AMD's Phil Rogers jumps ship to Nvidia

[Shehriazad]

Yea, well then why doesnt intel give us more of them? Stuck on 4 for what, nearly 10 years now?

More cores have been available for many many years now...I mean come on.

Phenom II 6 cores, 2010.
And I'm pretty sure you can right now buy a certain Intel 2011-3 chipset CPU that also has 6 cores.


But that's hella off topic now.


On topic: I'm not really surprised...AMD is radically shrinking these days, but I'd really like to know the circumstances here before any doom and gloom...but doom and gloom will always happen. Yes yes, AMD is gonna be bankrupt in a month, I get it. :whiste:

 

[raghu78]

A 17h core will beat a 15h core, but a 17h core will not beat a 15h module.

1C2T(17h) > 1C1T(15h)
1C2T(17h) < 2C2T(15h)

As usual you have nothing to back up your statements. you are just hot air. :whiste:

 

[LTC8K6]

Yea, well then why doesnt intel give us more of them? Stuck on 4 for what, nearly 10 years now?

http://ark.intel.com/products/58664/Intel-Core-i7-980-Processor-12M-Cache-3_33-GHz-4_8-GTs-Intel-QPI

 

[Arachnotronic]

Yea, well then why doesnt intel give us more of them? Stuck on 4 for what, nearly 10 years now?

Apparently they're going to do just that with Cannonlake...

http://forums.anandtech.com/showthread.php?t=2449674&highlight=

 

[superstition]

Yeah, Skylake is under-performing but AMD's efforts in this space have been just fantastic. I guess you will be buying an Excavator-powered APU next year for your next desktop PC?

Compared to the hype for Skylake (OMG! Going to revolutionize computing! Huge paradigm shift!) it is underperforming on the desktop. One set of reviews made me rather get Broadwell. But, my impression (although I didn't look at the review closely) is that it's more impressive in the Xeon space.

 

[AtenRa]

AMD actually gave up HSA years ago.
Zen is way more important.

Carrizo is the first fully HSA CPU and not the last one.

 

[Arachnotronic]

Carrizo is the first fully HSA CPU and not the last one.

Yeah and guess what they're hyping up as coming next? Zen standalone CPUs

 

[AtenRa]

Yeah and guess what they're hyping up as coming next? Zen standalone CPUs

And this has to do what with HSA ??? You do know there will be ZEN based APUs, so what are you talking about ??

 

[Arachnotronic]

And this has to do what with HSA ??? You do know there will be ZEN based APUs, so what are you talking about ??

I do, but they don't appear to be the priority to AMD.

 

[AtenRa]

I do, but they don't appear to be the priority to AMD.

They already have a very competitive HSA product (Carrizo), ZEN is more competitive for the Data Center and Cloud and the reason it starts its life as a CPU and not APU.

 

[Arachnotronic]

They already have a very competitive HSA product (Carrizo), ZEN is more competitive for the Data Center and Cloud and the reason it starts its life as a CPU and not APU.

How much good is HSA support doing for Carrizo? Where is the HSA software ecosystem? It seems to me that aside from a few token design wins, the majority of the OEMs are using Skylake for their systems during this product cycle.

 

[Good_fella]

AMD will keep Roy Taylor and Richard Huddy till the end. :whiste:

 

[AtenRa]

How much good is HSA support doing for Carrizo? Where is the HSA software ecosystem? It seems to me that aside from a few token design wins, the majority of the OEMs are using Skylake for their systems during this product cycle.

As i said before, Carrizo is the FIRST APU with full HSA 1.0 support. You dont expect software to be released with the first HSA product at the same time.
Also, AMD is not the only one designing HSA hardware.

And my initial response was to the commend that AMD gave up HSA years ago, which is simple false.

 

[ThatBuzzkiller]

How much good is HSA support doing for Carrizo? Where is the HSA software ecosystem? It seems to me that aside from a few token design wins, the majority of the OEMs are using Skylake for their systems during this product cycle.

Well there is OpenCL 2.0 out there to for any devs who want to explore some of those features but I do get your point ...

It'll be interesting to see what ray tracers will do with it ...

 

[DrMrLordX]

As i said before, Carrizo is the FIRST APU with full HSA 1.0 support. You dont expect software to be released with the first HSA product at the same time.
Also, AMD is not the only one designing HSA hardware.

And my initial response was to the commend that AMD gave up HSA years ago, which is simple false.

Okay, hold on here. Let's not go full-bore pro HSA without clarifying a few things.

Kaveri supports every major HSA feature except for context switching. Context switching is sort of a big deal in some situations, but not really overall. You can still have meaningful acceleration of applications via the iGPU without it. Kaveri's been out since 2014. The HSA .95 standard and the later 1.0 Provisional standard were both fully-supported by Kaveri. There was a fledgling software ecosystem based around drivers and a runtime that were pretty-much only available on Linux. That ecosystem did not mature, and some relatively high-profile projects like Project Sumatra went belly-up.

