from what I see it is a PC a MS branded kinect controlled windows 8 PC for the living room.
Predict: The Next Generation Console Tech
- Thread starter Acert93
- Start date
- Status
from what I see it is a PC a MS branded kinect controlled windows 8 PC for the living room.
Which would put in line with what we're seeing with Microsoft Surface devices. And now that they are unifying the Windows Phone 8 platform with Windows 8, I could see them doing the same with 720. There will be one architecture(Windows 8) for all devices: desktop, tablet, phone & living room.
Tommy McClain
Bulldozer uses AVX, not VMX. However it only packs 2 units which means at 3Ghz it's only a little over half the theoratical floating point performance of the 7 SPU's in PS3.
We're talking Steamroller cores though if I'm not mistaken? So if Steamroller uses AVX2 in the PS4 then it would exceed the SPU's in theoretical floating point performance.
I don't think they could make the simultaneous claim of being 10x the power of the PPE + more powerful than the SPU's at the same time though.
Last edited by a moderator:
I'm wondering how much benefit a huge amount of memory would really be. I've got a 670 at 680 performance which is quite likely going to be more powerful than either next gen console and yet it only has 2GB of memory. I've got a "meager" 4GB of system RAM but I hardly ever get over 75% full. I don't really understand why PC's come with 8 and 16GB these days since I can't imagine any gaming situation were you would use anywhere near that amount.
If the consoles come with 8GB then I find myself in the strange situation of having a PC with both a more powerful CPU and a more powerful GPU than either console, yet being severely memory limited in comparison. Surely a first!
Even if they come with 4GB unified the 2GB on my 670 could still be a contstraint event though it could be packing near double the actual performance of the console GPU's.
I'm guessing then the only reason PC GPU's haven't increased memory size in line with processing power is because they are constrained by console game engines which target a much smaller footprint??
If the consoles come with 8GB then I find myself in the strange situation of having a PC with both a more powerful CPU and a more powerful GPU than either console, yet being severely memory limited in comparison. Surely a first!
Even if they come with 4GB unified the 2GB on my 670 could still be a contstraint event though it could be packing near double the actual performance of the console GPU's.
I'm guessing then the only reason PC GPU's haven't increased memory size in line with processing power is because they are constrained by console game engines which target a much smaller footprint??
On PC's the 4GB limit is a function of almost every app being a 32 bit binary, many don't ever set the executable bit to allow access to more than 2GB.
A theoretical 8GB Console wouldn't have that issue.
There are plenty of ways to use lots of memory if you can guarantee it's going to be there.
The bigger issue to me is getting data off the storage and into that memory.
A theoretical 8GB Console wouldn't have that issue.
There are plenty of ways to use lots of memory if you can guarantee it's going to be there.
The bigger issue to me is getting data off the storage and into that memory.
I'm not concerned about BC.
I was thinking about CELL in terms of something internal that would make the next PS harder to develop for again.
french toast
Veteran
I have to say we agree on something
I have no idea the speed differences between gddr5 and ddr 4...nor the costs..but that might be the cheapest way to get 4gb with nice bandwidth and a slim bus.No way we are getting 8gb Ian console.
I have a question, why bother with simd engines on cpus?? Wouldn't it be far better to allocate that space to the gpu?
GPUs are only good for workloads that are massively parallel (tens of thousands of threads). GPU calculations take long time to start and long time to finish (long latency because of asynchronous execution and deep GPU pipeline). Basically you have to wait for millions of cycles before you get your results (back to CPU). Basically GPUs are only useful for large scale batch processing (but are very good at it).
In contrast, with CPU SIMD instructions you get your results ready in less than twenty cycles, and you do not need much parallelism (four lanes is enough, eight with AVX for peak performance).
Also GPU cannot feed itself. CPU must setup and start GPU execution. Setup phases often require vector math, and SIMD is perfect for this. New DX11 GPUs have additional (but limited) ability to control themselves. For example you have drawAuto that utilizes GPU stream out results to determine size of the next calculation (no CPU intervention needed) and now GPUs can stream out to constant buffers (to change static flow control constants of next stage calculation), but that's pretty much it. Forthcoming Tesla K20 boards (based on compute oriented Kepler design) can finally issue new kernels themselves, but so far there's not much documentation about that feature. It will be really interesting to see how efficient that feature is, and how soon it becomes commonplace in consumer products. It would be great to do visibility determination and draw call setup fully inside GPU (these tasks are commonly done on CPU and are tasks that benefit much from CPU SIMD instructions).
french toast
Veteran
@sebbi cheers. so having the 128bit simd engines on xenon was worth it for gaming then?
