Hacker News Reader

>We are in the era of TDP Computing.

Intel has been saying this for years, like almost 8 years. But this is just what non-innovator say. It is an excuse for mediocrity and failure and stagnation.

There is tons of uArch improvements in Apple's chips. That is why they consistently beat even the other ARMs chips on the same process nodes.

AMD also has been significantly innovating with uArch.

I love the innovation we are seeing these days. We are definitely not in an TDP era (where only TDP matters! but of course TDP is important), we are in another golden age of CPU/GPU innovation and it is beautiful. Reminds me of the 1990s and early 2000s.

Don't they beat others because they are almost always a process generation ahead like now with 5nm?

At some point, the process can't get any smaller. Physics has hard rules. Once we get to that limit, improvements will have to come from elsewhere.

We are butting up on the physical limit, but to be fair, that’s what people have been saying for over a decade, and we keep going.

According to Wolfram Alpha, 5 nanometers is ~23 silicon atoms across (~2.2 Å each). But that’s also assuming they’re actually fabricating 5 nm gates - which they’re not.

Like if you look at the rate of down scaling it has slowed some. It's also not the size of the whole transistor it's the size/thickness of the N, P and Insulator layers that will limiting factor in FET style transistors.

The oxide/insulator thickness on modern process is pretty insane. For instance for the intel 14nm++ the silicon nitride and silicon oxide layers seem to 4 to 5nm thick.

-Edit- For instance here is 2 fins for from the intel 14nm++. The grains/spheres you see this image are the atoms. Some parts of the a modern transistors are even smaller than process number may let you know.

https://imgur.com/a/hgeSOVV

Although Apple has been leading edge for a while, it's actually rather unusual for Apple to buy the entire production capacity of a new node.

The A13, Qualcomm Snapdragon 865, and Mediatek Dimensity 1000L were all on DUV N7P (Hisilicon Kirin 990 5G was actually on the more advanced EUV N7+). For more information on chips and their processes: https://www.eetimes.com/apple-huawei-use-tsmc-but-their-7nm-...

Here are some benchmarks of how the A13 and 865 (and the 2020 updates) stack up: https://www.phonearena.com/news/apple-a14-vs-a13-vs-snapdrag...

From the charts, it looks like particularly for single-core and browser/JS performance, the Apple designs are very strong, and beat out the Qualcomm competition even on the same process node.

While I'm disappointed by how Intel is executing over past 4 years, it's still impressive how they manage to increase performance on their 14nm node.

You're absolutely right. People already propagating Intel's death are too fast. They may do not the innovation we like, but it maybe will help them push their smaller nodes, too. And for many special cases and for a suckless Linux experience on new devices, Intel is still the best choice.

AMD beats Intel basically everywhere you need an x86-64 architecture. The latest CPUs beat Intel even on single core performance.

Intel is basically running a super legacy design and they are doing surprisingly well with it, but they are losing on all fronts at this time. What they have for them right now is momentum and enterprise deals.

That will hold up for a while yet, so it gives them time to get their act together.

Intel still beats AMD for idle (or lower load) power consumption (important on mobile).

While you may well be correctly reading between the lines, you are reading between the lines -- the article doesn't say anything about future "Mac Chips," as the headline puts it, definitively being SoCs. There's nothing about a unified memory architecture that requires the GPU and CPU to be on the same die, let alone requires the RAM to be in the same package as both processing units. (People may no longer remember that Silicon Graphics was using UMA in their workstations a quarter century ago!) I expect all their portables, and probably all the iMacs to be SoCs, but I'm not sure the Mac Pro -- and the iMac Pro, if it continues -- will follow that path.

> We are back to TrashCan Mac Pro again... we would have reached the thermal ceiling again...

This is kind of drawing new lines to read between. :)

Apple almost certainly developed their current uber-expensive, trypophobia-triggering Mac Pro design knowing that they were also working on Apple Silicon processors. While it's possible they intend that whole case design to be a one-time thing, it doesn't seem likely to me -- and if that wasn't the intent, then it's also not likely they want to ship a teeny tiny Apple Silicon logic board in that huge case with the rest of the space going unused. I think they're going to do their best to make an Apple Silicon-based Mac Pro that can be upgraded and expanded. (With all respect to Mark Gurman, I think that "half-sized Mac Pro" is going to be a different product entirely if it ships, with a different branding. Mac Mini Pro?)

