AMD CPUs, SoC Rumors and Speculations Temp. thread 2

[con635]

When do you guys think we will see HBM on budget APUs?

The only thing I've read is early 2017 for zen/greenland/hbm apu.

 

[de5_Roy]

AMD to offer mobile business-class Carrizo APUs
http://www.cpu-world.com/news_2015/2015081901_AMD_to_offer_mobile_business-class_Carrizo_APUs.html

 

[gamerk316]

When do you guys think we will see HBM on budget APUs?

Never, because it's never going to replace DDR as system RAM, due to cost reasons.

 

[jimmysmitty]

When do you guys think we will see HBM on budget APUs?

Never, because it's never going to replace DDR as system RAM, due to cost reasons.

If anything will replace DDR it will be HMC. Something will need to change in the memory standard as it is going to become slow with all the other standards out there.

That said, HBM could be used just for the APU much like Intels eDRAM cache for its iGPUs. But I doubt HBM will ever replace system RAM.

 

[gamerk316]

When do you guys think we will see HBM on budget APUs?

Never, because it's never going to replace DDR as system RAM, due to cost reasons.

If anything will replace DDR it will be HMC. Something will need to change in the memory standard as it is going to become slow with all the other standards out there.

That said, HBM could be used just for the APU much like Intels eDRAM cache for its iGPUs. But I doubt HBM will ever replace system RAM.

Problem again, is cost.

 

[juanrga]

When do you guys think we will see HBM on budget APUs?

Never, because it's never going to replace DDR as system RAM, due to cost reasons.

If anything will replace DDR it will be HMC. Something will need to change in the memory standard as it is going to become slow with all the other standards out there.

That said, HBM could be used just for the APU much like Intels eDRAM cache for its iGPUs. But I doubt HBM will ever replace system RAM.

HMC has been designed as replacement for JEDEC DDR4.

HBM is the JEDEC replacement for GDDR5.

HMC is already being used as replacement for DDR4 in Fujitsu CPUs. The KNL Xeon is using both HMC and DDR4, but this is due to limitations with current design of the CPU. One Intel lead engineer gave a talk at ISC15 about how plan to use more stacked DRAM in future and drop DDR entirely. Some believe he gave hints to the future KNH CPU.

 

[cdrkf]

When do you guys think we will see HBM on budget APUs?

Never, because it's never going to replace DDR as system RAM, due to cost reasons.

If anything will replace DDR it will be HMC. Something will need to change in the memory standard as it is going to become slow with all the other standards out there.

That said, HBM could be used just for the APU much like Intels eDRAM cache for its iGPUs. But I doubt HBM will ever replace system RAM.

Problem again, is cost.

We know they are developing a HBM apu for server. I wouldn't totally rule out a consumer APU with a single stack of HBM on it (netting between 128 and 256gb/s of bandwidth depending on which version), although such a part would be at the top of the product stack and wouldn't be cheap. If anything I think it could be a real winner in a high performance laptop part....

 

[Cazalan]

We know they are developing a HBM apu for server. I wouldn't totally rule out a consumer APU with a single stack of HBM on it (netting between 128 and 256gb/s of bandwidth depending on which version), although such a part would be at the top of the product stack and wouldn't be cheap. If anything I think it could be a real winner in a high performance laptop part....

Any form factor that can deal with fixed memory capacity could be replaced with HBM2.

Various embedded computers, NUCs, AIOs, laptops. Gaming systems of course. Especially if you want to get to smaller form factors and lower power.

There is no reason the cost can't be eventually brought down to a level similar with the premium Intel charges for the eDRAM versions of their parts.

 

[Cazalan]

AMDs experimentation with on package memory actually started with a CPU.

HotChips 2015 info, the long road to HBM.

http://wccftech.com/amd-reveals-radeon-r9-nano-performance-efficiency-numbers-hot-chips-2015-prototyping-hbm-july-14/

 

[de5_Roy]

AMD readies A8-7690K and A10-7890K APUs
http://www.cpu-world.com/news_2015/2015082501_AMD_readies_A8-7690K_and_A10-7890K_APUs.html

AMD Details Carrizo APUs Energy Efficient Design at Hot Chips 2015 - 28nm Bulk High Density Design With 3.1 Billion Transistors, 250mm2 Die
http://wccftech.com/amd-carrizo-apu-architecture-hot-chips/

 

[8350rocks]

AMD readies A8-7690K and A10-7890K APUs
http://www.cpu-world.com/news_2015/2015082501_AMD_readies_A8-7690K_and_A10-7890K_APUs.html

AMD Details Carrizo APUs Energy Efficient Design at Hot Chips 2015 - 28nm Bulk High Density Design With 3.1 Billion Transistors, 250mm2 Die
http://wccftech.com/amd-carrizo-apu-architecture-hot-chips/

Interesting that Godavari turned out to be the architecture name for those new APUs. I know there was some dispute over the naming of that new stepping for those APUs, and someone (honestly cannot recall who...) was mistakenly saying Godavari was the name of a GPU...

 

[cdrkf]

AMD readies A8-7690K and A10-7890K APUs
http://www.cpu-world.com/news_2015/2015082501_AMD_readies_A8-7690K_and_A10-7890K_APUs.html

AMD Details Carrizo APUs Energy Efficient Design at Hot Chips 2015 - 28nm Bulk High Density Design With 3.1 Billion Transistors, 250mm2 Die
http://wccftech.com/amd-carrizo-apu-architecture-hot-chips/

Interesting that Godavari turned out to be the architecture name for those new APUs. I know there was some dispute over the naming of that new stepping for those APUs, and someone (honestly cannot recall who...) was mistakenly saying Godavari was the name of a GPU...

