Hacker News Reader

Comment by paulpan | Hacker News Reader

> What happens to GPU? Having their own GPU for iMac and Mac Pro as well? Dual GPU options where Apple GPU for power efficiency? This feels like additional complexity.

I think GPU scaling will be much harder than CPU, so whereas Apple can surpass Intel CPUs for all but the highest segments, putting together a standalone GPU will be hard and very interesting to see. For an entry-level GPU? No issues. But what about a midrange (AMD RX 5700XT or Nvidia 2070S)? And not to mention the top-tier Nvidia 2080ti.

The other unspoken risk is that while Apple may be vertically integrating its SOC, it still relies on a fab like TSMC. Intel's recent problem is rooted in their inability to move off legacy 14nm fabrication process. TSMC may have done great in 7nm and now to 5nm transition, but what happens if/when they stumble? Would Apple also want to acquire them or build its own fabs to mitigate this risk?

View on HN

> The other unspoken risk is that while Apple may be vertically integrating its SOC, it still relies on a fab like TSMC. Intel's recent problem is rooted in their inability to move off legacy 14nm fabrication process. TSMC may have done great in 7nm and now to 5nm transition, but what happens if/when they stumble? Would Apple also want to acquire them or build its own fabs to mitigate this risk?

Surely this is an advantage to being fabless? If TSMC stumble, they can evaluate other options. Same for AMD, where would they be now if they were still tied to GlobalFoundries?

Valid point as a GF-tied AMD would not be in the same position as today.

That said, what are the other options if not TSMC? Besides Intel, Samsung is only other cutting-edge option. Intel's 7nm would be technically on par with TSMC's 5nm (marketing names aside). https://en.wikichip.org/wiki/7_nm_lithography_process

There is a chance, however unlikely, that TSMC's 3nm push will run into issues and be delayed. Would create an interesting scenario where Apple would pay Intel to fab their SOCs.

>Apple would pay Intel to fab their SOCs.

Why would Intel agree to that? Have they ever fabbed any other companies design?

Apple said they’ve produced 2 billion A-series SoCs in ten years. Would Intel turn down a slice of that very large pie if offered?

Intel currently doesn't have enough capacity to make their own chips and are rumored to be outsourcing to Global Foundries as a result ( https://wccftech.com/rumor-intel-moving-select-cpus-to-globa... - huge grain of salt on this one ofc, but the supply constraints on Intel's fabs are well known - they mentioned it even at their earnings call) - why would they stop making their own products to make Apple's instead? Apple'd have to pay an absurd amount for that to make sense.

Apple has an absurd amount of money.

Yes: https://www.intel.com/content/www/us/en/foundry/overview.htm...

There were rumours going around about its demise a few years ago, but a fair bit of that was simply their failure to ship 10nm parts on schedule AFAIK. They're still doing some degree of third-party manufacturing, and I don't doubt once they reach a point of having the capacity for their first-party products on 10nm we might seem them expand.

However, the inevitable flip-side of this is unlike TSMC/SS where Apple can bid the highest for the early production of a new node, Intel are highly likely to keep the new node for themselves to start with.

Even if TSMC slipped a year it would still be on par with Intel 5nm so Apple would still be better off

I will tell you one thing for sure that it's impossible for Apple to acquire TSMC. TSMC have a lot of customers other than just Apple. I think it's logical for them to come up with their own fab but honestly that is incredibly hard. Maybe in 10 years, I would say.

>I think GPU scaling will be much harder than CPU

We only have mobile SOCs as a reference point so far, but Apple is doing very well on that metric.

>On the GPU side of things, Apple has also been hitting it out of the park; the last two GPU generations have brought tremendous efficiency upgrades which also allow for larger performance gains. I really had not expected Apple to make as large strides with the A13’s GPU this year, and the efficiency improvements really surprised me. The differences to Qualcomm’s Adreno architecture are now so big that even the newest Snapdragon 865 peak performance isn’t able to match Apple’s sustained performance figures. It’s no longer that Apple just leads in CPU, they are now also massively leading in GPU.

https://www.anandtech.com/show/15246/anandtech-year-in-revie...

And what happens if China marches into Taiwan? This is a plausible consideration that must be one of Apple's worst nightmares.

TSMC has fabs in mainland China as well. China is a friend of capitalism.

