Next Generation Hardware Speculation with a Technical Spin [2018]
- Thread starter Tkumpathenurpahl
- Start date
-
- Tags
- 5 bytes
- Status
Rikimaru
Veteran
Maybe, but still much more expensive than HDD.
Yeah thinking more along the lines of a 128GB scratch pad.
Everyone is already mass producing them (could already be on markets too, didn't check), so obviously yes.
Samsung starts MASS Producing QLC/4Bit SSD -- https://news.samsung.com/global/sam...duction-of-industrys-first-4-bit-consumer-ssd
However, Samsung’s 4-bit 4TB QLC SATA SSD maintains its performance levels at the same level as a 3-bit SSD, by using a 3-bit SSD controller and TurboWrite technology, while increasing drive capacity through the use of 32 chips, all based on 64-layer fourth-generation 1Tb V-NAND.*
The 4-bit QLC SSD enables a sequential read speed of 540 MB/s and a sequential write speed of 520 MB/s, and comes with a three-year warranty.
Samsung plans to introduce several 4-bit consumer SSDs later this year with 1TB, 2TB, and 4TB capacities in the widely used 2.5-inch form factor.
However, Samsung’s 4-bit 4TB QLC SATA SSD maintains its performance levels at the same level as a 3-bit SSD, by using a 3-bit SSD controller and TurboWrite technology, while increasing drive capacity through the use of 32 chips, all based on 64-layer fourth-generation 1Tb V-NAND.*
The 4-bit QLC SSD enables a sequential read speed of 540 MB/s and a sequential write speed of 520 MB/s, and comes with a three-year warranty.
Samsung plans to introduce several 4-bit consumer SSDs later this year with 1TB, 2TB, and 4TB capacities in the widely used 2.5-inch form factor.
D
Deleted member 13524
Guest
4-bit QLC will seemingly be used for very large SSDs that run at SATA speeds, not for speed-oriented solutions.
The fastest NVMe drives do tend to be expensive because of the cell parallelism required to achieve the fastest throughput, i.e. only the 512GB+ models can approach the PCIe 3.0 4x saturation at the moment.
However, YMTC seems to have a solution for that:
The fastest NVMe drives do tend to be expensive because of the cell parallelism required to achieve the fastest throughput, i.e. only the 512GB+ models can approach the PCIe 3.0 4x saturation at the moment.
However, YMTC seems to have a solution for that:
More or less what we will get next gen only on a 7nm process.
The presence of a fast flash drive setting precedent is a good thing. Price seems fairly steep ($625 equivalent) but I guess that includes Windows license. I also wonder if there's a decent markup because this machine, one assumes given the market, will be playing lots of games that the hardware company gets no cut from. So I think it's more like a Steam PC in terms of price rather than console.
<$500 next-gen with NVMe would make me happy.
<$500 next-gen with NVMe would make me happy.
D
Deleted member 13524
Guest
Considering the fact that TSMC's 7nm brings a 70% area reduction and 30% Performance or 60% Power reduction compared to 16FF+, then a 400mm^2 chip at 7nm should get us:
24 CUs / (1-0.7) = ~80 CUs
1.3GHz x (1+0.3) = ~1.69 GHz (if ISO Power)
4 Zen / (1-0.7) = ~12 Zen cores
3 GHz x (1+0.3) = 3.9 GHz
That would result in a 17.3 TFLOPs GPU and 12 Zen cores at 3.9GHz.
Realistically, I think it's safe-ish to assume the next-gens will be able to boast >13 TFLOPs GPUs and at least 8 Zen cores at ~3.5GHz, as long as the SoC size remains at 350-400mm^2 and they don't spend tons of die area on eDRAM.
Extrapolating a memory subsystem with enough bandwidth to feed such a SoC and bring a significant capacity upgrade at a reasonable cost is a much harder task, IMHO.
I just hope the current RAM pricing won't force the console makers to implement on-chip eDRAM, unless they're willing to increase die size by a substantial amount.
D
Deleted member 13524
Guest
Then 7nm consumes 3.33x more power per mm^2 than 16FF+, considering similar architectures and clocks?
Then we should make the calculations at 60% reduction in power at ISO clocks.