Would having 8 low clicked OoO cores with 256bit simds attached be a good idea?, or rather having a higher clocked quad with simds and the rest allocated to the gpu?
Would having 8 low clicked OoO cores with 256bit simds attached be a good idea?, or rather having a higher clocked quad with simds and the rest allocated to the gpu?
the more I think about it the more I believe That jaguar cores are unlikely.
Bobcat cores suck as Atom do, even significant improvments from Bobcat to Jaguar may not get the perfs where MS/Sony wants.
Using TSMC process less than 400 sq.mm of silicon would be enough for a four module / 8 cores CPU based on Stream roller arch and the R1100 equivalent of something in between Cap verde and Pitcairn (think 16 CUs and a 128 bit bus).
They may even pass on EDram and go with 2 128 bit buses. Late GPU have reach pretty high speed for the VRAM even a 128bit bus would provide lot of bandwidth.
I could see 4GB of DDR3 or DDR4 (or more) and as few as possible fast GDDR5 (I don't know what the minimum would be on a 128 bit bus using the hight speed GDDR5 memory module).
The issue with such set up is power consumption, I could see it a bit too high for an always on device.
Bobcat cores suck as Atom do, even significant improvments from Bobcat to Jaguar may not get the perfs where MS/Sony wants.
Using TSMC process less than 400 sq.mm of silicon would be enough for a four module / 8 cores CPU based on Stream roller arch and the R1100 equivalent of something in between Cap verde and Pitcairn (think 16 CUs and a 128 bit bus).
They may even pass on EDram and go with 2 128 bit buses. Late GPU have reach pretty high speed for the VRAM even a 128bit bus would provide lot of bandwidth.
I could see 4GB of DDR3 or DDR4 (or more) and as few as possible fast GDDR5 (I don't know what the minimum would be on a 128 bit bus using the hight speed GDDR5 memory module).
The issue with such set up is power consumption, I could see it a bit too high for an always on device.
Last edited by a moderator:
anexanhume
Veteran
Neither do I (I think it's a ludicrous number), but I'm just repeating other rumors. I figure 4GB max. 128bit vs. 256 bit seems to be based simply on whether it's Pitcairn or Cape Verde based. I'm surprised there's little talk of Sea Islands influence in next gen console GPUs, especially given AMD's insistence on HSA.
Wouldn't an AMD driven design of CPU, GPU and ARM CPU for low power set-top tasks be the ultimate all-in-one solution for MS?
I agree 8GB seems pretty unrealistic. 16x the current generation is too high given the rest of the rumoured specs. 8x seems reasonable though so I'm betting on 4GB of reasonably fast GDDR5 + some edram. 2GB of very fast GDDR5 isn't out of the question though, possibly still with edram.
anexanhume
Veteran
I don't care what type it is. I think 8GB is bonkers.
The SIMD engines on the xbox 360 are heavily used. I've seen them used a lot by audio engines for effects processing, and they're also used a lot by physics engines.
Ditto I really don't see the need to keep an entire dl DVDs worth ofdaya in main ram.
The only reason for me would be if the rumours are bullshit and MS really have built a monster. Otherwise the system would seem totally unbalanced which I really can't see them doing.
And I am also totally convinced there won't be a straight forward AMD CPU GPU or APU, that doesn't make any sense to me. You have a clean slate every 5-8 years to really move things forward in a closed platform. I am sure there will be some creative thinking going into it.
The only reason for me would be if the rumours are bullshit and MS really have built a monster. Otherwise the system would seem totally unbalanced which I really can't see them doing.
And I am also totally convinced there won't be a straight forward AMD CPU GPU or APU, that doesn't make any sense to me. You have a clean slate every 5-8 years to really move things forward in a closed platform. I am sure there will be some creative thinking going into it.
8gb ddr3 might cost the same as 2gb ddr5.
Per the 8GB being unusable, insane, etc I posted this about 2 weeks ago:
TLDR Version: If you are going to use an optical drive you are going to have crazy long load times. SSD are expensive, a flash cash could be limited in size/need to be written to often/reliability, HDDs are marginally faster...
