StereoVsn
Member
And HPC market is growing fast and carries fat margins. Throw in VDI market and its expansion into cloud and Nvidia being the only provider there.
-
Hey, guest user. Hope you're enjoying NeoGAF! Have you considered registering for an account? Come join us and add your take to the daily discourse.
DF: Nintendo NX Powered By Nvidia Tegra! Initial Spec Analysis
- Thread starter DeathSparks
- Start date
- Status
StereoVsn
Member
Yeah HPC / server market is disappearing because Google plans to use custom ASICs... Wait, no, it's growing with or without Google. A lot of software utilizes CUDA, a lot of cloud vendors (like Azure and AWS) utilize Nvidia. A lot of corp/university HPC deployments utilize Nvidia. Their commercial side is growing and growing fast.
While utilizing custom ASICs or quantum computing (another thing Google is using) is something that is starting to be utilized for certain tasks that leaves a huge chunk of he market unable or unwilling or outpriced by such.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
That's good. I didn't check their last quarter so I was not aware they had that much of an increase. Though, as good as that sounds, they're facing there the classic problem - once you have a market to yourself, and you've cranked the margins to 11, what else can you do the get your next increase?
wonderful103
Member
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
All those 'new markets' pushes will be good as long as GPGPU is useful for them. Now, while general-purpose programmability might be an inherent requirement for some of them, it definitely is not for others. Again, for reference see the bitcoining which started off massively on GPGPU but in the course of a few years ended up on ASICs which were neither more expensive nor less productive than even the best-suited GPUs. In contrast HPC is firmly GPGPU territory, as general programmability will always be valued there.
dr_rus
Member
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
I never said anything about how simple building stream processors was. Google spent a good amount of resources on their designs.
Well, let's put things into perspective, shall we:
Tegra3 was the last Tegra that scored with mobile. Mind you, it wasn't a great success, just had some wins - T4 was a fiasco in comparison. Notice how their current Tegra revenue has not reached their 2013 levels? So at the end of the day, NV left mobile for greener pastures, only those were not as green. What is worse, automotive will abandon them at the drop of a hat for autonomous car purposes, once the tools of the trade settle down. That would leave Tegras in automotive infotainment alone - as fickle a business as it gets.
dr_rus
Member
Slides are out: Nvidia Tegra Parker : Neue Denver-2-ARM-Kerne treffen Pascal-Grafik in 16 nm
It's very much a Tegra for automotive industry. GPU is essentially the same (well, with Pascal additions obviously, but the same SP count), biggest changes are in CPU and doubling of memory bandwidth.
I think it's pretty safe to assume that Nintendo would be using something custom based on Parker design instead of using Parker itself.
It's very much a Tegra for automotive industry. GPU is essentially the same (well, with Pascal additions obviously, but the same SP count), biggest changes are in CPU and doubling of memory bandwidth.
I think it's pretty safe to assume that Nintendo would be using something custom based on Parker design instead of using Parker itself.
veggiesaurus
Member
half-precision?
Only 256 CUDA cores? That's not very encouraging.
Most likely. The GPU would need to be over 2GHz otherwise.
Wouldn't doubling of memory bandwidth be a very important factor for Nintendo?
dr_rus
Member
Sure but MC is something which can be changed pretty radically in a custom design anyway so it's hardly something which needs Parker itself.
We know the 1.25TF is not half precision. The Numbers provided by Nvidia are for FP32.
blu
Wants the largest console games publisher to avoid Nintendo's platforms.
Good find!
And where is that said? Even watercooled, there's no way the GPU is running at ~2.5GHz.
ethomaz
Banned
FP32.
nVidia site says 2.5TFs for 2 Tegra Pascal chip.
I'm not sure if they count the ARM processor flops together.
http://www.nvidia.com/object/drive-px.html
Kaffeemann
Banned
For idiots like me: what is the gaming performance of that chip?
It doesn't say which it is. There's no chance in hell that the GPU is running at 2.5GHz though, so the logical conclusion is FP16.
Nobody knows. It's not made for gaming, and there are some uncertainties with it.
The Bookerman
Member
You should probably temper those expectations.
Plinko
Wildcard berths that can't beat teams without a winning record should have homefield advantage
I don't think they're asking about NX--just the specific chip mentioned.
You are giving for granted that NX will use this latest Tegra chip, to me it's more likely they'll be using the older one, NX needs to be cheap, and rumors are indeed that it'll be cheap, i guess Nintendo will take advantage of the release of this new Tegra chip so that the older one will drop its price, perhaps a NX revision in 2 years from now could use the new Tegra.
After all that's what console always do, for example PS4 and Xbox One on their launch in 2013 used AMD custom VGA based on 2011 graphics cards
After all that's what console always do, for example PS4 and Xbox One on their launch in 2013 used AMD custom VGA based on 2011 graphics cards
I don't know where to post this but recently released Nintendo Company Book shows that they're studying the competition 
Source: https://twitter.com/NintendoMemo/status/767679843333005312/photo/1
Source: https://twitter.com/NintendoMemo/status/767679843333005312/photo/1
michaelius
Banned
Well yeah - how many chip providers do you know who would be not interested in possible 100 milion units contract ?
ethomaz
Banned
This is speculation from an update in old article in AnandTech.
I believe it is including ARM flops.
In any case the total system with 2x dGPU is up to 8TFs with 2x Tegra counting for 2.5TFs... all FP32.
DoctorFling
Member
I don't know where to post this but recently released Nintendo Company Book shows that they're studying the competition
Source: https://twitter.com/NintendoMemo/status/767679843333005312/photo/1
NX reveal delayed while Nintendo staff play Bloodborne confirmed
NX will launch somewhere in 2017. So Tegra Parker will also be "old" by then. Plus it would justify the delay of NX for next year. Not that it will use an off the shelf Parker, but the custom chip NX will use can benefit from few things from Parker, like 16nm and double memory bandwidth would be the obvious choices.
ethomaz
Banned
This new Tegra chip is cheap than old one... it is way small... it is the same 256SPs in 16nm.
