[Mynavi.jp] AMD reveals plans to 2020.

[Genx87]

So basically the same plan they have had since about 2008? It has worked out well for them so far.

 

[Silverforce11]

Hawaii is obviously not part of any APU from 2014. Secondly, it says dGPU on the slide so the red part is their discrete GPUs.
Pretty much confirms rebrands from AMD and most likely a 390X with high power consumption in 2015

Stop trying to draw conclusions on every sniff regarding the upcoming series. It's not long to go and we have plenty of leaks to draw any conclusion we want to already.

The point here is that HBM is the solution to the problem of APUs scaling up in performance. Bandwidth was the missing key.

With that solved, there's zero reasons why a 300W APU can't be a monstrous graphics accelerator along with everything else.

I used to disregard all the predictions that the dGPU's days are numbered, but this changes everything and it can happen very soon, a lot earlier than we all thought.

 

[Genx87]

Nvidia's Tegra is an effort to reduce their dependence on Geforce revenue. But in the last 2 years Geforce has generated a larger portion of revenue than traditionally. I am sure Nvidia will feel the pain starting in 2017 when the Zen based APUs with HBM basically makes it unnecessary or rather completely redundant to buy a 100 - 120 sq mm discrete GPU.

Nvidia was supposed to start feeling the pain of APUs in ~2010-2011.

 

[Genx87]

I think this pretty much confirms that AMD does not believe the dedicated CPU or GPU will survive. It will be on a single silicon in which the CPU size is x to 8 cores and the GPU can be scaled from y to 300 watts depending on gamer or VR need.

If ARMS don't go mainstream, Nvidia can kiss their future bye bye. I think that's the real reason why they've been pushing so hard for ARM SOCs and chrome, If x86 window maintains it's desktop and laptop dominance, we can kiss Nvidia bye bye in a few years.

Tegra is a diversification project. One that targets a high growth market. Unlike APUs that target a stagnant or shrinking market. x86 has already lost their dominance to mobile devices that are shipping ARM SOCs. It will only get worse moving forward.

 

[AtenRa]

Hawaii is obviously not part of any APU from 2014. Secondly, it says dGPU on the slide so the red part is their discrete GPUs.
Pretty much confirms rebrands from AMD and most likely a 390X with high power consumption in 2015

Pretty much doesnt confirm anything, Hawaii was a 2013 architecture and it seams the above roadmap is not up to scale. So dont draw any conclusions from it.

 

[positivedoppler]

Tegra is a diversification project. One that targets a high growth market. Unlike APUs that target a stagnant or shrinking market. x86 has already lost their dominance to mobile devices that are shipping ARM SOCs. It will only get worse moving forward.

Nvidia right now enjoys 75% of that shrinking x86 dgpu market that might vaporize once Intel can make a igpu that's good enough(and they will)

Nvidia competes with Intel, Qualcomm, Mediatek, Apple, and anyone who can afford a ARM license in the very overcrowded mobile market.

 

[raghu78]

Unless you need a high performance CPU. I doubt Intel will have anything even then that competes with nV's ~$100 dGPUs.

What do you mean by high performance CPU. Are you telling 8 cores and higher. btw are you assuming AMD Zen will not be competitive with Intel Skylake. I am optimistic (but not overly) that AMD will have a competitive Zen core. A quad core Zen APU should be more than enough for the vast majority of the mainstream PC gaming market.

 

[raghu78]

Nvidia was supposed to start feeling the pain of APUs in ~2010-2011.

No. the APU can never beat a dGPU of similar die size (i mean the GPU portion of APU and a dGPU of same size) until it has access to the same kind of bandwidth. HBM is the game changer and will be the nail in the coffin for the entry level dGPU.

 

[Yuriman]

What do you mean by high performance CPU. Are you telling 8 cores and higher. btw are you assuming AMD Zen will not be competitive with Intel Skylake. I am optimistic (but not overly) that AMD will have a competitive Zen core. A quad core Zen APU should be more than enough for the vast majority of the mainstream PC gaming market.

Kaveri is plenty for most already, but it's no i7.

I'm looking forward to seeing what Zen brings but I'm not expecting miracles considering AMD's R&D budget.

 

[Headfoot]

Basically a counter play to Xeon Phi. There was 0 indication in those slides of any anticipation of dGPUs dying... that's purely your own speculation

 

[VirtualLarry]

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.

