Nvidia N1X CPU leak unveils RTX 5070-sized integrated GPU

[-]

DerpSenpai@reddit

This is a 3nm 5070 GPU that runs on LPDDR5X. It's a super interesting chip

[-]

DerpSenpai@reddit

This chip will be sold with 32-128GB RAM range, not every laptop will have that much RAM. Personally if I am buying this I would go for the 64GB version.

[-]

Plank_With_A_Nail_In@reddit

Does the table show it has 60GB of VRAM available to it? I was hoping nearly all of the 128Gb could be addressed by the GPU.

[-]

Gloriathewitch@reddit

begun, the APU wars have. excited to see strix halo 2.0.

i've always loved the idea of laptops like this which can just flourish instead of walking on thermal eggshells due to the heat dgpus put out. the future is very exciting for midrange gaming laptops.

[-]

jhenryscott@reddit

I think APU high performance budget builds offering better performance that a current budget champ (9600x/9060xt) at a lower price point are gonna shake up DIY pc market soon.

[-]

996forever@reddit

How many more attempts will it take for Reddit to accept expecting “high performance APU” to ever be budget is ridiculously delusional?

[-]

Intel is going to drop its 12 Xe core Panther Lake SKU and they are going to claim victory as you’ll be able to buy it in a $800 mini PC. How performant that will be idk but Intel is the best shot at driving prices down as they win on volume.

[-]

Front_Expression_367@reddit

12 Xe cores are still very small and way less performant compared to the likes of Strix Halo, which is the reason why it can be that cheap to begin with.

[-]

Scion95@reddit

I mean, from what I've heard, and what testing has shown, Intel's tile-based CPUs are a little bit more efficient than AMD's chiplet implementation (I think they use an interposer or bridge or something instead of "just" a PCB) and.

By all accounts, Strix Halo doesn't idle very well at all. It has two Zen CCDs, and the I/O+GPU die has to maintain communication with both of them, and with the memory chips on the 256-bit memory bus. The memory is also on the board, but not on the package, so that's another link that uses more power than it otherwise might. And from what I've heard, that's even more true with a bigger, wider 256-bit bus than a smaller, 128-bit one.

I've heard that for future CPUs, like, AMD even knows this, hence why they're going to start integrating about 2 CPU cores on the I/O die itself. The plan for Panther Lake is also for Intel to start putting more of the I/O, certainly the memory controller, back on the compute tile.

Fundamentally, most operating systems and programs will always need at least one CPU core running at least one thread. And they also will basically always need I/O, memory and storage, to get information in for the CPU and other processors to run.

And, the more distance information travels to get to the processor, the worse. It adds latency and uses more power.

That's also part of the point of APUs and the advantage they have over dGPUs to begin with. The big issue I see with Strix Halo, though, is that it isn't actually a monolithic design.

Chiplets have advantages for cost, yes, but. They have some disadvantages as well.

Intel's foveros or whatever they call it seems to be working for them, performance and power wise, a little bit better for them than AMD's solution.

Basically, even if Strix Halo performs better. It it uses a lot more power, and devices with it need more cooling or get less battery life. There's a bit of an open question about how much it matters or not.

[-]

JRAP555@reddit

May not matter at 1080p though. 12 celestial cores may be enough.

[-]

GenericUser1983@reddit

The iGPU portion is going to put out exactly as much heat as a dGPU of the same generation & performance. Sticking the GPU hardware next to the CPU does not make it use less power when utilized. Granted, it may be easier to manage that heat; putting in one good cooling system may be easier than having two separate systems.

[-]

wtallis@reddit

Waking up a dGPU can easily add 10-15 Watts to a laptop's power draw. There are real efficiency advantages to eliminating the PCIe link and power-hungry GDDR, and once you've integrated the high-performance GPU onto a unified SoC there are opportunities to do a much better job of balancing the system's power budget than NVIDIA+Intel laptops do.

[-]

Scion95@reddit

As far as eliminating power hungry GDDR goes, would LPDDR have enough bandwidth to keep a 5070 running at 5070 levels of performance?

