How is Sony going to implement 8GBs GDDR5 in PS4? *spawn

[Arwin]

Let me turn another question around. For high performance jobs where the latency actually matters and that run on the CPU, wouldn't the Jaguar being an 8 Core processor also mean that performance critical work is more likely to be split over various paralel jobs, also making latency less important?

 

[dumbo11]

OOO CPUs are usually constructed with ROBs large enough to schedule around latencies comparable to a hit in LLC (last level cache). A LLC cache miss will stall *any* CPU core after 15-40 cycles.

Thanks for both explanations - makes things a lot clearer for me .

Looking around the web, it seems that jaguar was supposed to have a larger ROB, but I doubt that makes a lot of differences given the size of the gap.

 

[Gubbi]

Let me turn another question around. For high performance jobs where the latency actually matters and that run on the CPU, wouldn't the Jaguar being an 8 Core processor also mean that performance critical work is more likely to be split over various paralel jobs, also making latency less important?

With more cores, you can have more memory ops in flight in parallel. Each core is still stalled for longer. When stalled, a core can't issue new memory ops, so the longer latency directly influnces the utilization of the memory system. Orbis has a lot more main memory bandwidth available than Durango, but have a harder time utilizing it.

Cheers

 

[Arwin]

Thinking about that, you're probably right. Writing to RAM would save HDD thrashing, which constant recording could severely induce. I have one question though, and that's how a rolling 15 minutes of compressed video is structured? Once you've recorded 15 minutes, can you just shift the current memory address to the starting memory and just overwrite the beginning? How does the video codec handle that?.

So if it is 1GB or less that you need for 15 minutes, you can afford to have a 1GB of flash ram for this purpose, right? Some RAM buffer may also be necessary, but there may be enough in the compression chip ... And that then wouldn't take away from the HDD or anything else.

 

[Shifty Geezer]

Wouldn't flash RAM be subjected to burn-in? Then again, it'll only write a cell once every 15 minutes, so that really won't be a problem! So yeah, flash mem should suffice and a gig of RAM would probably be overkill.

 

[loekf]

OoOe CPUs can usually schedule around 10-15 ns of latency. DDR3 is between 50 and 100ns, GDDR5 more.

Cheers

N00b question perhaps...

If the latency penalty is that high and given the fact that the PS4 is not a real general purpose computer aka desktop and running most of the time a single application (game). Can't you apply tricks in the compiler (so scheduling) to make improve things ?

 

[Gubbi]

Can't you apply tricks in the compiler (so scheduling) to make improve things ?

The compiler can insert prefetch hints, which might give a bit.

Developers can make a bigger difference. Use data-structures that favor the memory system (wide fan-out trees, - or outright stream computing). They will effectively trade bandwidth for lower effective latency.

Cheers

 

[tunafish]

Good memory controllers with DDR3 can get you ~40-50ns. GDDR5 is 300-500ns. 8-10 times worse.

I don't think most of that is inherent in the ram. GDDR5 controllers in GPUs are very high-latency because sacrificing latency for bandwidth makes sense in GPUs, and the transfers are usually large.

 

[Arwin]

I found this on techspot:

The memory is also fundamentally set up specifically for the application it uses:
System memory (DDR3) benefits from low latency (tight timings) at the expense of bandwidth, GDDR5's case is the opposite. Timings for GDDR5 would seems unbelieveably slow in relation to DDR3, but the speed of VRAM is blazing fast in comparison with desktop RAM- this has resulted from the relative workloads that a CPU and GPU undertake. Latency isn't much of an issue with GPU's since their parallel nature allows them to move to other calculation when latency cycles cause a stall in the current workload/thread. The performance of a graphics card for instance is greatly affected (as a percentage) by altering the internal bandwidth, yet altering the external bandwidth (the PCI-Express bus, say lowering from x16 to x8 or x4 lanes) has a minimal effect. This is because there is a great deal of I/O (textures for examples) that get swapped in and out of VRAM continuously- the nature of a GPU is many parallel computations, whereas a CPU computes in a basically linear way.