Kaveri was supposed to be the Trojan horse that let HSA into a non-trivial number of PCs. Sadly, the software support for that future never showed up. Even one killer app, like HSA-enabled 7zip or HSA-enabled Firefox or . . . something, just didn't happen. AMD needed at least one real-world benchmark where they could say that the 7850k was faster than a Haswell i5 or i7. It never happened.

Looking to Carrizo to be the torchbearer for HSA is not a good idea. Now that Intel has moved from Gen7 to Gen9 graphics, Intel can do just about everything with Gen9 that AMD is able to do with Carrizo. The difference here is that Intel is reliant on OpenCL 2.0. The edge that AMD had (or could have had) in that department is at least partially vanished. 512 shader GCN iGPUs may still be faster than some of the cheaper Gen9 iGPU implementations in GPGPU scenarios.

AMD certainly appears to have given up on HSA. Many of the people who were involved with open-source HSA projects like Okra, aparapi lambda, and the like got axed last year. With OpenCL 2.0 maturing the way it has, it seems more practical for AMD to simply supply compliant drivers and be done with it. If AMD is getting any support from their "HSA partners", it isn't showing publicly. And I can't think of anyone but AMD that has bothered to actually produce HSA-compliant hardware.

 

[LTC8K6]

http://ark.intel.com/products/58664/Intel-Core-i7-980-Processor-12M-Cache-3_33-GHz-4_8-GTs-Intel-QPI

This one was a bit earlier:

http://ark.intel.com/products/47933...ocessor-12M-Cache-3_20-GHz-4_80-GTs-Intel-QPI

 

[dark zero]

Seems that AMD is finally knowing that they will die sooner than expected. Even the most loyal people are leaving. Seems that their end might come after Zen.

 

[William Gaatjes]

It is not that strange, I feel AMD was betting on HSA a lot. The idea is that the cpu would offload fpu calculations to the much faster gpu (on the promise that it can be done in parallel of course). That is the only reason i can think of that AMD went for 2 integer cores and one fpu core for a module. Combine that with an APU with HSA. It makes sense. Now that HBM is here and HBM2 is coming and that HBM is for now way to expensive to be put on the same package as a cpu, there will be dedicated CPU's again from AMD and discrete gpu's with HBM. After that, if AMD can get another design win for a console, i expect a HBM APU to be presented. And when that happens, sooner or later, there will be an AMD X86-64 SOC with HBM , gpu and cpu for mobile platforms and desktop pc's. Depending on the dissipation, the execution speed will be lower and higher. Ofcourse, ddr4 channels will still be there. The HBM memory would be for graphics or heavy HSA alike calculations. the ddr4 memory pool would be to put "slow" data in. Before it is swapped out to the ssd.

Too put it in short words : HSA has been put on hold at AMD.

 

[Blitzvogel]

HSA should not have been the priority. AMD should've focused on FPU throughput, hell they shouldn't have gone the BD route in the first place, but getting dual 256-bit FPUs in by 2013 would've made AMD chips a much better proposition for people who game, are into content creation, or even scientific computing.

 

[DrMrLordX]

Eh, HSA wasn't a bad idea. Launching HSA-compliant hardware with no software ecosystem for developers or end-users WAS a bad idea. I do not think that HBM on video cards really addresses the problems corrected by HSA and OpenCL 2.0: reduced copy overhead from system to card and latency.

If you have a block of thousands (or more) instructions you want crunched at predictable intervals and want to send it to a dGPU, great, you can probably benefit from that. Bonus if the memory footprint of the data involved is very small. If you have small blocks of 10-20 mathematical instructions you want crunched at erratic intervals and you need instant turnaround, you need/want your GPU to be an iGPU. Or you need something like NVlink.

 

[dark zero]

Even Open CL is underpromoted thanks to MS domain...

 

[RussianSensation]

Seems that AMD is finally knowing that they will die sooner than expected. Even the most loyal people are leaving. Seems that their end might come after Zen.

If someone offered you a promotion with much higher pay and a lot more stock options, wouldn't you leave your existing employer? There is always a lot more to the story of an executive leaving the firm that the typical DOOM/bankruptcy any time now.

I am pretty sure if all the Doom experts on AT were so sure of their own gospel, they would all be very wealthy from shorting AMD stock and buying thousands of puts over the last decade. Where are these wealthy financial chair experts driving Ferraris and McLarens who made $ by shorting AMD stock/buying puts over the years? Most of these armchair experts claiming AMD is doomed/going bankrupt don't even own a 5960X with Titan X Tri-SLI that can easily be purchased from a single successful short/annual put strategy, so we all know it's all just talk with nothing to back it up.

Your posts aren't even mildly entertaining, just the same baseless thing repeated over and over -- AMD/NV are soon dead, Intel will take over GPUs. :whiste: Serious question, how old are you?

Even Open CL is underpromoted thanks to MS domain...

