be a problem.

*CPU cooler fans are generally run at a lower rpm though.

This.

And with CPU AIO cooler around, I don't see where the problem is.

 

[Enigmoid]

HBM simplifies system architecture. With HBM you can eliminate DDR4. btw HBM is on the same package as the GPU. The GPU and HBM are built on the silicon interposer. HBM is not on the motherboard. Why is this very important detail lost on you ? With HBM there is no need for DDR4 memory slots. Power efficiency improves as you do not have to run memory I/O on motherboard traces. The power required to drive memory I/O on an interposer is much lesser than that required to drive on motherboard.

At 14nm FF with HBM2 you can get 32 GB HBM DRAM with 1 TB/s bandwidth for a 250W APU. Some really interesting HPC server designs are going to be possible with HBM.

Sure, as long as your designs use less than 32 GB ram.

The problem with HMB will be that you can't change the amount of RAM in the system...if all of a sudden your code needs to change (say a new project) that requires 40 GB ram well then you need to buy completely new hardware.

Otherwise I agree with you.

 

[NTMBK]

Sure, as long as your designs use less than 32 GB ram.

The problem with HMB will be that you can't change the amount of RAM in the system...if all of a sudden your code needs to change (say a new project) that requires 40 GB ram well then you need to buy completely new hardware.

Otherwise I agree with you.

Could always have a system like Knight's Landing- 16GB of high speed stacked memory on the package, but support >400GB of DDR4 as well.

 

[S.H.O.D.A.N.]

Erm, your average high end CPU cooler already has far more fin area and fan air flow* than your average high end GPU cooler, so at least that part of the equation shouldn't really be a problem

You average CPU runs at much lower temperatures across the load scale as well.

 

[Enigmoid]

Could always have a system like Knight's Landing- 16GB of high speed stacked memory on the package, but support >400GB of DDR4 as well.

Which is why there will be a number of boards with RAM slots and a lot of times those channels will be populated.

32 GB HBM is insufficient for a lot of tasks.

 

[AnandThenMan]

What a load of nonsense. Have you seen the typical cooler needed for a 250W+ TDP GPU?

A heavily overclocked CPU can draw 200+ watts and a good air cooler has no trouble dealing with it. The main impediment on a GPU cooler is the form factor which is generally not an issue for a motherboard/processor.

 

[ShintaiDK]

There is a difference on what OCed can do and what guaranteed function requires.

Also it will add extra cost to the mobo. Extra cost for cooling. Extra cost for cache. And outside a niche, for what purpose. On the consumer segment it will just yet again be a product without buyers. Weak CPU with a good GPU in a nuclear furnace without the ability to compete in the metrics that matters

 

[AnandThenMan]

There is a difference on what OCed can do and what guaranteed function requires.

Also it will add extra cost to the mobo. Extra cost for cooling. Extra cost for cache. And outside a niche, for what purpose. On the consumer segment it will just yet again be a product without buyers. Weak CPU with a good GPU in a nuclear furnace without the ability to compete in the metrics that matters

Your predictions are almost always wrong, so I'm confident that AMD is on to something big.

 

[ShintaiDK]

Your predictions are almost always wrong, so I'm confident that AMD is on to something big.

I can understand your frustration, and it shows in your posts when dealing with those you dont like. It just doesnt change anything. But again, hope doesnt cost anything. R&D however does.

 

[njdevilsfan87]

My dCPU already pulls ~250W and my dGPU ~300W. But a ~300W APU would probably be much more efficient, so we'll see...

 

[frozentundra123456]

Well, a huge high power APU with HBM does have appeal, but I see it definitely as a niche product for gamers and perhaps for other power users if they can use the igp for their productivity tasks. And personally, on the desktop, I probably would still prefer cpu plus discrete if the price/performace was similar, because of greater flexibility to change the gpu without replacing both.

But lets face it, the trend is toward mobile, low power, and away from desktop. Now if they can get HBM into an affordable laptop that will give performance like say GT860m at a cheaper price than intel plus discrete, that would be an application I would be interested in.

But with all the excitement over HBM and gpu performance, lets not forget the cpu performance needs to be good as well, and that is totally unknown at this point.

 

[antihelten]

You average CPU runs at much lower temperatures across the load scale as well.

And the fact that you're targeting a much lower temperature, is of course the entire reason why CPU coolers are generally bigger and more efficient than GPU coolers.