Predict: The Next Generation Console Tech

8GB along with Cape Verde seems so odd. They tend to skimp on RAM, yet 8GB is a hefty amount even by today's standards. Perhaps there is some truth to the dual but not xfire GPUs rumor back in April?

 

Who said he was in a high position, almost any senior PM at MS could have created/given that presentation. Presentations like this are a dime a dozen at MS unless you know the context of why it was given it's hard to weigh the value of the content.

FWIW I also don't like to equate individual competence with project success they are rarely related on large teams. There were probably a lot of smart people on the Zune project, I know there were a LOT of smart people on WinFS and that more or less managed to kill Vista as it was back then.

 

There are rather major limitations though. The shared data has to be very small in order to fit into core-local memory. The AI characters can do independent decision making only as long as their decisions do not depend on what any other AI is doing. Not to mention that AI is fundamentally about decision making, i.e. branching, which generally wreaks havok with very parallel architectures with their small local memories, long pipelines and typically light weight branch prediction/handling hardware.

So AI parallelizes nicely as long as you want to do comparatively trivial stuff on small data sets. As usual.
That's not to say it's useless. I'm just pointing out some limitations for the benefit of those who do not have much personal experience with parallel codes.

 

AI cooperating at 60 Hz time steps is plenty fast enough (which from a simulation point of view means they are completely independent). Unless you are trying to simulate nearly instantaneously communicating hunter killer robots. The edge case instabilities with timestep simulation in things like routing are an opportunity to improve your AI model and get more realistic behaviour (anyone who has been in traffic knows that humans have edge case instabilities in routing as well).

 

Pick a lane already dammit!!!

 

Can you give me a real life example where you get stuck? Large shared data is similar to a framebuffer, but that doesn't stop it being processed in parallel. You can have one step in the process isolate what AI is close enough to what other AI to group them together for chunking, you can bit-flag lots of state and decisions on AI and run them past decision logic, and distribute the data according to the key flags that were set. There's not a lot you cannot do - most of us just aren't used to it.

 

Wouldn't the NP nature of round robin style individual communications through all pairs of AI be the big show stopper, not how often they need to communicate?

 

I'll make an observation here, I would guess that even if you were 100% utilizing the output of your GPU to render a scene something like 20 to 60% of the ALU flops would be going unused.
ALU's aren't used at all when you;re rendering shadow, they are grossly underutilized when rendering post effects. If you do a deferred renderer then when you lay down the initial pass they are underutilized.

It's hard to say how much compute you could do without impacting rendering, and a lot depends on the mix of texture units, ROPS etc to ALU's, but I would guess it's a none trivial amount.

 

Well I don't work on a game team anymore, so you should take my opinion with a huge grain of salt.

Compute certainly has it's place, it's actually hard to write performant none trivial compute jobs.
Part of that today is tools, without a enough information to determine why your compute job is running slowly there is a lot of guess work involved.

I think next gen challenges are going to be a lot like this gens were, it's as much about dealing with team size growth and production issues as it is technology. What's the right balance for geometry, how complex should you're shaders be etc etc.
3d graphics algorithms will continue to move forwards and I think you'll see a big step forwards as the generation progresses as a result. I think compute will be a big part of that.

 

Obviously we are still working on early, vague details, But based on those details my non-fictional take looks like this.

Developers: "Give us more power."

MS: "We'll double the memory like last time."

Developers: "Not good enough."

MS: "Deal with it."

 

Devs asks for more ram, DICE want 8gb for example.

 

I'll take fancy shaders and lighting over texture resolution and loading times any day of the week. Someone should contact that GAF guy who knew Wii U specs and tell him to give a closer look because meltdown is about to happen in near future

In the end it could end up like this...

Developers : Give us more power MS!

MS : Here is 8 gigs of cheapest RAM available.

Developers : Huh...Its cool I guess. Give us more graphics processing power.

MS : NO! Here is 1 TFLOP and deal with it.

Developers : Sigh...Ok. Here is screen tearing, frame dropping sub hd multiplatform game for you. Deal with it!

 

Interesting the consumer doesn't seem to take part in those dialogues

That being said, SONY won't need more than 2GB of RAM if they're really going for cloud gaming in the long run.

All they need is a well-priced system that has enough power to sustain "traditional" console gaming for another few years - and is ready to be gradually integrated into next-gen cloud gaming stuff.

As far as I'm concerned, it's also time for more variable SKUs. Like having a "base model" without Bluray drive (but with a few USB3 ports for later upgrades).

 

This is along my thinking and something almost no one ever considers when totting up the specs, you don't need 10 times the units to get 10 times the power of a 8 year old console.

 

Interesting how people are so confident that they already "know the specs"

 

if the hd 4770 is the gpu inside the Wiiu, why Nintendo talk about only 1,5 times the raw power of the current gen console ?