AMD's Big HPC Swing: Exascale Heterogeneous Processor [Updated]

[Shakaza]

Post Updated. Sorry guys, bad news.

Well shoot. That sucks. A lot. I was hoping this might, ideally, give AMD just the boost it needs to get back in the game. Welp, guess not then. gg, AMD; it's been fun. 

[SurvivorNVL]

One can hope and dream

 

Never the less, if AMD stays alive to their expected release date, it could be a huge contender.

I hope AMD stays alive, or that their outcome becomes more certain far sooner.  Given the variable-sync market is becoming crowded rather quickly, or well, more displays are being announced and tested; I am currently likely to jump on G-Sync, especially given AMD's uncertain future at this current time.  Or maybe jump on Freesync, given Intel could utilize it in the future and especially so if they got ATI.  ohgod.  Now I don't know again.  Anyway.  Hope something good happens soon for them, an acquisition, or massive, massive success with Zen.  Like a curveball.

[Shakaza]

One can hope and dream

 

Never the less, if AMD stays alive to their expected release date, it could be a huge contender.

I'd like to hope so, but I I recall seeing an article a while ago that said with AMD in it's current state of losing money, it would be bankrupt by, I think, 2020. 

[Notional]

What a shame. However if that theoretical prototype is actually 7 nm, it could very well be 1 die. However 5 years from now is irrelevant as way too many things could change.

 

Taking the idea of let's say a ZEN APU with HBM2 could very well be a weaker model of this. However 10 TF SP does not sounds like that much considering Fiji is 8.6 TF @ 28nm. So just going to 14nm FF should make it possible to bump that up, even if the performance is limited by the ZEN cores. So 10 TF per APU does not sound unrealistic for next year.

[Albatross]

This is for servers right ? Hence the ludicrious price point ... 

 

This article gives me hope that AMD's future is bright and not dark as it seems . Could be false hope ...

 

I hope so. :lol:

[6God]

Anyone else find the announcement of this so shortly after the announcement of the US Gov't investing into an Exascale Supercomputer?

 

Coincidence? I think not.

 

I doubt it's guaranteed, but I have no doubt that AMD is pursuing the contract on that order heavily. They would just love to beat Intel to the contract for supplying that big of a project.

 

They need to beat out IBM too, who would probably lowball the government bid just for shits and giggles since they already own so much of the HPC segment to begin with. 

[patrickjp93]

I hope AMD stays alive, or that their outcome becomes more certain far sooner.  Given the variable-sync market is becoming crowded rather quickly, or well, more displays are being announced and tested; I am currently likely to jump on G-Sync, especially given AMD's uncertain future at this current time.  Or maybe jump on Freesync, given Intel could utilize it in the future and especially so if they got ATI.  ohgod.  Now I don't know again.  Anyway.  Hope something good happens soon for them, an acquisition, or massive, massive success with Zen.  Like a curveball.

If you want the nice outcome where Intel gets ATI and Nvidia gets x86_64, buy Nvidia now and don't look back. If you want to risk one nasty future where Intel gets control over Nvidia and can squash AMD like a bug without risk of a serious competitor, buy AMD now.

[ErickS89]

I hope zen is a success, but knowing amd.... this is a gamble.

[patrickjp93]

What a shame. However if that theoretical prototype is actually 7 nm, it could very well be 1 die. However 5 years from now is irrelevant as way too many things could change.

 

Taking the idea of let's say a ZEN APU with HBM2 could very well be a weaker model of this. However 10 TF SP does not sounds like that much considering Fiji is 8.6 TF @ 28nm. So just going to 14nm FF should make it possible to bump that up, even if the performance is limited by the ZEN cores. So 10 TF per APU does not sound unrealistic for next year.

10TFlop DP. Whenever you see the Linpack flops ratings for supercomputers, it's always for double precision. The full (96 page) paper makes that more clear. This is a good read actually despite all the technicals for a new interposer fabric needed for high-speed communication. It really acts like 2 separate CPU clusters with 1 shared iGPU (asynchronous compute engines in OpenCL 2.x make this possible) and a secure transactional high bandwidth cache (suggesting an HMC-like technology given HBM is not designed to be transactional memory). This thing is a monstrous change in engineering on paper. I'm having nerdgasms myself, but the financial feasibility is making me only cautiously optimistic.