To be honest the name Godavari could well have been re-purposed, AMD's plans on their 'existing' tech appear to be rather fluid lately (I guess as funds are being moved towards Zen and Artic Islands),

I wouldn't be surprised if Godavari was intended to be something else (e.g. a desktop Excavator based apu?).

Not that I'm complaining, Zen is what's important as Excavator whilst being a nice improvement, wouldn't be enough to impact AMD's competitive position really (I know many argue Zen won't because it won't have the IPC to match Intel, that said a 40% IPC boost + new process node has *got* to bring them into contention at more price points than they can currently compete imo).

 

[gamerk316]

OK guys, can we NOT get the thread locked down again? Don't kick the hornets nest please.

 

[Saga Lout]

Good advice gamerk316.

 

[cdrkf]

Saga, gamer, I feel that's targeted at me...? Can't really see how anything I've said is unreasonable and I can assure you I've no intent to upset anyone.

I don't think it's reasonable to have to censor legitimate thoughts on the topic because of the potential over reaction of another member. Frankly I should be free to think whatever, and present such thoughts so long as it's polite and doesn't contravene the forum guidelines.

I mean what's the point having a discussion forum about something if we can't explore (within reason) all aspects of it?

 

[gamerk316]

Was directed at the post above yours; don't want another 8350/Juan argument derailing the thread again.

 

[Saga Lout]

I didn't target anyone, cdrkf. It's a matter of reasoned debate where if anything is attacked, it's the post and not the poster. Opinions will always vary, often in the teeth of technical evidence, but it should still be argued in a civilised manner.

 

[juanrga]

I guess AMD and Tomshardware clarified things with the "Don't Call It Godavari" article

http://www.tomshardware.com/news/amd-kaveri-apu-update-not-godavari,29196.html

 

[juanrga]

Zen is what's important as Excavator whilst being a nice improvement, wouldn't be enough to impact AMD's competitive position really (I know many argue Zen won't because it won't have the IPC to match Intel, that said a 40% IPC boost + new process node has *got* to bring them into contention at more price points than they can currently compete imo).

My complaints and doubts are more related to the process node than microarchitecture. Zen could hit Haswell-like IPC, but I don't see how AMD will hit 4GHz base frequencies.

 

[-Fran-]

My complaints and doubts are more related to the process node than microarchitecture. Zen could hit Haswell-like IPC, but I don't see how AMD will hit 4GHz base frequencies.

Uhm... That is a fair question to have. I think I might have an answer...

Intel and AMD have been toying around with "Turbo" statements instead of "sustained" frequency, so AMD could do just that. Target 4Ghz+ for single core and 3.5Ghz for sustained. Those bursts can be using a lot of voltage, but they won't be for long. Rephrasing, they can play with the TDP to hit some target frequencies, but not constantly, since throttling will kick in.

Now, they can keep those target frequencies until the process matures and they can either bump it, lower TDP and keep the frequency or just bump base clocks.

At the end of the day, the process maturity has the answer, but I haven't read anything in regards "how well" Samsung's and GF'es nodes are coming out. Maybe we could try and find something in that regard.

Cheers!

 

[juanrga]

I said 4GHz base frequencies. I am discussing base frequencies, not turbos. My old prediction for Zen already contains ~3.5GHz base frequencies.

A mature process is not the answer, because Intel process will be more mature as well.

 

[-Fran-]

Ah, I misread your comment then.

And what does Intel having a more mature process has anything to do with GF or Samsung having a mature enough process to have 4Ghz target frequencies?

It sounds like you're implying that if AMD comes with 4Ghz base, Intel will answer with 4ghz+ right away. I don't see that as a bad thing. AMD might not be able to have the same base frequencies, but at least will be competitive. Rephrasing that, if AMD manages (somehow) to have 4Ghz base, Intel won't be able to pull much further away keeping TDP in check. 300Mhz maybe worth of difference at the same TDP if that were the case?

I'll stick to my intuition of checking on process maturity, since usually the uArch is tuned for how the process might turn out, right?

Cheers!

 

[ex_bubblehead]

Let's watch the arguing. It's very close to thread lock time.

 

[salgado18]

I said 4GHz base frequencies. I am discussing base frequencies, not turbos. My old prediction for Zen already contains ~3.5GHz base frequencies.

A mature process is not the answer, because Intel process will be more mature as well.

The probability that AMD will match Intel in base frequencies with the first product of a new process is very small. Intel is already at 4GHz base with Skylake, by next year they will be able to push farther even on the same process.

AMD could have some more strength if it can get turbo frequencies up to 4GHz at least, not much from a performance perspective, but a marketing one. Remember that the FX 9590 is labeled as the first 5GHz processor, even if on turbo.

But, as cdrkf said, having Haswell-like IPC with 3.5-4.0 GHz could be enough for AMD to be competitive in higher price points than it can now. So they may not reach same frequencies as Intel, but that may be enough to keep them afloat for longer.

 

[Cazalan]

Intel has base 4GHz in its flagship quad-core product. The bulk of the SKUs are still well below that, which is all AMD needs to remain competitive.

Especially the 8+ core parts which are in the 2.4(base)/3.2GHz(turbo) to 3.2(base)/3.6GHz(turbo) range.

If they can hit the 40% IPC gain and get near a 3.5Ghz base they'll be back in the game.