According to Wikipedia, TSMC's 9th largest fab is based in China. Their 1st - 8th largest fabs are based in Taiwan.

Could this lead to a large chip fab being built in the US? Or do they just hop to Japan / South Korea?

If we're in a scenario where Taiwan is getting bombed or invaded, South Korea and even Japan aren't exactly safe options either.

Their whole supply chain is going to grind to a halt in that scenario even if TSMC's fabs were somewhere else. China would at minimum get sanctioned and there'd be component and raw material shortages for a while.

The United Nations would issue a strongly-worded letter. Markets would fluctuate for a week. Then everyone outside of Taiwan will pretend that nothing happened.

This is why the US just struck a deal to have TSMC build a fab in Arizona. It's absolutely critical to have a backup plan.

Isnt its planned output tiny compared to the production volumes in Taiwan?

Yes. Everything points to it being done for a DoD order.

Having the foundations of a fab makes it a bit easier to expand production in the future compared to setting up the whole shop at once if needed.

China can bomb TSMC, no doubt.

But landing troops (they can't march, it's an island, and the difference between amphibious and land based operations matters a lot) would be extremely difficult, and not obviously in the PLA's favor. See, for example:

https://foreignpolicy.com/2018/09/25/taiwan-can-win-a-war-wi...

China's sheer scale means there aren't too many military shortcomings it can't address on a roughly 10-year timeframe.

The larger point is that it'd be way too messy for them to even try. Basically on the scale of US/SK invading NK.

Besides that significant causalities and economic damage to both sides, which of course they'd eventually win, but it'd be a geopolitical disaster that won't end with the quelling of the armed forces.

China has far more to gain not going this route.

I’d imagine China would more likely begin blockading Taiwan, at least until the US and the west responded. Given the strength of US naval power it’ll be a while before China consider even that.

If China mearly wanted to destroy Taiwan's civilian society, they could do it. But to take out the defences and make an amphebious landing across a large and dangerous sea is extremely hard. Perhaps possibly only with cyber-warefare to disable the defenses.

My understanding is that the actual machineries and materials necessary to make and run fabs are made in US, Japan and Europe, not in China or Taiwan. For example photo-lithography machines are made by the like of ASML (Netherlands), Nikon (Japan) and Canon (Japan).

Although it would assuredly take some time to ramp up, TSMC should be able to spawn fabs outside of Taiwan, out of CCP reach. They are already building one in the US, albeit with a small output.

I heard from random sources that their GPUs are actually (relatively?) very powerful. Better sources/experience appreciated.

"Apple claims the GPU in the iPad Pro is equivalent to an Xbox One S, although how they came to thise conclusion is difficult to say since we know so little about the underpinnings of the GPU." [1]

[1] https://www.anandtech.com/show/13661/the-2018-apple-ipad-pro...

Game console GPUs are middle-end at best. They compensate by huge amount of hardware-specific optimizations since game developer can only target few models of hardware during 5 years lifecycle and devkits for hardware become available for game engine developers almost 2 years in advance.

Apple sales for Mac are much smaller numbers compared to consoles, they change generations more quickly and amount of optimization in GPU-intensive apps is nowhere close to consoles.

So they might be able to compete with Intel iGPU, but that's nowhere close to AMD or Nvidia offerings.

That's less true than it once was. Both that the console APIs abstract way more of the systems than they once did so that console manufacturers can do perf refreshes halfway through the cycle, and that the whole point of Vulkan, et al. is to bring console like programming techniques to full computers since the GPU's MMU means that all you'll do is crash your own process anyway. That all adds up to them needing pretty nice hardware to keep up.

> Both that the console APIs abstract way more of the systems than they once did

And PC graphics APIs abstract way less of the system than they once did

I dont see how this helps Apple to compete with AMD or Nvidia GPUe. It's much easier for Apple to just use AMD GPUs for their Pro devices. Macs market is just too small to create custom high-end GPUs.

Apple might have huge leverage on mobile with iPhone because Android graphics drivers are horrible mess, but on Windows GPU drivers are decent and building custom driver stack sounds like too much.

I mean, it's not like the consoles are very custom GPUs. They were pretty standard GCN last gen, and pretty standard RDNA this new gen.

Apple's custom GPUs they use in their SoCs are arguably more effort on their part.

Why wouldn't Apple just reuse everything, including graphics hardware and drivers from iOS?

