NVidia announces Carmel ARM CPU

[NTMBK]

There's a TX2 now. Nintendo was too cheap to use it in the first run of the Switch. Hopefully they'll come to their senses and use the TX2 in a hardware refresh somewhere down the road. Should extend battery life and/or enable Nintendo to use higher clocks when the console isn't docked.

I thought the X2 was more focused on automotive?

 

[arandomguy]

There is automotive specific features (basically silicon budget) but that wouldn't preclude it being used in other applications. Tegra X1 has features in that I don't believe Nintendo would have been interested in or are used on the Switch either. Fundamentally it is still a newer/faster CPU and GPU combination at the end.

The consideration could be how much more efficient it actually is both in terms of cost and power. I don't believe Tegra X2 is used for any direct consumer devices? Bear in mind Tegra X1 was on TSMC 20nm. So the desktop Maxwell->Pascal uplift is likely not as comparable (28nm to 16nm FF+). Also on the CPU side I'd guess for maximum uplift you'd have to leverage the additional complexity of optimizing for the 2 Denver cores.

 

[Genx87]

So why nVIDIA doesn't dare to enter to the laptop market?

Hyper-competitive and shrinking? And did MS continue development of Win10 on arm? I honestly can't remember.

 

[Nothingness]

The consideration could be how much more efficient it actually is both in terms of cost and power. I don't believe Tegra X2 is used for any direct consumer devices? Bear in mind Tegra X1 was on TSMC 20nm. So the desktop Maxwell->Pascal uplift is likely not as comparable (28nm to 16nm FF+). Also on the CPU side I'd guess for maximum uplift you'd have to leverage the additional complexity of optimizing for the 2 Denver cores.

NVIDIA official figures are 10W for Jetson TX1 and 7.5W for Jetson TX2 in Max-Q mode (the mode where only the 4 Cortex-A57 are used while the 2 Denver are powered off). The Max-P mode is 15W.

Ref: https://devblogs.nvidia.com/jetson-tx2-delivers-twice-intelligence-edge/

 

[CatMerc]

NVIDIA won't enter the CPU market mainly because they don't have the know how and R&D budget to start doing anything remotely competitive with Intel or AMD.
You might say NVIDIA has a larger budget than AMD, but remember that AMD already has design methodology and expertise for a high performance x86 CPU in house for decades, NVIDIA would need to essentially start from scratch even if they did get the x86 license. That would be immensely expensive and simply not worth pursuing when they can just license ARM cores and be done with it. Intel and AMD just have too much built up for NVIDIA to try and enter in the middle and fight for scraps.

I'm not even sure why they keep trying with their in house VLIW cores, they're not exactly impressive next to beasts like the A76.

 

[Qwertilot]

They must think they're well suited to their specific needs in compute?

Either that or they set the R&D running a while ago when they were trying to compete in the Arm market and we're mostly seeing the inertia working through.

 

[DrMrLordX]

I thought the X2 was more focused on automotive?

I think @arandomguy said it best. But the main thing I was getting at, is that it can do everything the TX1 does, but faster and at lower power.

Also they're a fairly big customer, so NV could easily spin new version of the chip for handhelds. Not sure if that would be cost-effective versus using an "off-the-shelf" TX2.

 

[Thala]

And did MS continue development of Win10 on arm? I honestly can't remember.

There is not separate development branch for ARM. Builds including latest insider builds are available for x86/x64/ARM64 this includes versions with installer as well as home and professional editions.
In fact you can just download a Windows 10 for ARM version and install it under QEMU (or on your Raspberry Pi)

 

[darkswordsman17]

I think @arandomguy said it best. But the main thing I was getting at, is that it can do everything the TX1 does, but faster and at lower power.

Also they're a fairly big customer, so NV could easily spin new version of the chip for handhelds. Not sure if that would be cost-effective versus using an "off-the-shelf" TX2.

I'd think Nvidia would be willing to do a custom Tegra design for Nintendo, but who knows. Nintendo might balk at the price. I think the only reason they went Nvidia is because AMD didn't have a mobile focused SoC, and NVidia was probably looking to dump old TX1 chips since I don't believe they sold terribly well. But basically every company that uses Nvidia ends up moving on it seems like (well outside their bread and butter markets), so I won't be surprised if Nintendo just moves to some new ARM design 3-4 years down the line (Nintendo likes to stick with consoles, especially when they're successful, so I don't know that we'll see them update the Switch like Sony and Microsoft have started to do).

