Hacker News Reader

Comment by api | Hacker News Reader

How much effort is it really? Make a version of this with more cores and cache and maybe all high-performance cores instead of heterogenous. It's just a remix of the same already designed components.

The only really tough thing would be supporting things desktop users want like external video cards, but people may be willing to forego that if the CPU smokes everything else and if the built-in GPU is at least fast enough for serious work.

View on HN

> Make a version of this with more cores and cache and maybe all high-performance cores instead of heterogenous.

Just a side note. It's unlikely that Apple will eliminate their "slow" cores from their desktops. Mac Pros spend big chunks of their time doing mundane things which their efficiency cores do just fine. By using their efficiency cores, they can bring down total power requirements and their desktops can essentially run continually as opposed to entering a dedicated sleep more.

I suspect the Mac Pro will sip power and run cool as a cucumber most of the time when it's not getting hammered for high end tasks.

I don’t think it’s that simple. This design is so good because Apple designed a system, and didn’t throw together a few parts they could buy.

They should revisit all choices they made earlier. Users might want more generic ML hardware, so that they can develop the algorithms that can get burned into specific hardware on some future consumer-level hardware, a GPU that can ray trace better, more memory than you can fit on a SoC, etc.

That last issue is particularly problematic. They might have to give up the idea of having (only) unified memory.

If so, the entire design could change. In some cases, that could mean task-specific hardware isn’t (much) faster anymore than doing it in the CPU, in which case discarding it to make room for extra cores, cache, etc. might be the better choice.

It’s a focus and vertical integration thing. Is the market big enough to justify the increased costs of SW and HW development (not just direct incremental cost but also indirect costs like more support staff, office space, etc). Then are the developments you make to support that market transferable in some way. Are there techniques you prototype in the more expensive/less power-constrained space that help your other markets in the long term?

Im sure Apple could produce something in the HPC space. The question is whether it makes financial and/or strategic sense to compete there.

Probably mostly this, but I would wager more work goes into providing more PCI lanes because I/O tends to be more demanding in pro workloads (connecting all kinds of things)

I'd be surprised if they didn't still include 2 or 4 high efficiency cores to continue touting power efficiency - the business customers would love to cut their power bill in the long term.

What you’re describing is the iMac, iMac Pro, Mac mini pro, etc.

But I think they’ll have the Mac Pro continue to be RAM, GPU and PCI upgradable.

How they are doing that is the most interesting secret in the industry.

There's not really any magic to it, they just need to implement PCIe and include memory slots instead of hardwired chips on the package. There are other ARM systems that already do this.

Yeah, but then you lose the benefits of the integrated memeory.

How they're gonna solve that is the big question.

There actually aren't really any benefits outside of manufacturing. Trace length is a marginal factor in memory performance/latency, and having the memory on-package versus only slightly further away doesn't mean much. I'm not sure exactly where the idea that having the memory on-package was so beneficial came from, especially considering the M1 actually has fairly poor memory latency (though within the expected normal range for LPDDR4X):

https://www.anandtech.com/show/16252/mac-mini-apple-m1-teste...

As far as HSA/unified memory architecture, PCIe 4.0 has enough bandwidth to comfortably access main memory for applications that support it (and vice versa with GPU memory), although I imagine most high-performance GPU workloads will do better primarily operating on dedicated GPU memory (especially given that a single M1 core can already saturate the memory controller on its own).

This is not my area of expertise, but I’ve seen concerns of implementing the unified memory architecture off chip. But maybe it’s just that it hasn’t been done yet.

That doesn't have any impact. The memory isn't on-chip, it's just on-package. This makes it no different from if it was fully external.

This has also not been a problem for any other unified memory system (which typically also do not include memory on-package), and I can't see any reason why it would matter.

Reply on news.ycombinator.com