24 CUs / (1-0.6) = 60 CUs at 1.3 GHz
4 Zen cores / (1-0.6) = 10 Zen cores at 3GHz -> more likely 8 cores at over 3GHz.
60 CUs @ 1.3 GHz (or 64 CUs @ 1.25GHz) = ~10 TFLOPs.
This doesn't sound excellent, to be honest. It would also result in a ~300mm^2 chip.
anexanhume
Veteran
I don’t think we should assume anything over 3.0 for CPU cores. The graphs out there show Ryzen really starts to ramp in power above this point, and we know from ARM that there’s essentially no frequency scaling on 7nm. Besides, I’d much rather the GPU CUs be the thing driving yield rather than trying to push CPU clock speed as well.
eDRAM is a losing proposition because you’re always paying for it with die space and memory prices can be guaranteed to fall as GDDR6 becomes more of a commodity. Cerny said they considered eDRAM for PS4 and outright rejected, and MS dispensed of it with the Xbox One X design. I hope that means it’s dead for good.
If I hear about anything other than cache on an APU, it better be some cool memory on logic next gen architecture rather than more eDRAM or eSRAM.
eDRAM is a losing proposition because you’re always paying for it with die space and memory prices can be guaranteed to fall as GDDR6 becomes more of a commodity. Cerny said they considered eDRAM for PS4 and outright rejected, and MS dispensed of it with the Xbox One X design. I hope that means it’s dead for good.
If I hear about anything other than cache on an APU, it better be some cool memory on logic next gen architecture rather than more eDRAM or eSRAM.
The performance/power improvements are per device, - transistor. If you cram three times the transistors on a die, you end up with three times the power consumption.
It is still a large improvement, you can increase clocks slightly, double the logic and end up with more than twice the performance at the same power level.
Cheers
https://www.eetimes.com/author.asp?section_id=36&doc_id=1329887
Unless that reality was untrue or has changed drastically over the last two years, I think 400 mm chips at 7nm is a pipe dream for next gen consoles.
Unless that reality was untrue or has changed drastically over the last two years, I think 400 mm chips at 7nm is a pipe dream for next gen consoles.
Probably still too expensive for console adoption, but PC consumer space will benefit a lot from the arrival of QLC NAND.
https://www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives
https://www.pcper.com/reviews/Storage/Intel-SSD-660p-1TB-SSD-Review-QLC-Goes-Mainstream
This is only the first outing of the tech, consumer prices will be even lower when other QLC NAND manufacturers arrive [namely Samsung].
https://www.anandtech.com/show/13078/the-intel-ssd-660p-ssd-review-qlc-nand-arrives
https://www.pcper.com/reviews/Storage/Intel-SSD-660p-1TB-SSD-Review-QLC-Goes-Mainstream
This is only the first outing of the tech, consumer prices will be even lower when other QLC NAND manufacturers arrive [namely Samsung].
I'm a bit pessimistic about future cost reduction for nand. QLC is a nice 33% higher capacity per cell but comes at the expense of cell reliability, which in turn requires more overprovisioning (and only applicable to limited write workload, forget using it as caching or continuous recording gameplay). It's a great cost cutting improvement but it already reached the point of diminishing returns. Number of layers is also seeing a difficulty increasing from now on, and process nodes are reaching a limit of how small the cells can be.
Not sure what the next cost cutting method could be, maybe new materials?
If memory suppliers would start competing again, that would help a lot to bring the prices down in the near term. More of a problem with ram but flash seems to have held the prices for a bit too long.
Not sure what the next cost cutting method could be, maybe new materials?
If memory suppliers would start competing again, that would help a lot to bring the prices down in the near term. More of a problem with ram but flash seems to have held the prices for a bit too long.
Maybe, just maybe, Sony will include a secondary ARM CPU (again) for the PS5, that actually works correctly, taking care of background task, application and streaming needs. This will allow developers full access to all cores on the main CPU, and give them the ability to choose between full compute throughput (i.e., 8 cores @2.6Ghz) or faster clock frequencies (i.e., 6 cores @2.8GHz or 4 cores @3GHz). But of course, this is very dependent on Sony SDKs allowing developers more granular access to Ryzen's CPU core count & frequency setup. Just a possibility...
- Status
Similar threads
- Locked