>With all respect to Mark Gurman, I think that "half-sized Mac Pro" is going to be a different product entirely if it ships, with a different branding. Mac Mini Pro?

The ever illusive xMac, aka "the affordable Mac tower"

Somebody once said that that Apple’s business model is charging a very high markup for flash memory (currently it’s $50 per 64 GB to upgrade storage on an iPhone or MacBook). The margin on other parts is also high. The user-upgradeable Mac Pro provides an escape hatch from these expensive upgrades. Therefore the entry price has to be much higher to make up for it. An affordable tower would cannibalize Mac Pro sales by providing businesses with a cost-saving opportunity that’s too good to refuse.

One does not simply create a $2trn business by selling flash memory.

This is good analysis, but there's some other effect I'm unable to put my finger on.

The iPad is at least an order of magnitude better than comparable netbooks (though chromebooks, depending on manufacturer, can be competitive thanks mainly to ChromeOS's reduced footprint). So even though the margins are high, the perceived quality, regardless of raw benchmarks, is still something. It's not just marketing to me to say the marriage of software and hardware is unique. (Gruber's observation about NSObject alloc's being a lot faster on Apple Silicon, for instance).

Now that I think about it, honestly Google is the only other company playing by these rules... ie pixelbook, pixel phone, etc. But they're much earlier in the evolution, and have less upstream control in software (especially since Fuchsia seems to be somewhat lower priority than before, though this is second hand knowledge).

Adjust reference counts on an NSObject, allocating one is mostly just a calloc ;)

That somebody is probable only partially right about a partial business unit within Apple. People tend to buy it for the package deal, not for individual specs. At volume, those are the much more valuable customers than the 'buy a single computer' crowd that often talk about their modifications they want to make.

Exactly - Apple "sells" experience, not parts or specs.

Considering how the M1 SOCs are, the possibility of a future "affordable Mac tower" looks more remote than ever. Apple seems to be aiming towards piling as much into their dedicated SOCs as they can. I expect what "pro" capabilities they do build for it going to be in the vein of the cool stuff they were doing in the Mac Pro, like the FPGA "Afterburner" and will probably cost way more than any person with generally "consumer-level" use-cases would want to spend.

As a technology enthusiast I'm excited to see what bonkers stuff they'll be announcing. But, since I am not doing climate models or rendering hollywood movies, I don't expect I will never work on any of them since I don't want to drop a new car's worth of money on one.

The "Mac" is what I want.

> There's nothing about a unified memory architecture that requires the GPU and CPU to be on the same die, let alone requires the RAM to be in the same package

No, but unified memory does require starving the GPU of bandwidth. There's a reason high end GPUs use different types of memory than CPUs.

Either Apple will be content with Mac Pro not having the absolute graphics performance lead, or they will make a discrete GPU with dedicated memory. I'm not sure which they will choose but I suspect the latter.

I agree, i'd also expect more integration of the Afterburner tech with support for multiple codecs beyond the current version.

The Xbox Series X (retail price $499) uses a 7nm SOC, where the GPU has 12 TFLOPS of single precision performance. The fastest single GPU Apple sells on the Mac Pro is 14 TFLOPS (with a step-up fee of $2,400!).

There doesn't seem to be any reason why Apple couldn't integrate a 16 TFLOPS GPU on an SOC using a 5nm or better process. The M1 GPU is 2.6 TFLOPS, they can x6 or more it, and obviously change the memory subsystem (HBM?).

Xbox Series X is very far from the top graphics available. GTX 3090 at 35 TFLOPS is what Apple needs to compete with for Mac Pro. Those chips are reticle limited which means there's simply no room to add a CPU, let alone 32 cores. You need separate dies.

Also, Xbox uses GDDR which trades off CPU performance for GPU performance. Apple probably wouldn't do that for a Mac Pro. The high end needs separate memory so the GPU can have bandwidth and the CPU can have low latency.