I guess because one can't just reuse hardware. They cant just take iPhone GPU, make it 5 times bigger and get 5X more performance with 5X of power budget. So more likely they'll just gonna sell laptops with GPU performance closer to iPhone which is not at all impressive.

GPU is actually relatively easier to scale by adding more cores/ALUs. The difficulty lies in figuring out how to handle peak current requirement of the whole IP.

Still a current-gen console grade GPU in a form factor like the iPad is pretty impressive. But then again maybe it wouldn't be able to sustain console-like performance due to thermal constraints

Agreed, plus benchmarks and measures like TFLOPS are always highly subjective.

I'd be curious if Apple's non-mobile SOC roadmap emphasizes on CPU development while still allowing for eGPU setups or even some built-in integration (e.g. Mac Pro with Apple CPU and AMD GPU) initially. Maybe in 5-10 years they'll shift to focusing on the GPU front and bring out their own dedicated GPU.

So it can compete with an old, lower-end APU using unreleased metrics. Great. Meanwhile, AMD is set to release much more powerful integrated graphics.

I don’t see how an integrated GPU can compete with the top of the line chips from nvidia and amd. The discreet GPUs(Radeon vega) in the Mac Pros have 13+ billion transistors and have 1 TB/s memory bandwidth with specialized memory.

The GPU in the Xbox One S was far from top of the line, it's pretty mediocre. A middle of the road GPU from that time frame was easily 200% faster than it.

https://www.techpowerup.com/gpu-specs/?architecture=GCN+1.0&...

That being said, frankly I'm amazed at what most modern mobile GPUs are capable of and for most people who are casual gamers, that level of performance will be more than enough. What Apple will bring to the table will certainly be better than that, it's already better than Intel GPUs and they can still support 3rd party GPUs if necessary.

The demo they showed of Shadow of the Tomb Raider running at 1080p looked great for a game, AND it also looked awful if you compared it to the PC version on Ultra settings.

Apple kept mentioning pro apps like maya where you need highend GPUs. Gaming isn’t a big use case on a Mac. Our video production team uses highend macs with discreet gpus. If they don’t plan to offer comparable solutions to current line of desktop machines, we’ll be forced to switch platforms. I don’t see how an integrated gpu can compete with a discreet GPU with specialized dedicated ram and a memory bus that’s 10x of a CPU using ddr4x.

I suspect they aren't going to rule out external GPUs, they were just demonstrating the platform capability without them.

The whole presentation struck me as a big middle finger to Intel and just Intel.

Yeah, but that demo was running under emulation through rosetta 2...

The game natively supports Metal, I suspect that's why they chose it.

Yeah, did you notice they also ran it at 1080p and it really didn't look all that great?

I was actually a bit surprised they would break that kind of a demo out at WWDC -- I've seen that game not running in emulation and it was beautiful. That demo wasn't.

I agree the game definitely did not look like it was intended to and from that perspective it looked bad.

Looking at it without context however, it demonstrated the capability of the platform to run a AAA game without developer optimization/consideration.

And it demonstrates someone could make a stylized game similar to Fortnite.

Yeah, I would agree with that. It definitely demonstrated that gaming is possible even under Rosetta, which is an accomplishment. Watching some other stuff from them this morning, it seems like they're passing the metal calls directly through to the GPU (which makes sense), even while translating CPU calls from x86 to ARM.

I mean, it looked awful compared to what I run on my PC, but considering as well that it was running using Rosetta I thought it was genuinely quite impressive.

An A12x has 10 billion transistors, and being a SoC the CPU, Neural Engine, ISC and GPU are all integrated.

And that was released in 2018. By the time Apple put out desktop class silcon in end of 2020, surely it will be significantly scaled up?

Considering xbox is 1/4 the price, it's not saying much about it. On the contrary.

TSMC is too intertwined with Taiwan and Taiwanese independence to make that feasible.

Building your own fab would be a gargantuan task.

A future JV with Intel seems more likely (despite it being very unlikely in absolute terms).

There's no reason why Apple can't use Nvidia or AMD GPUs with an ARM CPU.

Some examples here https://linustechtips.com/main/topic/917482-arm-and-pcie-lan...

I wonder if Apple had anything to do with TSMC's announcement of a fab in the US?

Reply on news.ycombinator.com