Hacker News Reader

Comment by Tloewald | Hacker News Reader

I can't believe you can't believe it. Have you heard of Apple before? If anything I'm a bit more perturbed that they max out at 32GB of RAM. My aging Mac Pro has more than that in it.

I imagine that they preferred shipping with a default 8GB and selling a $200/600 RAM upgrade than having a higher headline price.

(Incidentally, their 1GB phones still outperform phones that are one year newer with 2GB of RAM, double the cores and nearly double the GHz, so maybe they're onto something.)

View on HN

32GB is a limitation of iX Haswell, Ivy Bridge could use 64GB - Apple can't do anything about it. Let's hope Intel allows it for Broadwell-K (-Y,-U,-H variants will be restricted to at most 16GB)...

It's not a Haswell vs. Ivy Bridge distinction, it's a consumer vs. server distinction. Consumer CPUs only have dual-channel memory controllers and Intel's DDR3 controllers can't use 16GB UDIMMs or have more than 2 UDIMMs per channel. The server chips (a handful of which get some features cut off and sold under the i7 brand) have 4 memory controllers and also support RDIMMs and LRDIMMs.

>Intel's DDR3 controllers can't use 16GB UDIMMs

That is a software not a hardware limitation (in particular, I think it is in Intel's MRC). ASUS has firmware with modified MRC code that can support it.

That is in fact incorrect. Consumer desktop Sandy Bridge, Ivy Bridge and Haswell all support 32 GB of RAM via 2 DDR3 channels. Enthusiast desktop processors (Sandy-E, Ivy-E, Haswell-E) support 64 GB, because they feature 4 channel memory controllers.

Once the consumer chipsets move up to DDR4, you will have support for 64 GB of RAM on the desktop. http://en.wikipedia.org/wiki/Skylake_%28microarchitecture%29

Haswell (as opposed to the pseudo-Xeon Haswell-E) caps out at 32GB.

I don't think this is supposed to align against Mac Pros (plus what people are already saying with the chipset memory limitation)

It seems to me that a top-of-the-line consumer machine should be able to compete with a 3 year old workstation (that hadn't received a significant update in two years -- effectively 5yo hardware).

What 3 year old workstation are you talking about?

Read my original post.

Ok - so just an arbitrary old machine you picked. Somehow I thought there was more logic to the choice of benchmark that I was missing.

Why aren't you comparing your old mac pro with a current mac pro?

12GB of RAM (up to 64GB) https://www.apple.com/mac-pro/specs/

The mac pro line always had xeon processors.. so it seems odd that you would expect the same memory support on a desktop.

Thanks for the info -- didn't know Haswell has that limit. (I guess that's one of the benefits of buying pro stuff.)

For what do you need more then 32G RAM ever on a desktop (excluding time travelling forwards). If you are asking yourself that question you need to get a proper server or rendering farm.

You'd be surprised how fast you can run out of memory if you are mastering your own, non-professional 5k/4k video and applying some advanced non-linear operations such as video stabilization. It's not GPU accelerated, parallelism is limited by the nature of the algorithm which makes rendering farms useless. Do you need a power hungry and noisy server with slower single thread performance than a top end i7 just for additional RAM?

On the other hand, beyond 32GB, ECC seems to be very important so you probably won't have any choice but to buy a proper workstation unless AMD makes a miraculous processor on par with i7 (as they do support ECC in all models).

> non-linear operations such as video stabilization

What do you mean by non-linear? Are you referring to an equation containing a linear combination of the dependent variable and its derivatives, or something else?

Usually a non-linear video operation would mean one that doesn't just process each frame of the video beginning-to-end, like moving clips around.

Stabilization seems pretty linear to me, though.

If you didn't need the 5k display but you did need more memory and rendering capacity, the Mac Pro has you covered maxing out at 64 GB of RAM or 128 GB is you use non-Apple RAM upgrades. So there's that.

Virtual machines behaves best with as much ram as a normal machine. Assuming you have 3 machines running, osx host and windows+linux vms and you want to give them 12gb each you are already at 30+. Might sound like a crazy setup but some times a compiler is only supported on one platform and then you want a replica of your clients server in one VM etc etc.

No, not really. I work in a games studio and our usage on individual workstations goes above 32GB when compiling our project. Due to technical and licencing issues it's also not something that can be relegated to a remote server farm for compilation.

Im guessing this is C++ with a lot of template abuse and header interdependencies? C++ has a powerful compiler, buts its also easy to shoot yourself in the foot with it.

Yep. Also the fact that the project cross compiles on 7 different platforms from the same code base.

If it's that gigantic, then a distributed build on local servers makes a lot of sense. Much faster than each developer doing it on their own single workstation.

We do distributed builds on local workstations across the studio with Incredibuild. That's the only way we could get away with licencing for consoles. That reduces the build times(from over 40 minutes to less than 5) but the RAM usage is still very very high.

Reply on news.ycombinator.com