[SurvivorNVL]

If you want the nice outcome where Intel gets ATI and Nvidia gets x86_64, buy Nvidia now and don't look back. If you want to risk one nasty future where Intel gets control over Nvidia and can squash AMD like a bug without risk of a serious competitor, buy AMD now.

The nice outcome sounds more like a pitch for AMD/Intel than for nVidia.  And the nasty future sounds more like a pitch for nVidia to me.

[patrickjp93]

They need to beat out IBM too, who would probably lowball the government bid just for shits and giggles since they already own so much of the HPC segment to begin with. 

IBM and Oracle both, and scale-up workloads are still something Intel sucks balls at because it doesn't provide for direct accelerator access to system RAM, whereas IBM and Oracle both provide a passthrough around the CPU to reduce latency and get around bandwidth constraints greatly.

[patrickjp93]

The nice outcome sounds more like a pitch for AMD/Intel than for nVidia.  And the nasty future sounds more like a pitch for nVidia to me.

No, because if AMD collapses (you support Nvidia, don't help AMD), then who do you think will step forward with the money to pay off creditors and get its precious IP? The FTC would ensure Intel doesn't get the CPU IP and that Nvidia doesn't get the GPU IP. Both situations would represent toxic monopolies. However, those two would be the biggest North American companies to go after it. Likely the FTC would ensure these two companies get the opposing IP, giving Intel the freedom to make dGPUs without having to dance around both AMD's and Nvidia's GPU IP to do it, and giving Nvidia access to x86 like it wanted for the Denver Tegra. Getting AMD's Zen IP and then adding Denver on top would represent a huge issue for Intel and would also represent the first time in history a VLIW architecture didn't fall on its face in high performance needs situations. Denver doesn't actually execute ARM assembly. It reads ahead and creates VLIW instructions for the cores to execute. That's why its 2 cores outclass the A8 so readily but can't quite match the A8X. For Nvidia's 3rd try, that was damn good. Add that to what AMD has now and provide the x86 ubiquity along with a jump in clock speed, and Intel will have one nasty competitor in CPUs.

 

If You buy AMD and keep it alive long enough for Intel to knock Nvidia out of the accelerator market (while simultaneously eating up low-end and mid-end dGPU sales via its iGPUs) then Intel will not be in competition with Nvidia in any market, and the investors will be looking to make back the money lost in the process. Nvidia will be wide open to a buyout by Intel. At that point, AMD being financially weak would be an easy thing for Intel to crush outright, and no one would be willing to try to pick up the smoldering heap left behind. Intel will have won total dominance in everything HPC, Desktop, and mobile, leaving only phones and microcontrollers as markets it didn't dominate in computing.

 

One option guarantees competition. The other potentially dooms it in the long run.

[SurvivorNVL]

No, because if AMD collapses (you support Nvidia, don't help AMD), then who do you think will step forward with the money to pay off creditors and get its precious IP? The FTC would ensure Intel doesn't get the CPU IP and that Nvidia doesn't get the GPU IP. Both situations would represent toxic monopolies. However, those two would be the biggest North American companies to go after it. Likely the FTC would ensure these two companies get the opposing IP, giving Intel the freedom to make dGPUs without having to dance around both AMD's and Nvidia's GPU IP to do it, and giving Nvidia access to x86 like it wanted for the Denver Tegra. Getting AMD's Zen IP and then adding Denver on top would represent a huge issue for Intel and would also represent the first time in history a VLIW architecture didn't fall on its face in high performance needs situations. Denver doesn't actually execute ARM assembly. It reads ahead and creates VLIW instructions for the cores to execute. That's why its 2 cores outclass the A8 so readily but can't quite match the A8X. For Nvidia's 3rd try, that was damn good. Add that to what AMD has now and provide the x86 ubiquity along with a jump in clock speed, and Intel will have one nasty competitor in CPUs.