[-]

wtallis@reddit

The GB10 in DGX Spark has less than half the memory bandwidth of a desktop 5070, not even as much as a 5050. But if you're not running at the same clocks as a desktop 5070, then you definitely don't need all the bandwidth that GDDR7 provides.

The GB10 and N1X may end up somewhat bandwidth-constrained, but it's not really surprising that their first silicon in this category sticks with a 256-bit bus rather than trying to squeeze in a 512-bit bus: the 5090 FE shows how crowded a small PCB gets when you're trying to put that much memory down on the board instead of integrating it on-package.

(GB10 is apparently only able to hit 9400MT/s for its LPDDR5, when the standard speed the memory chips are likely rated for is 9600MT/s. So NVIDIA/Mediatek are probably running into the same kinds of trouble we've seen others like Qualcomm and Intel have with getting laptop chips to run at the same memory frequencies used by smartphone chips.)

[-]

Scion95@reddit

Running at lower clocks than a 5070 also means it will have lower performance than a 5070, doesn't it?

There's been a lot of discussion about people assuming that having the same number of CUDA cores as a 5070 and the same arch means it will have the same performance as a 5070, and. What I'm getting at is that it's. Unlikely.

[-]

okoroezenwa@reddit

Yes, but it’ll probably take up a lot of space in any laptop. Apple uses 1024-bit LPDDR5 for the M3 Ultra and that’s 800GB/s.

[-]

Shadow647@reddit

That's very unlikely in laptops, Strix Halo is 256-bit and even that is quite a packaging challenge. Even 512-bit is quite a stretch.

[-]

okoroezenwa@reddit

Definitely, although for 512-bit Apple is doing that for their laptops so it is at least available.

[-]

1soooo@reddit

Most of AMD's iGPU is linked via PCIE and Intel's through DMI aka branded PCIE, how will Nvidia do it differently?

[-]

wtallis@reddit

No, they're not. They just look to software like they're connected via PCIe so that they can be discovered by the OS and use the same drivers as discrete GPUs. There's not a pair of on-die PCIe PHYs talking to each other to link the iGPU with the rest of the chip. That would be hugely wasteful of silicon area and power for signals that never need to leave the chip.

In much the same fashion, early PCIe SSDs appeared to the OS to be a SATA controller connected to exactly one SSD, because at the time every OS had AHCI drivers but most didn't yet have NVMe drivers. But those SSDs delivered performance far in excess of the 6Gb/s limit of the SATA link they pretended to contain; the SATA link was a fiction for software, not an indicator of how the hardware was actually laid out.

And Intel's DMI is a strictly off-chip interface, used to connect a desktop CPU to a southbridge with PCIe as the physical layer. There's no need or use for it in the context of a single SoC package.

[-]

Scion95@reddit

As far as eliminating power hungry GDDR goes, would LPDDR have enough bandwidth to keep a 5070 running at 5070 performance?

[-]

asdfzzz2@reddit

Sticking the GPU hardware next to the CPU does not make it use less power when utilized.

Classical CPU+GPU system - 3 physical links (cpu>mem, gpu>vram, cpu>gpu), 2 sets of memory (ram, vram).

Integrated CPU+GPU - 1 physical link (cpu/gpu>unified mem), 1 set of memory (unified ram). Power savings come from this.

If you sacrifice upgrade possibility and add memory to the chip too (apple way), you get rid of all long-range energy-consuming physical links, saving even more power. Single cooling system simplifies things too, as you already mentioned.

[-]

xternocleidomastoide@reddit

this is the way...

[-]

asdfzzz2@reddit

This way also limits maximum realistic performance of a single unit and increases cost of packaging. Not a silver bullet.

[-]

xternocleidomastoide@reddit

Sure there is never a free lunch.

single package cost goes up. But you reduce overall number of packages and PCB complexity significantly. Performance vs dGPU goes down, but overall system is much cheaper, so you may end up with a better price/performance ration... it's all tradeoffs.

