Discussion nVidia's ARM SoC Coming Back?

[Tigerick]

Highly Confident:

When Reuter reported NV and AMD are going to launch ARM SoC for PC:

Nvidia dominates the market for artificial intelligence computing chips. Now it is coming after Intel’s longtime stronghold of personal computers. Nvidia has quietly begun designing central processing units (CPUs) that would run Microsoft’s Windows operating system and use technology from Arm Holdings, two people familiar with the matter told Reuters.

Nvidia and AMD could sell PC chips as soon as 2025

I contacted my source, turned up there might have some truths there. Apparently, MS Surface which presumably will upgrade Surface Pro X to Snapdragon X Elite this year. Next year 2025 and later, NV and AMD have won the contract for upcoming Surface X series. PS: And I believe here is the source of Reuter's leaks. If Microsoft going to launch new Surface in 2025, the specs must be better than upcoming Surface Pro X w/ X Elite.......or not? See below:

	X Elite G1	Tegra T239	NV's ARM 64-bit MC	Lunar Lake
Launching Date	Q3 2024	?	2025 ?	Q4 2024
Node	N4P	Samsung 7LPH	Intel 3-E ?	N3B + N6
Memory Supported	128-bit LPDDR5X-8533	128-bit LPDDR5-6400	64-bit LPDDR6-12800	128-bit LPDDR5X-8533
Memory BW	136 GB/s	102 GB/s	102 GB/s	136 GB/s
CPU	12P Oryon	8 x A78	8 cores ?	4P + 4E
GPU	Adreno	Ampere	Blackwell ?	Arc 8 XE
CUDA		1024
RT	N	?	?	Y
FP32	4.6 TF	~ 1.88 TF	~ 4 TF	3.8 TF
AIE (TOPS)	45		?	?

If MS choose 64-bit version, then NV SoC will be used for Surface Go 4 ? Anyone remember Surface RT????

Qualcomm, AMD and NV are all creating their ARM SoC because they are coming after Intel's longtime stronghold of personal computers. As ARM's Cortex X series are getting faster each generation especially with the arrival of LPDDR6 memory, ARM PCs are getting strong enough to challenge x86's performance.

	Intel 's MTL-U	VS	NV's ARM SoC
Launch Date	Q4 2023		2025 ?
Process	Intel 4 + N5 + N6 + N6 w/ Base Tile		Intel 3-E ?
SoC Design	4-tile		Monolithics
CPU	Intel x86		ARM
Processor Base Power	9 - 30 W
Model	Core Ultra 7-164U
Total Threads	14		8 ?
Performance Cores	2P / 4T
Clock Speed	1.1 - 4.8 GHz
Efficient Cores	8E / 8T
Clock Speed	0.7 - 3.8 GHz
LP Efficient Cores	2e / 2T
Clock Speed	0.4 - 2.1 GHz
SLC Cache	12 MB
GB6 1T	2172
PPC	452
GB6 MT	8552
Max Size	64 GB		12, 18, 24 GB
Memory Types	128-bit LPDDR5-6400		64-bit LPDDR6-12800
Memory BW	102 GB/s		102 GB/s
GPU	Intel Graphics		Blackwell ?
ALU	512
Max Frequency	1.8 GHz
FP32	1.8 TF		~ 4 TF ?
NPU	Intel AI Boost
Max Frequency	1.4 GHz
TOPS	10 ?		40 ?

 

[SteinFG]

Watch as nvidia just puts switch2 chip in there lol

 

[Tigerick]

@Kepler_L2

Steam Deck and upcoming Switch 2 are having same memory bandwidth, their raw performance (ie. excluding DLSS) should be similar. There is no way Switch 2's performance is up to XSS which has double memory bandwidth. MLID does not know what he is talking about. We are talking about rasterization performance, DLSS of course will make difference in FPS.

PS: XSS has 4TF while T239 should have around 1.88 TF (depends on final clock speed), slightly higher than Steam Deck's 1.6TF.

 

[Saylick]

@Kepler_L2

Steam Deck and upcoming Switch 2 are having same memory bandwidth, their raw performance (ie. excluding DLSS) should be similar. There is no way Switch 2's performance is up to XSS which has double memory bandwidth. MLID does not know what he is talking about. We are talking about rasterization performance, DLSS of course will make difference in FPS.

PS: XSS has 4TF while T239 should have around 1.88 TF (depends on final clock speed), slightly higher than Steam Deck's 1.6TF.

One thing to note is that Ampere's SM isn't as performant on a per-TFLOP basis as RDNA 2, which is what the Steam Deck uses. When Nvidia transitioned from Turing to Ampere, TFLOPS per SM doubled but performance was only like 30% higher, not double. That 1.88 TF might behave closer to 1.25 TF in comparison to the Steam Deck.

