Ok memory card doesn't equal cartridge which means I'm still expecting load times (look at Vita). I want cartridge confirmation and zero load times
It's worth clarifying the difference between old-style cartridges, modern 3DS-style game cards, and memory cards (i.e. SD cards).
Cartridges - Used in NES, SNES, N64, etc - ROM - Directly addressable via the memory bus, meaning data doesn't need to be loaded into memory when starting up the game, giving close to zero load times. Impractical on a modern system for a variety of reasons.
Game Cards - Used in DS, 3DS, NX - ROM - Accessed via its own bus, assets must be loaded into memory before use. Modern cards should be comfortably faster than optical media, and potentially faster than HDDs for contiguous reads, and orders of magnitude faster than both in terms of latency. Not zero load times but potentially faster loading than systems with disc-based media.
Memory Cards - SD Cards, etc. - NAND Flash - Not used in games consoles. Slower than game cards due to the need to run everything through a flash controller chip.
NX will be using game cards, not cartridges or flash memory cards. The load times are going to depend on the interface Nintendo uses for the cards. The DS and 3DS use an 8-bit parallel interface, but this was chosen back in 2004 for a handheld with limited 3D capabilities, and had to accommodate writing saves to the card. If they're designing a new game card interface for a modern system it would make sense for them to switch to a serial interface and drop support for on-card saves. Hopefully they adopt a variable-speed serial interface, potentially allowing games that need it as much as 250MB/s or even 500MB/s (or higher for NX2) without increasing costs for games that don't.
Maybe they really did postpone the release to get a custom Tegra X2, and there's no update because they are in fact only getting built now.
The design of the SoC would have been pretty well finalised by the time they announced the March 2017 release date. It's possible they made a switch from TSMC's 16FF+ to their new 16FFC process, which would potentially allow them to bring down power consumption and reduce costs a little, but there wouldn't have been any major redesigns possible at that stage.
I am going to draw out some scenarios so that people understand what could happen with NX's GPU.
We currently have dev-kits which are not finalised and contain a Tegra X1. It is rumoured from NateDrake that the NX will have a Pascal-based Tegra GPU.
Nvidia license out their GPUs, they don't do Semi-Custom designs like AMD which has been mentioned by other posters.
I will show what scenarios could happen. This ignores customisation.
- The GPU is unchanged and Nintendo have a stock Tegra X1 as the GPU for the NX. Keep in mind that 20nm chips do not have much manufacturing lines because it was a die size most companies passed on or stopped early because the performance benefits weren't significant.
- The Tegra X1 GPU has a die-shrink to 14nmFF, this gives a performance boost to save on energy but keeps the Maxwell Architecture.
- The GPU is Parker which is a 14nmFF design and has Pascal architecture.
Nvidia's GPUs are licensed, asking to have a Tegra X1 stay 20nm while changing to a Pascal architecture means designing a whole new chip. It is cheaper to pick Parker because it is already a GPU design with Pascal architecture.
If Nintendo decided to stick with some random 28nm chips, they would still have to do a die-shrink later on because it becomes expensive to manufacture 28nm over time when smaller nodes become cheaper to manufacture. 28nm has been around for a few years, things have finally picked up with 16nmFF/14nmFF that 7nm is expected by 2019.
Those are the scenarios I have drawn out.
There's really no reason to frame Nintendo's options in relation to existing Tegra chips. The NX will almost certainly be the biggest selling Tegra-powered device ever made by a substantial margin, and the chip that powers it will be made in far larger quantities, perhaps by an order of magnitude or more, than either TX1 or Parker. If anything, it would make far more sense for Nintendo and Nvidia to design NX's SoC from scratch and then use a variant of that for Parker, rather than vice versa.
Maxwell and Pascal are near-identical graphics architectures, and are both inherently modular and scalable, based around 128 ALU "Shader Modules" or SMs. Really there are only three questions when it comes to NX's GPU hardware:
- What manufacturing process will it be made on (ie 28nm, 20nm or 16/14nm)?
- How many of each functional unit will it have (ie how many SMs, ROPs, etc)?
- What clock speed will it operate at?
Nintendo's decisions on each of these are pretty much completely independent of the decisions Nvidia made for TX1 and Parker, and will depend on their own preferences when it comes to cost, performance and power draw. Both Nvidia and Nintendo have ample experience designing mobile SoCs, and in the grand scheme of SoC design this is definitely a case where the expected production run is so large the designers can do whatever they like without feeling constrained to simply modify an existing design.
(Also, while Nvidia haven't operated on a semi-custom business model recently, there's no reason to believe they wouldn't if it were profitable for them, just as AMD operate both semi-custom and licensing models depending on the client).