But it makes strategic sense for nvidia to start having an iGPU SKU, since the way things are going, dGPUs are going to be a very stagnant market for laptops.

[-]

goldcakes@reddit

Plus NVIDIA gets even more money in the BOM since they’re doing the cpu too, if not also memory.

[-]

Gloriathewitch@reddit

Thats not true, hx370 can get that perf with 15w and strix gets its perf with around 35w iirc from memory.

higher watts translates to higher temps.

it also allows you to run a larger rad and a more central heatsink. you're focusing 100% of a system on one component, not splitting it across the chassis.

strix halo under load uses approx 140w, theres dgpus that use that by themselves.

[-]

GenericUser1983@reddit

The (current gen) dGPUs that use 140w by themselves are much faster than the Strix Halo iGPU; the mobile 5070 ti is a 140w part and easily beats Strix Halo on anything that fits in its 12 GB VRAM.

To get the same performance as a dGPU the iGPU of the same generation will need the same amount of silicon with the same amount of power running through it, producing the same heat. Yes, as I mentioned that heat may be easier to manage with one big cooler vs two small ones, but the same amount of heat is going to be produced.

[-]

Gloriathewitch@reddit

strix halo is roughly 4060/4070m rtx which is more than enough for 1080p, if they add fsr4 it'll run even better

[-]

FitCress7497@reddit

These are very misleading - Roughly 4060/4070 mobile? AT WHAT POWER CONFIG. watch Jarrod's review. It's still slower than a 100W 4060m, while consuming estimated 70~90W for the 8060S alone. It's nowhere near a 100W 4070m. - If they add FSR4? How exactly?. It's RDNA 3.5, not RDNA 4. At best it will get a crappy version. - Why do you need crazy specs in a portable gamer? With the same money for any laptop with a 395+ (rog flow z13 and hp zbook g1a), you can even buy a laptop with 4090m lol

[-]

Gloriathewitch@reddit

ive already told you three times: Radeon 890m 15w(sole gpu, 45 package) 8060s roughly 35w(100+ package)

why are you just ignoring what im saying?

[-]

996forever@reddit

Because everything you’ve mentioned is demonstrateably false

[-]

BloodyLlama@reddit

also allows you to run a larger rad

??? I have as many rads as I could fit in my case, and had to cut big ass holes to do so. How is reducing my number of water blocks going to increase my rad space?

[-]

Gloriathewitch@reddit

we're talking about laptops, and they have inbuilt radiators and cooling pipes.

[-]

BloodyLlama@reddit

If you're talking specifically about laptop it's probably best to actually say that rather than assume everybody can read your mind. Either way I don't understand how you can magically use a bigger heatsink* because you moved the GPU.

presuming you mean heatsink when you're typing radiator.

[-]

ParthProLegend@reddit

Hx370 supremacy. Best cpu for laptop under 1L

[-]

lintstah1337@reddit

HX 370 has anemic 16MB L3 cache for 4x Zen 5 and 8MB for 8x Zen 5c cores.

As a comparison, a regular desktop 1 CCD Zen 5 (Ryzen 7 9700x) has 32MB L3 Cache and an X3D variant (Ryzen 7 9800X3D has 96MB (32MB+64MB 3D V-Cache) L3 Cache

L3 Cache size makes a huge difference in gaming which is why X3D cpus are so much better than other CPUs.

[-]

Gloriathewitch@reddit

yeah i really want one. that or a 258v

[-]

Raikaru@reddit

Those DGPUs are faster i don’t understand your point here ngl

[-]

Gloriathewitch@reddit

890m is 1660/2060 maxq 8060s is 4070

they use like a tenth of the power.

the efficiency is why they are so good, and they can play 1080p games well.

[-]

loczek531@reddit

890m is 1660/2060 maxq

In what, synthetic benchmarks? For sure not in games

[-]

GenericUser1983@reddit

The 1660/2060 are older chips, made on a less advanced fab process than the 890M; that's why they need more power for similar performance. The 8060s iGPU chup is made on a similar fabrication process as the 4070, and when you run the same amount of power through them the performance is also pretty similar.

[-]