 

[DrMrLordX]

(Nintendo likes to stick with consoles, especially when they're successful, so I don't know that we'll see them update the Switch like Sony and Microsoft have started to do).

Nintendo respins hardware on their portable platforms all the time.

Gameboy/Gameboy Color, for example. I think there was a refresh of the Advance, and of course you had the DS, 3DS, and 2DS. I can totally see Nintendo releasing an updated Switch that's 100% software-compatible with the release model.

 

[moinmoin]

I'd think Nvidia would be willing to do a custom Tegra design for Nintendo, but who knows. Nintendo might balk at the price. I think the only reason they went Nvidia is because AMD didn't have a mobile focused SoC, and NVidia was probably looking to dump old TX1 chips since I don't believe they sold terribly well. But basically every company that uses Nvidia ends up moving on it seems like (well outside their bread and butter markets), so I won't be surprised if Nintendo just moves to some new ARM design 3-4 years down the line (Nintendo likes to stick with consoles, especially when they're successful, so I don't know that we'll see them update the Switch like Sony and Microsoft have started to do).

When their involvement was publicized Nvidia claimed that it was the start of a 10 years partnership. Also the TX1 in Switch has no customization at all. So my guess is Nintendo gets TX2 for comparable cheap sometime down the line when it's essentially worthless to Nvidia, just like TX1 was to them.

 

[NostaSeronx]

When their involvement was publicized Nvidia claimed that it was the start of a 10 years partnership.

Twenty years.

There are two upgrade paths to order though:
- The enhanced dock with eGFX. 1060 in dock + 8 GB RAM Switch ver. 1.1
- Introduction of Denver/Carmel.

 

[NTMBK]

New blog on Xavier: https://devblogs.nvidia.com/nvidia-jetson-agx-xavier-32-teraops-ai-robotics/ Not a great deal of detail on Carmel:

Jetson AGX Xavier’s CPU complex shown in figure 10 consists of four heterogeneous dual-core NVIDIA Carmel CPU clusters based on ARMv8.2 with a maximum clock frequency of 2.26GHz. Each core includes 128KB instruction and 64KB data L1 caches plus a 2MB L2 cache shared between the two cores. The CPU clusters share a 4MB L3 cache.

The Carmel CPU cores feature NVIDIA’s Dynamic Code Optimization, a 10-way superscalar architecture, and a full implementation of ARMv8.2 including full Advanced SIMD, VFP (Vector Floating Point), and ARMv8.2-FP16.
The SPECint_rate benchmark measures CPU throughput for multi-core systems. The overall performance score averages several intensive sub-tests, including compression, vector and graph operations, code compilation, and executing AI for games like chess and Go. Figure 11 shows benchmark results with a greater than 2.5x increase in CPU performance between generations.

Dynamic Code Optimization sounds like this is indeed an evolution of Denver.

 

[Nothingness]

New blog on Xavier: https://devblogs.nvidia.com/nvidia-jetson-agx-xavier-32-teraops-ai-robotics/ Not a great deal of detail on Carmel:



Dynamic Code Optimization sounds like this is indeed an evolution of Denver.

Yeah I guess it's just some updated Denver with ARMv8.2 support.

As far as perf goes, if we assume A57 is half the speed of Denver (that's what I measured, but not on SPEC; I''ll run SPEC when I have time) that means TX2 has ~4 Denver cores, so scaling ot 8 cores, that means ~120 vs 160 for Xavier, so that'd be ~30% faster per core. OK that's wild guessing and likely completely stupid

 

[darkswordsman17]

Nintendo respins hardware on their portable platforms all the time.

Gameboy/Gameboy Color, for example. I think there was a refresh of the Advance, and of course you had the DS, 3DS, and 2DS. I can totally see Nintendo releasing an updated Switch that's 100% software-compatible with the release model.

That is somewhat true, but they don't tend to change things up too drastically and its usually not a fast change. The DS Lite was years down the line though wasn't it? Looking it up guess it was a bit over a year. And it was mostly a form factor change. The DSi was years later and had a processor upgrade. The 3DS was form factor changes (XL and 2DS) for 3.5 years before changing the internals.

Which the move to X2 would probably be minimal, but I'm not sure I'd expect it terribly soon.

When their involvement was publicized Nvidia claimed that it was the start of a 10 years partnership. Also the TX1 in Switch has no customization at all. So my guess is Nintendo gets TX2 for comparable cheap sometime down the line when it's essentially worthless to Nvidia, just like TX1 was to them.

