akamateau's Avatar

akamateau profile

4
Dave James Avatar
Dave James responded to akamateau's comment in
8 Months ago
Intel won't be selling you a Kaby Lake chip with an AMD GPU in it this yearIntel won't be selling you a Kaby Lake chip with an AMD GPU in it this year
akamateau Avatar

David;


I never once said in my post that "people will be able to pick off the shelf and build systems around themselves. " You made that up all by yourself. Don't put words in my text.


You missed my point completely. AMD is likely intending to sell Intel dGPU silicon that Intel can place on an Interposer along with an Intel server CPU.


This would not impact AMD APU sales as it is not the same thing. AMD is doing the very same thing with it's exascale Tarnhelm and Naples for the Chinese server market.


However it also does not mean that a dGPU would not be placed on an Interposer with HBM and a CPU for workstation uses.


AMD has remade itself into a graphics company that happens to sell x86 CPUs and APU's rather than a CPU company that happens to sell GPU's.


This is a huge difference. AMD is not going to turn it's back on dGPU sales to Intel as Intel has 95% of the server market. The best AMD can hope for server penetration is maybe 5% growth. However it makes a hell of a lot more sense to sell dGPU for servers to Intel as now AMD not only has penetration into a new market but that is a hell of an endorsement for the Radeon Tech.

Reply
Dave James Avatar

Fair enough, from what Bennett was saying in the thread it didn't seem like he was expecting it to be something for the server space.


Still, if it is going to be as you say it's still not going to be a Kaby Lake processor you'll be able to buy and drop into your gaming rig. Which is the thrust of the story - this AMD/Intel collaboration isn't going to be a chip for the general PC upgrade market.


If it were that would be creating competition for themselves which they've said they won't do when licensing out IP.

sign in to comment
Dave James Avatar
Dave James responded to akamateau's comment in
8 Months ago
Intel won't be selling you a Kaby Lake chip with an AMD GPU in it this yearIntel won't be selling you a Kaby Lake chip with an AMD GPU in it this year
akamateau Avatar
@David James

"If that was what this licensing deal was proposing it would essentially mean AMD giving notice they were killing off their APU division."

If you believe that an add-on slab of Radeon silicon is going to come even CLOSE to a fully integrated Heterogeneous System Architecture AMD Ryzen APU then you do not know what you are writing about.

While an on-die gpu will certainly exhibit less latency than a discrete gpu, a similarly spec'd APU will perform circles around it. In fact ONLY AMD supports Asynchronous Compute. What that means is an APU with the same compute units will crush the on-die version handily. That does not mean that Intel may be using a larger gpu. In fact AMD has announced a very thin gpu silicon for interposer and slim notebooks and tablet use.

Likely what AMD might be building for Intel is a similar GPU they intend for use with Ryzen: an Interposer mounted GPU with additional V-ram. AMD has patents regarding this particular architecture and it something NVidia can not build as they do NOT have this patent. In fact NVidia is not able to build an interposer with memory controller on die, they must use a controller at the bottom of each HBM stack. Again read GH Loh's patents.

Just in case you disagree then I would suggest that you do a google search for the patents for Gabriel H Loh. He is an AMD fellow and is co-owner along with AMD of many stacked and HBM memory architecture patents including the use of programmable cache, interposer memory controllers and interposers with discrete gpu's.

Then run over to Wcctech for this piece: it is 2 years old.

"AMD Reveals the Monsterous ‘Exascale Heterogeneous Processor’ (EHP) with 32 x86 Zen Cores and Greenland HBM2 Graphics on a 2.5D Interposer....
The APU, dubbed an “Exascale Heterogeneous Processor” or EHP for short is the mother of all APUs with 32 Zen Cores, an absolutely huge Greenland graphics die and upto 32 GB of HBM2 memory – all on a 2.5D interposer."

I can see AMD building silicon for an Intel CPU and why not? They are still selling silicon. And likely the GPU will not be Greenland!!

An on-die GPU will certainly better Intels laughably poor iGP but it can not do better than Heterogeneous System Architecture. unless of course the GPU is a discrete GPU running 2000-4000 shader cores.Reply
Dave James Avatar

I'm not disputing a closed, on-die solution, closely integrating the GPU and CPU silicon in a way that they can be accessed interchangeably by the system would be faster than simply tacking a graphics chip onto an existing package, but I can't see AMD willing to let that happen in a space they already operate in.


Given Intel's historic dominance of the CPU space that would result in unnecessary competition for their Raven Ridge chips, competition they would've then helped foster.


My contention is that this is not going to result in a hybrid Core/Radeon chip that people will be able to pick off the shelf and build systems around themselves. It's a chip that's more likely to occupy a segment of the tech market where the RTG won't be in competition with AMD in a wider context - hence dropping it into a closed system like a Mac or a sealed HMD.

sign in to comment
akamateau Avatar
akamateau responded to Belimawr comment in
2 Years ago
AMD: "Maxwell is not capable of asynchronously executing graphics and compute"AMD: "Maxwell is not capable of asynchronously executing graphics and compute"
Belimawr Avatar

give them time, it's still new and Nvidia can do way more with software fixes to help performance than AMD ever has, so I would just give it time as they will likely still be working on perfecting the drivers since Nvidia has always been about the software support and it's the reason Geforce cards have nearly always been more power efficient as AMD just went for brute force.

akamateau Avatar

Asynchronous Shader Pipelines and Asynchronous Compute Engines is hardware IP.


NVidia tries to implement it likely they will be sued.


NVidia is also not a member of HSA. Async SP and compute engines are both derived from IP shared by HSA members. Of course actual hardware implementation of the standard is AMD intellectual property.


It would serve NVidia right to get sued. They like to sue everyone else!!

sign in to comment