[tostitos1979]

What you folks are saying is mostly true. VR and 4K are the drivers for a whole new compute cycle in the home. I think the current core counts for Xeon are good enough for cloud (by this I mean that I think CPU isn't the bottleneck for the average EC2, Google Compute, Azure instance). Any one care to comment?

[noir_lord]

In the sense of VR/4K the CPU is way less important than the GPU and we still have some headroom on those even with existing process, what will be interesting is when everyone else's process catches up with where Intel is now, they've generally stayed out in front of everyone else for a long time (except AMD for a spell).

I'm really looking forwards to VR if it catches on though, having an insanely high resolution headset so I can dump multiple monitors for programming is a big win, combine that with something that has the portability/form factor of a MS Book/Macbook Pro and you'd be able to program as capably from a hotel room as at your desk at home/work.

That would be the biggest shift in my work habits since I went from Windows to Linux in the late 90's.

Also I think once everyone can get down to the same size as intel we might start seeing more exotic architectures, Intel has often won with the "with enough thrust a brick will fly" approach to engineering, it doesn't matter if your chip is clock for clock more efficient if Intel is operating at a level where they can put 5 times as many transistors down in the same unit area and ramp the clock speed way up.

[tostitos1979]

Good point. Nvidia seems like the exciting company in this regard. I have a bad feeling they're going to tumble because the latest announcements related to Pascal have focused on deep learning instead of 4K for consumers. As far as I know, they haven't even announced their consumer Pascal cards yet. The rumor I've read online is announcements in June.

[mac01021]

What would it mean for Nvidia to focus on 4k? Do the new screen resolutions require architectural innovations in the GPU? (Honest question. I'm not a graphics engineer)

[noir_lord]

Me either so take this with a pinch of salt.

Mostly it's about shuttling around 4 times as much memory for each screen as well as 4 times as much processing.

1920x1080 has ~ 2 million pixels. 3840x2160 has ~ 8 million pixels.

Internally iirc this is often done as vectors before been rasterised out and having various filters and shaders applied but that step requires that you store multiple buffers etc, same reason a card that will play a game just comfortably at 1024x768 will run away crying at 1920x1080 I guess.

[mac01021]

I can see how higher resolutions will require the GPU to have more memory or more FLOPS or both. I guess my question is, how, if at all, are the improvements required to support 4K different from improvements that target something like deep learning performance?

[digi_owl]

Not quite about that last part. Clock rates have been more or less stationary over the last decade or so. What Intel has been focusing on for some time is cache, and how to practically always have the right data in the cache at the right time. but yes, having more transistors to work with do allow them to pack more cache onto the die.

[noir_lord]

Yeah that wasn't perhaps very clear, I meant historically they ramped the clock up (back in the P4 days).

[digi_owl]

And ran into a brick wall, iirc.

[eru]

We still need more CPU in the cloud. There's an almost unlimited appetite; and we still often look for optimizations to our program that reduce CPU load. (At least at Google. Not sure how much I am allowed to say in more detail, though..)

[tostitos1979]

Touche. I've also been involved in building clouds but ours was focused on customers building web-apps. Those are probably not compute bound. I can see why data processing workloads would need more CPU.

[justin66]

> VR and 4K are the drivers for a whole new compute cycle in the home.

My i7 desktop will be four years old in June. I've done the research and all I need to run the Oculus or HTC Vive (or the upcoming games that look good) is a graphics card upgrade.

[z3t4]

As a software developer I constantly find myself optimizing and scaling across CPU's. Sure those programs I optimized 10 years ago run blazingly fast now. But as computers got faster the possibilities and expectations also increased.

I would like to compare CPU's with roads: Roads increase traffic, not the other way around. So if you make a faster CPU, the demand for a faster CPU increases.

So if Intel stops making faster CPU's, the demand will (unexpectedly) go down.

[codeonfire]

Good point except facebook is targeting 6k per eye. People are definitely going to buy hardware, but maybe not in huge numbers.