Small = cheaper in silicon terms and for that older chips trend to be more expensive than new ones.
Nostremitus
Member
Has Nintendo ever used a bog standard GPU solution?
Whatever the NX has it's almost certainly custom designed, so the core count of this chip wouldn't matter, right? They typically license the architecture and help design their own custom silicon for their consoles. Or am my mistaken?
Whatever the NX has it's almost certainly custom designed, so the core count of this chip wouldn't matter, right? They typically license the architecture and help design their own custom silicon for their consoles. Or am my mistaken?
Jacethemindtaker
Neo Member
Who cares what the graphics are like as long as the games are good, seriously..this obsession is beyond my ken..
It seems that you're in the wrong thread.
We need more information. You need a very special chip and LN2 to hit 2.5GHz. Including the CPU wouldn't explain it either.
It's not really surprising for an automotive SoC designed to be used in conjunction with a dedicated GPU. There's not much point going crazy on the integrated GPU at that point.
As previously discussed, I wouldn't read too much into benchmarks for Denver, the microarchitecture (or more specifically the use of dynamic code optimisation) is basically built for benchmarks, and how it handles real-world code could be a different matter entirely.
Interestingly, slide 6 actually provides a fascinating little nugget of insight into this:
This seems to be a confirmation of something I speculated a few months ago, that Nvidia's decision to combine Denver and A57 cores on Parker is a recognition that Denver has some fairly significant performance weak spots, and they're offloading threads which perform poorly on Denver to the A57 cores. I'd be interested to know what kind of heuristics they're using to allocate threads, although I very much doubt that's something they'd ever talk about.
Yes, Blu pointed out in the Eurogamer NX thread how the HMP it is using is important and noted it is 'proprietary coherent interconnect' from slide 4.
dr_rus
Member
That'd mean 2.44GHz for FP32 which is basically impossible without active cooling and in mobile power envelope. 1.22GHz with FP16 being the basis for this number OTOH sounds completely possible on 16FF+ and a Pascal shader core. So we're looking at ~624 GFlops of FP32 and 1.25 TFlops of FP16.
Then again it's rather unlikely that Nintendo will use Parker without any customizations as lots of Parker features make sense only in automotive applications.
Not sure what you mean as it handles benchmarks in the same way it handles real world applications, there is not difference. The offloading to A57 partition is there for tasks which benefit from multithreading more than from a high single thread performance. I would assume that a thread which needs high performance processing will be mapped to Denver2 core while lighter threads will go on A57 which are worse in ILP but are likely better in perf/watt.
I'd think Denver 2 would be overkill from a CPU standpoint, plus the slides don't mention anything about power draw, so it's likely not as power efficient as some of the other ARM options that would still match or beat PS4/XB1 in CPU performance.
Of course it's hard to say without more info about these particular Denver processors
Of course it's hard to say without more info about these particular Denver processors
Nintendo could take out the Denver cores, leaving just the 4 ARM cores on there. Saves a bit of money, space, watt power. Lower sustained clocks for the CPU and GPU cores, to keep it within thermal envelopes. Say, 800Mhz. Call it custom at that point? On 16nm that should be around 3-4 watts according to my calculations. If they want a small 3DS sized console anyway. They can increase power slightly for a bigger device.
dr_rus
Member
That first line sounds like Parker should consume less than TX1 while providing more CPU performance - but this is hardly surprising considering the move from 20SOC to 16FF.
With the size of TVs increasing fast, 720p is starting to show its age, 1080p still seems like the sweetspot, I really want to know what resolution they will go for NX docked, I'm having hard time looking at Wii U games with my 40" TV, most of them are 720p, I need to sit about 3 meters away from the screen. In general I'm very tolerant to low resolutions, the problem is TV manufacturers are moving quickly to UHD, and those panels are not ideal to watch sub 4k content.
I don't know where to post this but recently released Nintendo Company Book shows that they're studying the competition
Source: https://twitter.com/NintendoMemo/status/767679843333005312/photo/1
That's good, they have been living in their bubble for far too long
ethomaz
Banned
I found some interesting info.
nVidia page spec for Drive PX (the first one) says 2.3 TFLOPs for a dual configuration... that means they are talking about FP16 that means 575 GFLOPs FP32 for each Tegra X1 inside it.
Let's works with that way of thinking...
Drive PX FP32: 575 GFs
Drive PX FP16: 1.15 GFs
Now how about Pascal Tegra?
Drive PX2 FP32: 625 GFs?
Drive PX2 FP16: 1.25 GFs?
I don't know if nVidia add CPU flops to the math or not.
nVidia page spec for Drive PX (the first one) says 2.3 TFLOPs for a dual configuration... that means they are talking about FP16 that means 575 GFLOPs FP32 for each Tegra X1 inside it.
Let's works with that way of thinking...
Drive PX FP32: 575 GFs
Drive PX FP16: 1.15 GFs
Now how about Pascal Tegra?
Drive PX2 FP32: 625 GFs?
Drive PX2 FP16: 1.25 GFs?
I don't know if nVidia add CPU flops to the math or not.
that feels like you have those reversed
My guess is that the chip in NX is something a lot closer to TX1 but featuring 16nm Finfet and some extra work in the memory setup, I dont expect extra cuda cores or Denver CPU cores. This is what makes most sense to me with all the available information.
But we already know that Nintendo is known to make a lot of decisions that don´t make sense, so there is that.
But we already know that Nintendo is known to make a lot of decisions that don´t make sense, so there is that.
He doesn't.
- Status