Best solution is a ton of RAM. Load once (mask your initial load time) and you get the best of all words in terms of latency and speed once the data is cached into system memory.
I had deposited in the Alt. memory/storage thread this, to summarize:
Optical Media: (a) Size: Large (25-50GB), (b) Cost: Media is cheap, Drives are $40-$50, (c) Transfer: slow, in the tends of MB/s, (d) Latency: very poor at >100ms, (e) Other: They tend to be loud at high speeds and have reliability issues. Getting content from an optical disk into a game is the pits.
Hard Disk Drives: (a) Size: Large (250-750GB), (b) Cost: Drives have jumped into the > $50 wholesale range, game distribution only works with DLC, (c) Transfer: slow, but better than Optical, as it nears about 100 MB/s for best case scenarios, (d) Latency: poor at ~ 10ms, (e) Other: While HDDs cut down on transfer speed to memory they are still not exceptionally fast or quick. Serviceable although the idea of populating 4GB of memory from a HDDs averaging 20-50MB/s is still cringe worthy (80 seconds to fully populate at 50MB/s... yes, gameplay starting before load is possible but so are fragmented loads in the 10s of MB/s, too).
SSDs: (a) Size: Medium (60-128GB for console budgets), (b) Cost: Drives are expensive running up toward $100 but with some potential for cost reduction, game distribution only works with DLC, (c) Transfer: Good, in the hundreds of MB/s for best case scenarios, (d) Latency: Good at ~ <0.1ms, (e) Other: Some hurdles, especially in durability, in future reductions. SSDs are fast but they have limited size (100GB would seem small with caching for games on 50GB disks) so you are stuck re-loading content.
So what are we to do? The industry is at a rough inflection point. The future looks to be large SSD's with DLC, but that future is not today.
I had suggested a cheaper solution: A ton of RAM. I even specifically recommended 8GB.
The idea that you could stream from your Optical Drive or HDD (or both) to a large RAM pool was this: One long load time, but enough memory that entire segments of games could be buffered into memory. So lets say each level is 2GB. You could start level 1 (2GB) and by the time you are done level 2 (2GB), level 3 (2GB), and level 4 (2GB) could already be sitting into memory. Cut it however you want (MP and SP and the next level loaded, whatever).
With 8GB of memory (a) Size: 8GB is large for memory so close to the processors; not good for long term storage but amazing for application use , (b) Cost: RAM has dropped through the floor, I have seen 4GB going for $25 and 8GB for under $50, (c) Transfer: Great, > 20GB/s, (d) Latency: Great in the ns range.
The only real issue is getting content into the RAM. But you are stuck with optical for distribution and optical and HDDs for storage anyways and a SSD, while a nice upgrade, is on the order of a magnitude SLOWER in bandwidth and latency than RAM. Instead of spending > $100 for a SSD why not spend an extra $25 to move from 4GB to 8GB memory and go with TSRs that state: gameplay must start within X seconds and then focus on content buffering. The idea you could access up to 8GB of content within a frame or two is pretty amazing.
That, right there, is why I think 8GB is a strong chance to being a real rumor. It is the cheapest trade off compared to a SSD, it addresses caching, buffering, and loading on a lot of levels (except initial load), and the performance benefit is WAY bigger than all the competing technologies. Getting bandwidth > 40x over a class leading SSD (and it gets only higher if the system bandwidth is over 20GB/s for the 8GB) and easily >200x over an optical drive (realistically nearing >1000) and the latency is a joke (e.g. RAM with 6ns latency = .000006ms; put another way 100ms optical drive latency is 100000000ns; RAM, not can be in the tens of ns of latency so 1x10^6 better if I did my math right).
Bring on the 8GB consoles
TLDR Version: If you are going to use an optical drive you are going to have crazy long load times. SSD are expensive, a flash cash could be limited in size/need to be written to often/reliability, HDDs are marginally faster...
Best solution is a ton of RAM. Load once (mask your initial load time) and you get the best of all words in terms of latency and speed once the data is cached into system memory.
aaronspink
Veteran
Must not run many modern games.
Civ5 will easily hit 2GB. Same with BF3 on larger maps.
Pretty much. Also the PC GPU's VRAM size has more to do with PCI-E bandwidth/latency than anything else.
- Status
Similar threads
- Locked