Is that why Apple is using AMD generation after generation? Is that why OpenCL acceleration has been added to hugely popular productivity apps like Photoshop?

Keep trolling and posting the same drivel that over the last 6-8 years has basically destroyed the once amazing and thriving CPU & overclocking sub-forum. Today, 95% of posts are AMD is dying, doom and gloom and it's honestly stupid to read them every year for the last 10 years. You are like guy #100 that repeats the same thing over and over and over and AMD is still alive in 2015, and will be in 2016. 1-2 executives leaving doesn't mean the firm is about to go bankrupt....and even if they do leave, there are many reasons for people to leave such as they want a new challenge, higher salary, promotion, more responsibility heading a larger company, etc.

If the top talent of Naughty Dog were to leave eventually, does it mean Naughty Dog is DOOMED? Going bankrupt? All future Uncharted games are DOOMED?

Any negative news on AMD, and the usual suspects are declaring impeding bankruptcy. Right.

 

[William Gaatjes]

Eh, HSA wasn't a bad idea. Launching HSA-compliant hardware with no software ecosystem for developers or end-users WAS a bad idea.

I agree. The idea is good. AMD should have done what NVidia does properly : Proper software support. Nvidia has all these ....works solutions to promote the features of their videocards. AMD had Mantle which is in sense absorbed into directx12. But AMD should have made more effort in working with software developers to show the benefits of HSA.

I do not think that HBM on video cards really addresses the problems corrected by HSA and OpenCL 2.0: reduced copy overhead from system to card and latency.

If you have a block of thousands (or more) instructions you want crunched at predictable intervals and want to send it to a dGPU, great, you can probably benefit from that. Bonus if the memory footprint of the data involved is very small. If you have small blocks of 10-20 mathematical instructions you want crunched at erratic intervals and you need instant turnaround, you need/want your GPU to be an iGPU. Or you need something like NVlink.

I agree that the unified memory idea were gpu and cpu can read write into each others memory directly is a good idea. And i wondered what happened to that idea now that the top models are HBM based discrete gpu cards.

I do think that AMD has plans for a HBM based apu / soc in the feature. Intel brings out cpu's with embedded dram. Now that i think of it, why could AMD not deliver a top model ZEN with HBM ? Just 512 MB would be enough. As fast L3 ? The question is : Will the cpu be able to do something useful with that much memory bandwidth ?

 

[RussianSensation]

I agree. The idea is good. AMD should have done what NVidia does properly : Proper software support. Nvidia has all these ....works solutions to promote the features of their videocards.

That's because they co-market games with developers/publishers which means NV is paying for some of the game's marketing. They also send free programmers on site to help implement these features into the game. The last time NV sighted, there are more than 300 engineers devoted ONLY to GameWorks.

AMD could never afford such a strategy with HSA in the last 3-4 years and if we go back 6-8 years ago, the management that ran the company wasn't smart enough to bribe developers with co-marketing $ and software engineering resources. So there you go.

Look at the reality -- the key founders of HSA do not include Intel and NV:

That means for any company lacking marketing $$$ and engineering resources, it's going to have to literally make HSA software / programs because 3rd parties sure as hell aren't going to do it given Intel's global market share. You are not going to have a push towards HSA unless all the key parties are on-board OR you throw huge marketing $ behind it and have 300+ software engineers ready to go just for this task.

The interesting part is NV doesn't have an x86 CPU and their ARM Tegra SoCs aren't exactly selling like hot cakes in servers. It will be interesting to see what exactly Rogers will do to benefit the server market as a Server Architect.

"HSA" is nothing more than a pipedream, quite frankly.

No, it's not a pipe dream but it's just very costly and technically difficult to implement. Even Samsung with tens of billions of dollars in cash still has not managed to implement HSA in their custom ARM processor that could go into the next gen of Galaxy line-up of smartphones, but it doesn't mean it won't happen at some point.

"According to recently released information, Samsung began to work on its custom general-purpose ARMv8-compatible core in 2011. For about four years now, the company has been hiring microprocessor developers from companies like Advanced Micro Devices. "
http://www.kitguru.net/components/c...sor-core-already-supported-by-software-tools/

We might start to see Samsung implementing HSA gen 1 in M1 starting with Galaxy S7 or it might take another 3-5 years but it's clear that other major players in the world outside of AMD do understand the benefits of HSA in the future. Now unless someone is 80-90 years old on this forum, they might never see the benefits of HSA but for the rest of us, we have decades to see if HSA can actually become reality in smartphones, tablets, laptops, PCs.

Even if AMD never implements HSA successfully, we also have ARM, Samsung, Qualcomm.

 

[Phynaz]

As i said before, Carrizo is the FIRST APU with full HSA 1.0 support. You dont expect software to be released with the first HSA product at the same time.
Also, AMD is not the only one designing HSA hardware.

And my initial response was to the commend that AMD gave up HSA years ago, which is simple false.

HSA failed. Period. End of story.