I called it. I suggested that was the future of APUs, several months ago, in another thread. I said that AMD would have to "go big or go home", in order to make their APU proposition compelling in the face of discrete graphics, and to get their APUs out of their current "no-mans land".

http://forums.anandtech.com/showpost.php?p=37132693&postcount=616

 

[sandorski]

Nvidia was supposed to start feeling the pain of APUs in ~2010-2011.

I kinda wonder if NVidias high priced Vid cards is a result of the loss of the Lowend? No longer can defective chips be simply sold as lesser models, meaning that they have to eat that loss through markups on the Highend.

 

[AtenRa]

I called it. I suggested that was the future of APUs, several months ago, in another thread. I said that AMD would have to "go big or go home", in order to make their APU proposition compelling in the face of discrete graphics, and to get their APUs out of their current "no-mans land".

http://forums.anandtech.com/showpost.php?p=37132693&postcount=616

Hint, those 200-300W APUs are not for the consumer market

 

[NTMBK]

Hint, those 200-300W APUs are not for the consumer market

Bet they'll be released in the consumer market though. It'd make a great "games console" running Windows.

 

[Paul98]

This will be great, only problems right now are lack of memory bandwidth and weak CPU core. Those should be fixed between now and 2017.

 

[.vodka]

Bet they'll be released in the consumer market though. It'd make a great "games console" running Windows.

Exactly. These APUs would make for an excellent steam machine, DIY or otherwise.

 

[Genx87]

Nvidia right now enjoys 75% of that shrinking x86 dgpu market that might vaporize once Intel can make a igpu that's good enough(and they will)

Nvidia competes with Intel, Qualcomm, Mediatek, Apple, and anyone who can afford a ARM license in the very overcrowded mobile market.

Hence why Nvidia is diversifying. Intel wont make a dedicated GPU due to the resources it takes, and the margins arent where they need to be. Plus there is little driving factor for them. They own 70% of the GPU market producing crapware. So why triple or quadruple the effort for the final 30%?

 

[Cloudfire777]

Stop trying to draw conclusions on every sniff regarding the upcoming series. It's not long to go and we have plenty of leaks to draw any conclusion we want to already.

The point here is that HBM is the solution to the problem of APUs scaling up in performance. Bandwidth was the missing key.

With that solved, there's zero reasons why a 300W APU can't be a monstrous graphics accelerator along with everything else.

I used to disregard all the predictions that the dGPU's days are numbered, but this changes everything and it can happen very soon, a lot earlier than we all thought.

People put way too much weight on HBM being the savour which makes AMD cards trumph whats out there. If it was so easy we would have 512bit GTX Titan X. That they instead used 384bit bus and improved memory compression instead, means they dont need more bandwidth. HBM is for saving space, reducing heat and power from the VRAM while preparing for the future. Its a natural step to prepare for virtual reality and 8K resolution, but it will do very little performance vs a proper bus width thats made for the specific GPU today, which Im sure both AMD and Nvidia did a good job scaling earlier with GDDR5.

I dont deny the possibility of existance of APUs with huge IGPs and that they will hit the consumer market sooner or later anyway. Its unfortunately the way things will go. Soldered CPUs and GPUs in notebooks have become the way things work today mostly. They have already begun taking away our freedom. And both PS4 and XBox have huge APU die's with big IGPs that you originally found only as discrete graphic cards for desktops.

I think its sad that people brags about these products, AMD fan or not. Because once allinone solutions hit the market, they are here to stay and overclock enthusiasts and similar minded will have much fewer options out there and AMD and everyone that makes these APUs will have more control on development cycle and choices and margins and price of products while we consumers lose from it

 

[Genx87]

I kinda wonder if NVidias high priced Vid cards is a result of the loss of the Lowend? No longer can defective chips be simply sold as lesser models, meaning that they have to eat that loss through markups on the Highend.

Nvidia has had high priced GPUs since well before APUs showed up.

 

[Genx87]

No. the APU can never beat a dGPU of similar die size (i mean the GPU portion of APU and a dGPU of same size) until it has access to the same kind of bandwidth. HBM is the game changer and will be the nail in the coffin for the entry level dGPU.

We will have to wait and see. dGPUs in 2017 will have also greatly increased in performance.

 

[Zodiark1593]

The increasing igpu realestate in Broadwell over the cpu should have Nvidia really worried in the sub $100 gpu market. VR might be the only thing that can jumpstart the need for dgpu though.