 

[Tigerick]

One thing to note is that Ampere's SM isn't as performant on a per-TFLOP basis as RDNA 2, which is what the Steam Deck uses. When Nvidia transitioned from Turing to Ampere, TFLOPS per SM doubled but performance was only like 30% higher, not double. That 1.88 TF might behave closer to 1.25 TF in comparison to the Steam Deck.

NV got deal mostly because Samsung 8N is cheap

 

[Saylick]

NV got deal mostly because Samsung 8N is cheap

One can also say they got the deal because Nintendo is cheap, too.

 

[soresu]

One can also say they got the deal because Nintendo is cheap, too.

To an extent, but mostly that they would rather risk most investment on the games rather than the hardware since their last truly big step forward console Gamecube was a bit of a flop for them.

I'd expect what follows Switch to be >=3x GPU tho, and likely >=2x CPU.

 

[Ghostsonplanets]

Switch 2 at 1.88 TFLOPs would be 612MHz? That's the clock I would expect for Portable at most. Specially when compared to current Switch clocking at 768MHz Docked.

Also, it depends on their BC solution, but surely they would need to at least match the Docked clock of current Switch for BC games running on Switch 2 to not break, no? Assuming the same 768MHz Docked, Switch 2 would be around 2.35TFLOPs. I guess that is what can be expected.

 

[soresu]

Switch 2 at 1.88 TFLOPs would be 612MHz? That's the clock I would expect for Portable at most. Specially when compared to current Switch clocking at 768MHz Docked.

Given clocks have more than doubled for APUs since the Switch SoC was originally in the SHIELD TV it's a toss up what would be viable today - largely depending on a combo of cooling (to prevent throttling), battery, ALU count and fab node.

 

[soresu]

Ooof wow, taking 2015's Godaveri APU as a reference the clockspeed of APUs has almost quadrupled from its 720 Mhz to the 2,800 Mhz of Hawk Point.

 

[Ghostsonplanets]

Given clocks have more than doubled for APUs since the Switch SoC was originally in the SHIELD TV it's a toss up what would be viable today - largely depending on a combo of cooling (to prevent throttling), battery, ALU count and fab node.

I expect Switch 2 to be lower clocked. Nintendo has a power budget for the entire device of around 5 - 10W Portable and 10 - 13W Docked. I don't think this will change with their next device. But 612MHz would certainly be low for the GPU Docked clocks in my opinion. I think in the upwards of 800 to 900MHz is something more likely to be their choice.

But as you said, it depends on multiple factors. Switch form-factor performance is largely held back due to battery and cooling solution into a small body. High Performance from higher power budgets like PC Handhelds isn't to be expected at all.

 

[soresu]

Switch form-factor performance is largely held back due to battery and cooling solution into a small body

It's also using a 16nm node roughly equivalent to what Zen1 was fabbed on - aka TSMC's first finFET node.

So ye, frequency curve isn't exactly going to be stellar.

But it will be cheap as it comes.

 

[jpiniero]

Given clocks have more than doubled for APUs since the Switch SoC was originally in the SHIELD TV it's a toss up what would be viable today - largely depending on a combo of cooling (to prevent throttling), battery, ALU count and fab node.

Thing is, the Switch 2 is likely to be fabbed at Samsung 8 nm, ie: the same as Ampere. There's a chance it's fabbed at a better Samsung node... but I wouldn't bet on it.

 

[soresu]

Thing is, the Switch 2 is likely to be fabbed at Samsung 8 nm, ie: the same as Ampere. There's a chance it's fabbed at a better Samsung node... but I wouldn't bet on it.

That does fit with Nintendo's habit of using older hw and on cheap nodes.

 

[TESKATLIPOKA]

Steam Deck has up to 1.6GHz TFLOPs.
That means up to 1.5625GHz clockspeed for the 8CU IGP.

Switch has supposedly 6x more Cuda than the previous version. Link
That's 1536 cores, If It's Ampere based, It would mean 12SM.
For It to be on par with Steam Deck It would need only ~800MHz.
If It has only 8SM, then ~1.2GHz.

I personally think Nintendo will have to increase TDP to 15W, but It will be compensated by a bigger battery.

 

[Ghostsonplanets]

Steam Deck has up to 1.6GHz TFLOPs.
That means up to 1.5625MHz clockspeed for the 8CU IGP.

Switch has supposedly 6x more Cuda than the previous version. Link
That's 1536 cores, If It's Ampere based, It would mean 12SM.
For It to be on par with Steam Deck It would need only ~800MHz.
If It has only 8SM, then ~1.2GHz.