 

If You buy AMD and keep it alive long enough for Intel to knock Nvidia out of the accelerator market (while simultaneously eating up low-end and mid-end dGPU sales via its iGPUs) then Intel will not be in competition with Nvidia in any market, and the investors will be looking to make back the money lost in the process. Nvidia will be wide open to a buyout by Intel. At that point, AMD being financially weak would be an easy thing for Intel to crush outright, and no one would be willing to try to pick up the smoldering heap left behind. Intel will have won total dominance in everything HPC, Desktop, and mobile, leaving only phones and microcontrollers as markets it didn't dominate in computing.

 

One option guarantees competition. The other potentially dooms it in the long run.

Hmm.  I understood that Intel would acquire ATI, and nVidia the X86 in most situations, but I did not think that nVidia could actually truly challenge Intel to that degree.  I am sure it would take time, effort, and trial/error to get the D-GPU segment on Intels part running en mass, as well as ensuring they co-operate with developers to ensure they have zero-day drivers for programs and games.  But--it seems like from your point right now - nVidia would be some terrifying competition.  Hmm.

[patrickjp93]

Hmm.  I understood that Intel would acquire ATI, and nVidia the X86 in most situations, but I did not think that nVidia could actually truly challenge Intel to that degree.  I am sure it would take time, effort, and trial/error to get the D-GPU segment on Intels part running en mass, as well as ensuring they co-operate with developers to ensure they have zero-day drivers for programs and games.  But--it seems like from your point right now - nVidia would be some terrifying competition.  Hmm.

Intel has programming man power to spare. Giving a few a few more to drivers (not to mention Intel would likely pick up AMD's driver team to smooth over the transition) is of little consequence.

[HalGameGuru]

nVidia is already a problem in the streets with Denver. Emulators.com did a write up on the possibilities of the architecture and binary translation. The future could get very interesting very fast with new avenues of leveraging architectures across instruction sets. And I don't think AMD would go quite so quietly into that good night. THey hold too much IP and have shown too much promise in their upcoming designs. Between Apple sitting on 200B, Microsoft being schizophrenic right now, and so much uncertainty in the face of a resurgent Intel I think AMD has a lot of back room champions behind them, to at least keep the waters muddied.

[patrickjp93]

nVidia is already a problem in the streets with Denver. Emulators.com did a write up on the possibilities of the architecture and binary translation. The future could get very interesting very fast with new avenues of leveraging architectures across instruction sets. And I don't think AMD would go quite so quietly into that good night. THey hold too much IP and have shown too much promise in their upcoming designs. Between Apple sitting on 200B, Microsoft being schizophrenic right now, and so much uncertainty in the face of a resurgent Intel I think AMD has a lot of back room champions behind them, to at least keep the waters muddied.

Apple wouldn't care if AMD dropped dead.

[HalGameGuru]

Apple would LOVE to have more IP to sue people over.

[Doobeedoo]

That APU though...
But I can't wait to see what' they'll offer on desktop side, what will be the top tier. Cause I would like to get a 8-core with good clock speed.

[SurvivorNVL]

I really hope AMD gets the government contract, and things just go up from here.  We need competition.

[Fetzie]

If you want the nice outcome where Intel gets ATI and Nvidia gets x86_64, buy Nvidia now and don't look back. If you want to risk one nasty future where Intel gets control over Nvidia and can squash AMD like a bug without risk of a serious competitor, buy AMD now.

 

The second scenario would see Intel being forced to break up into competing companies by the US government.

[Prysin]

Everyone thought the Console deals were the big saviour of AMD - and surely, at that time they were - but that's going to be nothing compared to the profit AMD is likely to make if they can secure this.

 

AMD surely does make a profit on every APU sold to the console manufacturers, but it's really not that much. However, at the top-of-the-top end, they'll be able to add in much higher margins.

a bit off-topic.

AMD also secured a massive contract with several hollywood movie studios...

 

 

so they definetively got some good stuff going on in the enterprise market. If they nail this exacomputing project, they will probably be cutting down their debt notably

[patrickjp93]

The second scenario would see Intel being forced to break up into competing companies by the US government.

No it wouldn't. When you gain monopoly by raw competition, you're legal and free to do as you please as long as you don't abuse that position.

[BuckGup]

Oh that is sorta cool

[Deathroned]

Can it run crysis 140fps 8k?

[Fetzie]

Can it run crysis 140fps 8k?

 

If it's running ARM or PowerPC then no, because Crysis is an x86 application