Or anyone looking to game
Or anyone abusing gpu compute
Or any sort of 3D work

The igpu would need to be superior to dgpu not only for the consumer (to kill off dgpu that is), but for the professional as well.

 

[RussianSensation]

If AMD had a new architecture ready now, they would have changed the entire lineup.

I've seen you repeat this over and over. R9 390X was never going to be a brand new architecture. Not 1 rumour in the last 1.5 years even hinted at such a case. From day 1 everyone knew R9 390X is either GCN 1.2/1.3 or 2.0, but it's essentially the same GCN architecture, just an enhanced version of the foundations AMD built with Tahiti, then Hawaii and then Tonga. The so-called "Post-GCN" architecture was rumoured for 2016 and beyond, but even that is not certain.

Not sure where you even got this idea that R9 300 series was supposed to be some new non-GCN architecture. It generally takes 3-4 years to design a brand new architecture and AMD doesn't have enough teams or financial resources to do that like NV can. However, this only goes to show just how amazing GCN design was by Eric Demers that it will last from Dec 22, 2011 when HD7970 launched at least all the way until 14nm AMD cards in 2016. 7970Ghz competed well with 680, R9 290X with 780Ti and all rumoured specs point to R9 390X doing the same against Titan X. AMD should come in +/-10% of the Titan X for much less $. That's incredible from what is essentially the same underlying GCN architecture.

That doesn't matter though because GCN still has a ton of life left in it as AMD can still more than double the geometry performance of R9 290X, increase pixel fill-rate/colour fill-rate efficiency at least 40%, improve the efficiency of ACEs/compute units, enlarge L2 cache, and nearly double the effective memory bandwidth. With such dramatic increases in perf/mm2 for GCN still possible, there is no need to replace GCN in 2015.

At the same time, retaining GCN doesn't mean that R9 300 series can't be competitive and it certainty doesn't at all mean most R9 300 series desktop cards will be re-brands. For example, if AMD adds Tonga's improvements and uses a more mature 28nm node for R9 380/380X series, even with identical SP/ROP/TMU specs, they could squeeze 10% more performance with 15-20% lower power usage. Would you call a chip with 2816 SPs and 64 ROPs but those characteristics a rebrand of R9 290X? :hmm:

In reality, AMD could probably use GCN for yet another generation post R9 300 series courtesy of 14nm/16nm shrink + HBM2. GCN has proven to be an extremely flexible architecture both for gaming and compute, something that was absolutely not the case for Fermi or Kepler. Both of those NV GPU architectures proved to be inferior for DirectCompute and DP performance compared to GCN on a perf/mm2 basis. GCN was designed from day 1 as a compute architecture (Fermi and Kepler were not, and I think R9 390X will beat GM200 in compute as well when it launches). Because GCN is much more scalable, AMD can just improve specific areas of the design separately -- ACEs, geometry units, ROPs. Why should AMD ditch GCN if it still can be improved greatly to be competitive? It costs billions of dollars to design a brand new architecture and there is a risk it won't be as good (Phenom/Bulldozer, etc). Simply stated, AMD and NV have different approaches to GPU design based on the resources available to them.

NV spends a disproportionate amount of R&D relative to its sales compared to other top 9 firms in the semi-conductor business. R&D accounted for 31% of NV's 2014 sales. In comparison, look how much more efficient Samsung and TSMC are.

 

[Techhog]

Personally I can't wait for these "300w APU" tiny Mini-ITX's. They've got the required 1m high external cooler that will be dangling out the back of those tiny cases designed already... :biggrin:

"The dGPU is dead. The future is 300w APU's stuffed under one heatsink in a 2" high case". If this were two days later, I'd be calling April Fool...

Or, get this... multi-rad CLCs.

 

[redzo]

Isn't this going to kill the upgrade path? I mean, they will have to increase the variety of apu products a lot. Today you can combine any cpu with any gpu which is a great thing. Maybe you want a specific gpu which is available only on a specific cpu which is it not what you want(the cpu part). How can you solve that? This why I do not think that the dgpu will ever die. Midrange might shrink a lot, but the dgpu will continue to live on at the high end of the market.

 

[AtenRa]

Bet they'll be released in the consumer market though. It'd make a great "games console" running Windows.

This will be great, only problems right now are lack of memory bandwidth and weak CPU core. Those should be fixed between now and 2017.

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.