That was just Nvidia saying stuff and doesn't actually mean anything (JHH actually said two decades even...). I still have a hunch that Nvidia's chips are too expensive (with that being a big reason why the Switch is using the X1 instead of X2). The security issue with the X1 didn't please Nintendo.

The first version might not have been (even after Nvidia claimed it was...) but since they had to fix that hardware flaw they would've had to do a new production run (not really custom but that was more work Nvidia had to do). If it can just slot in without much if any change, I'd definitely see there being an X2 based Switch, or maybe they wait a while and release it with a bit of an overhaul of the design.

The real question will come when Nintendo is deciding on a new chip (don't think Nvidia has announced a consumer based newer Tegra chip?).

I wouldn't be totally shocked with Nintendo sticking with Nvidia, but we'll see. Nintendo is pretty fiscally conservative and tends to like having control. I think there were benefits to using an API made by someone else, but if Nvidia tries to hold that over them financially, it could become a problem. The thing is, if ARM keeps making progress, future versions might have GPU compatibility in line with the Tegra chips, making a transition away easier.

Even if they go with the X2, the CPU is still a bit lackluster, and while I'm sure Nvidia could make an updated one with newer ARM cores, it'd be about the cost. If another company could offer close to equal GPU performance with better CPU performance, for a lot less money, I think Nintendo might consider it.

IIRC rumors put the X1 at like $40-50 per chip for Nintendo. I'm pretty sure they can find someone to make an ARM chip for a lot less than that.

 

[moinmoin]

That was just Nvidia saying stuff and doesn't actually mean anything (JHH actually said two decades even...).

Yeah, two decades it was. The meaning behind that was that Nintendo stuck with the same graphics folks (Silicon Graphics -> ArtX -> ATi -> AMD) for nearly two decades before switching to Nvidia, so JHH sure wants to outdo them. Looking back I'm pretty sure the Nintendo-AMD link severed when the all that long time veteran staff including David Wang left in 2012. Now he's back at AMD.

 

[Andrei.]

I actually have an AGX for review. I'll be publishing detailed SPEC2k6 scores for it. Though don't expect much microarchitectural stuff.

 

[NTMBK]

I actually have an AGX for review. I'll be publishing detailed SPEC2k6 scores for it. Though don't expect much microarchitectural stuff.

Thanks for the article Andrei! The Carmel CPU honestly doesn't look very impressive, and the software cost of maintaining the dynamic code generator must be nasty. I'd be amazed if they don't use ARM designs (possibly with customisations) next time out. https://www.anandtech.com/show/13584/nvidia-xavier-agx-hands-on-carmel-and-more/3

 

[NTMBK]

I'd be amazed if they don't use ARM designs (possibly with customisations) next time out.

Called it: https://www.anandtech.com/show/1524...-orin-a-herculean-arm-automotive-soc-for-2022 RIP the Denver line.

 

[NostaSeronx]

RIP the Denver line.

The team has restarted from scratch. It will be based on a proprietary Nvidia CPU ISA not based on ARM, but closely related to x86 apparently. It isn't x86/x86-64, nor is it ARM.

-> New CPU
-> New ISA

Since, they aren't continuing Denver/Carmel, what is the point of Orin using Carmel2(v1.5)?

 

[NTMBK]

The team has restarted from scratch. It will be based on a proprietary Nvidia CPU ISA not based on ARM, but closely related to x86 apparently. It isn't x86/x86-64, nor is it ARM.

-> New CPU
-> New ISA

Since, they aren't continuing Denver/Carmel, what is the point of Orin using Carmel2(v1.5)?

Got any links? Because I've not seen anything to indicate anything like that.

 

[DeathReborn]

Got any links? Because I've not seen anything to indicate anything like that.

All I found was RISC-V.

Is NVIDIA Doubling Down On RISC-V? | Synced

For years NVIDIA has built its reputation on specialized graphic processing cards (GPU). The US chip giant now appears to be exploring the terrain beyond GPU architecture, raising the stakes on other dedicated electronic circuits. Synced recently observed a series of job postings seeking experts...

syncedreview.com

 

[Nothingness]

All I found was RISC-V.

Is NVIDIA Doubling Down On RISC-V? | Synced

For years NVIDIA has built its reputation on specialized graphic processing cards (GPU). The US chip giant now appears to be exploring the terrain beyond GPU architecture, raising the stakes on other dedicated electronic circuits. Synced recently observed a series of job postings seeking experts...

syncedreview.com

All this points to NVIDIA using RISC-V as a micro-controller. This won't replace the ARM CPU any time soon.