| JustBeingReal said: 1)No DDR3 cannot run faster than GDDR5, not without running into issues with incredibly high voltage. DDR3 just isn't designed to operate at the kind of speeds that GDDR5 does. |
You don't need high voltages.
There is also MORE to bandwidth than just the amount of Mhz that something operates at.
| JustBeingReal said: I didn't say 12GBs of memory wasn't possible, I said it doesn't make sense to use 12GBs of DDR3, alone. |
There are plausable reasons why it actually does make sense.
| JustBeingReal said: Newer GPU tech actually requires more bandwidth, which is why companies have invented GDDR5X and HBM, you aren't going anywhere without any more data actually getting to the GPU for it to actually work on. Fast memory is essential. |
I know this. I am not attempting to refute that.
| JustBeingReal said: 384 Bit DDR3 doesn't exist. The biggest bus width available is 256 Bit, unless you clock it higher than 2133Mhz, you're not getting higher than 68GB/s theoretical performance, as you clock it higher you run into latency problems. |
It doesn't need to exist.
384bit memory is a memory configuration, you could have 1024bit DDR3 if you wanted, only the memory controller needs to support it, it's transparent in every other aspect.
And for the record... There is no limit to the bus width. Only what is economically feasible.
And no. You don't run into latency problems.
Do you have any idea how latency on RAM is even calculated? CAS Latency is latency per cycle.
You could have two sticks of RAM, one running at 2133mhz, the other running at 3000mhz, both for example are CAS 10. - Guess which actually has the lower latency? The higher clocked RAM, funny how that works isn't it? The more you know.
| JustBeingReal said: 2400Mhz DDR3 on a 256 Bit bus would only offer 76.5GB/s, which is pitiful for a GPU 2X stronger than XB1's, hell 68GB/s wasn't enough for XB1, which is why Microsoft added 32MBs of eSRAM and there are still bandwidth problems. This hypothetical NX potentially has a 2.7TFlop GPU, which is more than 2X the GPU performance of XB1, 2X the bandwidth would be the minimum needed compared to XB1's GPU. |
Bandwidth doesn't work like that.
You don't need 2x the bandwidth because of colour compression.
Secondly... These consoles are only topping out at 1080P, you don't need HBM levels of bandwidth because it would be wasted.
| JustBeingReal said: You're not going up to 3000Mhz on DDR3, the voltage problems would be ridiculous, better to just go with GDDR5, which can have a 384 Bit bus and latency would be much lower than clocking DDR3 through the roof, hell even DDR4 would have more latency problems than GDDR5 would, 4GBs of HBM would be even better, especially for a 2.7TFlop GPU. |
I am going to assume you are ignorant about PC hardware.
I had a Samsung Ram kit a few years back that would reliably hit 3000mhz @ 1.5v. I would assume DDR3 fabbed at 28nm or lower would fare even better than that today.
You keep throwing the latency "issue" around. When there isn't one, I don't think you truly understand what it actually means.
| JustBeingReal said: No DDR3 can't provide more bandwidth than GDDR5, not even if you had 2 separate 256 Bit buses, each with a 6GB pool clocked at 2400Mhz, PS4 would still outperform it for bandwidth, with 176GB/s theoretical vs 153GB/s for NX. NX would have a lot of latency issues clocking it's memory that high, considering that the timings for PS4's memory are comparable to 2133Mhz DDR3, if you go higher then latency gets longer. |
If you had 153GB/s on the NX and the NX had a more modern GCN core... Then that 153GB/s would be of bigger benefit than the PS4's 176GB/s.
Again. Memory latency is in terms of clock cycles.
AMD for example... Released Tonga, cut down the memory interface from 384bit to 256bit, increased the memory clock but with a reduction in total bandwidth of 240GB/s down to 176GB/s. Yet... Thanks to efficiency gains Tonga managed to beat Tahiti Pro. Despite it having less than 64GB/s of bandwidth.
| JustBeingReal said:
You can't hide bandwidth deficits, latency sure, but bandwidth is literally like the fuel of a car, without data the GPU has nothing to work with, compression means you loose data, better to have high bandwidth, hell HBM would be a better option in part, because it has lower latency. Part HBM, part GDDR5 would probably be best if you can't get all HBM. |
You can hide bandwidth deficits.
Why do you think so many consoles have used ESRAM, EDRAM and other derivites? Why do you think Intel has it's own for it's Iris Pro? It hides bandwidth deficits.
Not only that... But with more efficient scheduling and prediction you can move bandwidth-heavy data when your workset isn't bandwidth heavy allowing for greater levels of utliziation.
And no. Compression doesn't mean you loose data.
Lossless compression is the act of reconstructing compressed data which is identical to the original.
HBM won't happen in the NX. It is far too expensive, consoles are cost sensitive remember? At the moment HBM is reserved for Graphics Processors that cost 2-3x that of a console. To put simply... Console gamers aren't rich enough to afford it.
| JustBeingReal said:
This leak is as fake as they come. |
I don't disagree. But the points which you are using to claim it as inplausable, simply aren't.
--::{PC Gaming Master Race}::--
