Hacker News Reader

Comment by gwern | Hacker News Reader

> With 8Gb per DRAM die and 2 Gbps speed per pin, we get approximately 256 GB/s bandwidth per HBM2 stack. With four stacks in total, we will get 1 TB/s bandwidth on NVIDIA’s GP100 flagship Pascal which is twice compared to the 512 GB/s on AMD’s Fiji cards and three times that of the 980 Ti’s 334GB/s.

> The Pascal GPU would also introduce NVLINK which is the next generation Unified Virtual Memory link with Gen 2.0 Cache coherency features and 5 – 12 times the bandwidth of a regular PCIe connection. This will solve many of the bandwidth issues that high performance GPUs currently face.

To point out the obvious, this sounds like it could be fantastic for deep learning. Not just is the RAM big enough to hold a lot of current datasets, it'll help alleviate the latency bottlenecks in updating stuff.

View on HN

Yup, this is exactly the marketing angle Nvidia CEO was using in his GPU Technology Conference presentation (10x speed-up on deep-learning vs their current Maxwell architecture):

http://blogs.nvidia.com/blog/2015/03/17/pascal/

This also seems to be the motivation behind the already available Titan X with its 12GB of RAM. Some of the gaming hardware reviewers are scratching their heads as to why anyone would need 12GB attached to one die. I chuckled when I saw those reviews that were totally oblivious to the deep learning applications.

It's easy to get lost in the noise of the "enthusiast" webring with their semiliterate "benchmarks" and overly paginated "reviews". They punch well above their actual purchasing weight and thus think they're a lot more important then they actually are.

Versus the data scientists and HPC communities that buy hundreds of Nvidia GPUs for CUDA and associated applications.

Well right, but compared to the number of units bought by IT departments, the "enthusiast" market is nothing but decimal dust.

IT departments are not buying Titans and they are not going to be buying Pascals.

IT departments aren't even buying Nvidia at all, they are buying Intel integrated.

That was my point. If all you do is read what passes for news on gamer sites, you might think that the discrete GPU overclocking market was important.

The notion that only the biggest of something is "important" is offensive and wrong.

Sure. I'm a fan of esoteric hardware and software and no big believer in the market success equalling virtue religious ideology, but I'm also not the one invested in the importance of how well this memory stick performs overclocked in Gamer Game IV: Games.

Probably, but remember that if you're maximizing profit then one $100 margin customer is as important as 1000 $0.10 margin customers.

The 1000 low-margin customers also come with the added benefit of diversification.

Not to mention margin decreases linearly not as a percentage. So those 1000 $0.10 margin customers can easily turn into -$0.50 margin customers.

No, you are just totally oblivious to Nvidia's product segments.

The deep learning applications and those that want massive amount of RAM buy Teslas, which have exactly that. Titan X is a gaming card and it's marketed as such.

We do SP GPU development requiring big memories and Titan X features heavily in nVidia's marketing towards us. Tesla is nowhere to be found.

The community as a whole is completely oblivious. It's pretty funny to see the youtube reviewers get all worked up over how nvidia and amd are going at it again and such. As if gaming is what's driving this battle. Anyone working in computer science knows the battle is over machine learning, not first person shooters.

Soon headless, socketed solutions will be the preferred form factor for HPC. I image the desktop and server product lines will diverge at that point. It'll be curious to see what will happening to PC gaming at that point.

The gaming hardware junky community is aware, if only as a result of most Titan Black/Z reviews arriving at the conclusion, "a gamer's money is better spent elsewhere, as all a Titan really provides over the latest consumer offerings is good workstation performance (double precision floating point)."

For reference: Titan Z DP FP: 2707 GFLOPS ATI 5870 (released sept 2009): 544 GFLOPS Titan X (current generation - NVidia recognized that gamers don't value DP performance): 192 GFLOPS

I sometimes wonder how it came to be that the Titan Z preceded the Titan X...

E: to clarify, review sites are aware that STEM people need double precision and 6/12GiB GPU memory for something

One note on Nvlink: To get that you will AFAIK have to buy a Power8 system. They would need a cooperation with Intel to get support for this on their architecture - and Intel so far likes to keep them on a slow bus.

Yes.. with the current deep dreaming fad, this could attract quite a few hobbyists eager to get near-real time video processing through large nn.

The current deep dreaming fad is going to last about one more week before everyone gets sick of looking at those stupid pictures. They really serve no purpose other than "look how crazy this stuff is" - deep learning is cray-cray.

It's not over until it shows up in a few mainstream media pieces.

Maybe some music videos, or the intro of an HBO show..

Until then, enjoy your eyes and dog faces on family pictures.

any thoughts on why all the current pix look like dog faces? Is that peculiar to the training set?

Also I think there's a bit more than just ZOMGtrippy. The whole project points out that "if you're looking for it, you'll find it." The system amplifies noise until it sees something familiar. Understanding that humans do this too (to a point), is a critical point in comprehension and discussion of the world, opinions and people. But yeah, it will seem rather "iconic" and dated 10 years from now.

The reason you mostly see dog faces is only because everyone seems to be running the default code google posted. It optimizes layer 4c, which seems to be biased to producing dog faces. Running the code on other layers, or other networks produces other things. They also recently posted code where you can provide a "guide" image to guide it to producing features that are in that image.

I'm thinking that using CNNs to generate or change images could be a huge game changing technology. It could potentially be bigger for hollywood than 3d graphics. In 10 years, we could see this stuff everywhere, albeit in a more developed state. It would be used to actually transform or enhance images in a more directed way.

Reply on news.ycombinator.com