I personally think Nintendo will have to increase TDP to 15W, but It will be compensated by a bigger battery.

Wouldn't 1536 CC@800MHz be 2.45TFLOPs though?

For 1536 CC/12 SMs to match SteamDeck 512 SC/ALUs / 8 CU @1.6GHz 1.6TFLOPs wouldn't it just need around 530MHz?

Of course just broadly speaking about the TFLOPs value. Not comparing real world performance nor specific Ampere and RDNA 2 layout or quirks.

As a thought experiment, if Nintendo just use the same clocks as Switch (Which doesn't make sense as it's a new SoC design with an entirely different manufacturing process), it would leave Switch 2 with:

Portable:
1536 CC@307.2 MHz: ~943 GFLOPs
1536 CC@375MHz: ~1152 TFLOPs
1536 CC@460MHz: ~ 1.41 TFLOPs

Docked:
1536 CC@768MHz: ~ 2.35 TFLOPs

 

[TESKATLIPOKA]

Wouldn't 1536 CC@800MHz be 2.45TFLOPs though?

Yes, but I meant real performance not TFlops.

 

[Tigerick]

Charlie @ SemiAccurate reported that NV's ARM SoC is made by Intel 3 process for 2025 launching which aligned with my leak.

I am not sure IFS going to make both 64-bit and 128-bit version, I tend to believe only 64-bit version available. The performance wise of course is much faster than T239. GPU should have around 4TF depends on clock speed and CUDA cores.

And this SoC also likely to be used on NV's handheld console to fight with AMD's custom APU. Can't wait....

PS: Anyone interested on upcoming Surface X RT Edition?

 

[SpudLobby]

Charlie @ SemiAccurate reported that NV's ARM SoC is made by Intel 3 process for 2025 launching which aligned with my leak.

I am not sure IFS going to make both 64-bit and 128-bit version, I tend to believe only 64-bit version available. The performance wise of course is much faster than T239. GPU should have around 4TF depends on clock speed and CUDA cores.

And this SoC also likely to be used on NV's handheld console to fight with AMD's custom APU. Can't wait....

PS: Anyone interested on upcoming Surface X RT Edition?

There’s a 64 and 128-bit bus version?

 

[SpudLobby]

Oh I bet it’s for handhelds.

 

[Ghostsonplanets]

64-bit bus will kill any potential of the iGPU. Unless it's a very cheap SKU like the Snapdragon 7c series.

 

[FlameTail]

64-bit bus will kill any potential of the iGPU. Unless it's a very cheap SKU like the Snapdragon 7c series.

Snapdragon 7c series actually had a 32 bit bus iirc. Horrible.

Still, 64-bit + LPDDR6 is actually 102 GB/s, which is a decent amount of bandwidth.

That's the same amount as Apple M3, and more than the M1 (68 GB/s).

 

[FlameTail]

Why is Nvidia partnering with Microsoft for the next Surface X?

 

[Ghostsonplanets]

Snapdragon 7c series actually had a 32 bit bus iirc. Horrible.

Still, 64-bit + LPDDR6 is actually 102 GB/s, which is a decent amount of bandwidth.

That's the same amount as Apple M3, and more than the M1 (68 GB/s).

Good call with the M1 comparison. Had forgotten that used LPDDR4x. So even a 64-bit bus will provide a good amount of bandwidth with LPDDR5/X/T. As long as Nvidia provides a reasonable cheap SoC, that's good for me.

Why is Nvidia partnering with Microsoft for the next Surface X?

WoA deal with QCOM expired and rumor says MS is pissed off about QCOM not fostering the WoA ecosystem despite having exclusive rights to it until now. So they're partnering with AMD for a custom Arm SoC for Surface and Nvidia is also developing an SoC to enter the WoA market. Not sure if they also have a deal with MS, but they might have.

 

[FlameTail]

WoA deal with QCOM expired and rumor says MS is pissed off about QCOM not fostering the WoA ecosystem despite having exclusive rights to it until now.

Damn. Guess we should have seen this coming.

After more than 10 years of WoA's debut, it still has <0.1% marketshare.

However, the unfortunate thing is that they are doing it just at the time Qualcomm seems to be finally getting their # together, with their Nuvia acquisition bearing fruit and the debut of the new Snapdragon X series.

Even as a Qualcomm fan, I am not complaining. I welcome Nvidia as a competitor to Qualcomm in the WoA space. It will ensure that Qualcomm doesn't become complacent or stagnant.

So they're partnering with AMD for a custom Arm SoC for Surface and Nvidia is also developing an SoC to enter the WoA market. Not sure if they also have a deal with MS, but they might have.

Well, I can see how Microsoft would like to give Nvidia a leg up to enter the industry.