Well, depends. If they're looking to replace the low- and medium end of the Mac Pro offerings, a SOC would be fine. If they are looking to exceed it in every way, then a separate GPU with 16x the current M1 GPU cores is necessary. (The GPU on M1 takes up about 30mm2-40mm2). Put that on a TSMC N5P process, and that will be quite the setup...

Reticle limited on Samsung 8nm means plenty of room is available on TSMC 5nm.

Also, GDDR does not have appreciably higher latencies than DDR memory when measured in nanoseconds. It's just more expensive than DDR and much more limited in terms of capacity.

If Apple does go the "huge SoC" route I'd expect to see HBM2 memory with socketed DDR4 or DDR5. It'd provide the best of both worlds - extremely high bandwidth and low latency for a small portion (say 32-64GB) of the memory, and high capacity for the rest (say 1-2TB), all without compromising the unified memory concept.

This is not without precedent - recent Xeon Phis, for all their other shortcomings, have had a similar memory hierarchy.

Yeah but when discrete GPUs get to 5nm they will still be reticle limited. They're not going to stop getting bigger. If Apple maintains a process node advantage maybe they could do SoCs to compete with discrete graphics at the previous node but I think Nvidia will catch up.

I agree that a combination of HBM and DDR sounds pretty good for a unified memory architecture. Are you imagining it as just another layer in the cache hierarchy or something actively managed?

How is it a unified memory architecture if one part is faster than another? Sounds rather non-unified.

"Unified" in this case refers to the CPU and GPU having the same access to memory. This could still be true even if part of the memory is faster, as long as it's faster for both CPU and GPU.

Is there any reason the Mac Pro couldn't or shouldn't combine an SoC with upgradable components?

So you get a certain baseline level of specs, but also have the opportunity to tack on a beefier GPU (which may or may not use UMA), additional slower memory, additional slower storage, maybe even additional CPUs, etc.

> Is there any reason the Mac Pro couldn't or shouldn't combine an SoC with upgradable components?

I don't think so, with the asterisk that I don't know enough to say whether mixing higher-speed and lower-speed RAM is really feasible in that scenario. (IIRC, if you do that in computers now, all the RAM will operate at the lowest speed.)

> if you do that in computers now, all the RAM will operate at the lowest speed

That’s correct, but irrelevant in this context. In that scenario, the memory is all on the same memory controller, and the memory controller used wants to run a single speed, so it has to be the highest universally supported speed.

Apple could have multiple tiers of memory with different characteristics, but just because you can, doesn’t mean you should. I don’t have access to such a test bed, so I can’t speak to what the experience would be like.

One thing to note here that's a bit ambiguous in your comment - an M1XX with 128 GPU cores would still have substantially fewer GPU cores than the 2x Radeon Vega II Duo.

The Radeon Vega II Due has 128 compute cores and 8192 stream processors, so that maxed out mac pro would have 256 compute cores and 16384 stream processors. I don't have any benchmarks to work off of, but I think it's safe to say that a new Mac Pro with the 32-core SoC would still be substantially far behind a current Mac Pro unless it also had a dedicated GPU.

I believe GPU core count between different products isn't comparable. Comparing TFLOPS would be better than core count.

Assuming linear scaling (which is not a good assumption), Apple would need 6x the number of GPU cores in the current M1 to get in the range of performance of the Radeon Vega II Duo.

If they stick with UMA, gpu performance is going to be severely limited, unless they can make the unified memory GDDR6 or something.

Why couldn’t they? The 3090 ships with 24GB of GDDR RAM today.

A future Mac Pro should support 1-2 TB of RAM which isn't really possible with GDDR.

Given the premium Apple charges for RAM, I don't even want to "napkin calculate" how much a 2 TB ram Mac pro would cost!

It was £200 to go from 8GB to 16GB which makes 2TB ... £51,200 or thereabouts.

Well, on the current Mac Pro, going from 32GB to 1.5TB is $25K, but you also have to upgrade to the higher spec CPU @ $6K.

I would love to know as well. I think it could be HBM2, something similar to Fujitsu A64FX [1], the problem with both GDDR6 and HMB2 is that are limited in capacity to ~64GB, or at best 128GB. Which is far from the 1.5 TB of memory the current Mac Pro is capable of.