 

[mczak]

Good memory controllers with DDR3 can get you ~40-50ns. GDDR5 is 300-500ns. 8-10 times worse.

Frankly I don't believe that for a second, I'd be quite surprised if it's more than a factor of 2.
Unfortunately I can't quite figure that out from datasheets, just about the only thing in there wrt to latencies are the published CAS latencies, and they look quite similar (at the same command clock). But of course there's much more to total latency.
With more bandwidth at your disposal you could also try more aggressive prefetching to counter longer latencies if they really were a big issue (but I doubt it's worth it).

 

[mczak]

Heck, are 4gbit GDDR5 chips even available?

Samsung and Hynix largest chips are both 2gbit chips with x32 interface.

If Sony needs to use custom packages, it's added cost to a pile of added cost.

Cheers

My guess is these will be brand-new but ordinary 4gbit chips, someone just promised they could deliver them by then.
Hence probably why most rumors were saying 4GB gddr5, they weren't available in time for developer systems.

 

[Kb-Smoker]

Seeing that dev kits have double the ram. Wonder if they are using a "dev kit" design to make the 8GB of work. I have seen no report of dev kits with over 8GB of ram and most dev didnt even know about the 8GB of ram.

Hmmmmm

 

[Xenus]

There were rumors of dev kits with 16GBs of RAM floating around

 

[MfA]

So the choice to go with 8 GB GDDR5 going to be too costly? Can this be a short term solution until a redesign comes along allowing Sony to use 2.5/3d stacking and achieve same bandwidth and latency?

GDDR5 is expensive because it's boutique ... using memory which is even moreso won't bring down costs.

 

[MfA]

Frankly I don't believe that for a second, I'd be quite surprised if it's more than a factor of 2.

500 ns at a 1 GHz clock would be 500 bloody cycles ... you can't even program the CAS delay that high ...

 

[mczak]

There were rumors of dev kits with 16GBs of RAM floating around

That is interesting, but if true it must have been ddr3 in this case.

 

[dumbo11]

Towards the end of last year, I guess people were looking at a high demand for GDDR5 in 2013:
- Win8 being released would push new system and GPU sales...
- AMDs 2013 lineup was poised to compete against NVIDIA's Titan in a glorious battle...

Instead the PC/laptop market is still getting hit hard by tablets, Win8 sunk, the world economy is still in a mess, a new console generation may reduce GPU sales, and AMD postponed their product lines...

If people have planned sensibly, then the GDDR5 market could be a complete nightmare... either AMDs partners or the GDDR5 producers would be left with an awful lot of useless chips/bad contracts/overcapacity?

I wonder if Sony were able to get this stuff "significantly cheaper than they planned".

 

[Brad Grenz]

So if it is 1GB or less that you need for 15 minutes, you can afford to have a 1GB of flash ram for this purpose, right? Some RAM buffer may also be necessary, but there may be enough in the compression chip ... And that then wouldn't take away from the HDD or anything else.

Writing 1MB per second to the hard drive is hardly a problem worth engineering around. If any smartphone or point and shoot camera can manage to effortlessly encode and save 1080p videos with their meager RAM and storage speeds, PS4 won't even feel it.

 

[TheWretched]

Writing 1MB per second to the hard drive is hardly a problem worth engineering around. If any smartphone or point and shoot camera can manage to effortlessly encode and save 1080p videos with their meager RAM and storage speeds, PS4 won't even feel it.

To a degree, PS3 doesn't even feel it. Play TV and Torne already give you the ability to DVR live TV while gaming. Well, it's not HD footage, but (from what I remember) broadcast TV sends at rather high bitrates (because they are not using the latest codes and just at a baseline encode, quality would suffer if they went lower).

 

[Wynix]

This is my first time posting but i've been an occasional lurker for a while now; anyway i just wanted to clear the uncertainty surrounding the density of the GDDR5 using a PDF from SK Hynix located here:

https://www.skhynix.com/products/support/databook.jsp

Q1 2013 for 4Gb density GDDR5, no doubt sony will be using this density.