[1] https://www.anandtech.com/show/15885/hpc-systems-special-off...

What if you gave the GPU its own dedicated memory again?

Then it would make sense to have the two chip separate and not on the same Die. Which is essentially what the current PC system are doing. The whole hypotheses is built on Mark Gurman's suggestion of Mac Pro using something similar to SoC like the current M1.

Why? Clearly, I'm not a chip designer, so maybe it's a dumb question. If you can place GPU cores, CPU cores, Neural Engine, and RAM on the same die, why can't you then use a second bank of RAM dedicated for the GPU cores? Why would this not be considered similar?

There's nothing stopping a chip from having HBM and, say, DDR5 DIMM slots but that has only been done once before by a doomed Xeon Phi.

It's more likely that the future Mac Pro will use Lifuka which everyone in this thread is ignoring.

Xeon Phi is actually a pretty nice system to work with but as picky about performance as a GPU while more expensive and harder to find. Combining regular tool chains for code with high core count and SMT on top of AVX512 was nice. 68 cores, 4 way SMT, and 16 way SIMD, provides 4k “threads” in the GPU sense, except the memory hierarchy isn’t nearly as harsh because of the MCDRAM.

Because Lifuka; the GPU, hasn't had any other source other than the first Taiwanese publication. And those publication has been known to pop up all kinds of rumours only to pump up TSMC's stock price. And has relatively poor track record of any Apple information, or as a matter of fact, any information.

It would be interesting if Apple did make a discrete GPU, that would means they are making a separate GPU die for their Mac Pro and iMac with volume estimated to be no more than 1M unit per year. It makes sense from a technical perspective, it surely doesn't from a cost perspective. Although Apple could happily charge $1000 just for the GPU to make up for ROI.

Yeah, it worries me a little that Apple thinks they can charge anything for the Mac Pro and probably iMac Pro. $10,000 starting price anyone?

Whether CPU and GPU share ram is orthogonal to whether the gpu and cpu are on the same die. You can do any combination of same die same ram, different die same ram, etc etc.

The advantage of GPU having it's own memory is you can use super high throughput ram that the GPU needs, while using higher capacity/lower energy/cheaper ram for the cpu.

The downside of using separate ram is that you have to shuttle data back and forth from one ram to the other quite a bit, as well you get less flexibility in how your total memory is used.

This is interesting.. is the M1 chip's GPU a tile-based chip similar to other mobile/ARM SoCs? If so, that could be a problem for truly higher end graphics work as certain operations such as real time shadows, fullscreen shaders, etc are less efficient when applied on such a GPU

I wonder if for bigger, desktop macs they will stay with ddr4/ddr5 memory sticks or maybe they will introduce something custom. I assume they won't go with mac pro with soldered, non-removable memory

Does UMA (unified memory architecture) prevent anyone from also adding additional conventional RAM too (apart from what is integrated in the CPU)? Have today's SSD reached RAM's read / write speed?

UMA has nothing to do with the RAM sitting next to CPU, which is what half of this thread is getting mixed up with. They are two different thing. So yes you could have UMA and conventional DRAM, the problem is that there are no DRAM, even with 8 Channel DDR5, capable of feeding enough bandwidth for the hypothetical 128 Core GPU.

That is unless Apple decide to use DRAM in additional to HMB or GDDR5 as 2nd layer of Memory.

So is Apple's UMA similar to the nearly 2 decades old SGI O2 UMA - http://www.sgidepot.co.uk/o2/1352.pdf ?

Apple is probably going to be first on TSMC's 3nm which will go into volume production in 2H 2022 so perhaps they can hit 300W easy.

Well, in terms of dollars per getting shit done, a GPU on ARM that only supports Metal 2 isn’t going to get any software for it in a really really long time, no matter how fast it is, perhaps so long that better conventional GPUs will come out by then.

Who do you think writes all the code for this stuff? It’s guys in tracksuits in Visual Studio C++ on Windows. The top games for iOS are also either written by a few guys in Unity or by an absolutely huge team of people porting an existing, low graphics quality PC game.

A lot of software will end up using OpenGL->Metal, Vulkan->Metal, or D3D->Metal translation layers but it won't matter if the GPU is fast enough.