Hacker News Reader

Comment by LorenDB | Hacker News Reader

LPCAMM is the best thing to happen in the laptop space since Framework. Forget Apple's ridiculously expensive soldered storage and memory; LPCAMM is a step in the right direction.

View on HN

That was true when they had Intel CPU's and soldered the RAM chips on the board.

Now they have the memory chips "inside" the "CPU", aka: system in a package.

Big Tim is way ahead of the game when it comes to ruthless product targeting and segmentation.

Apple's ridiculously expensive fast & wide on-package memory is a big reason why the M series chips are so fast. Expect AMD & Intel to adopt a similar strategy in their high end chips soon.

Intel is already there at the ultra-high-end, the Xeon Max is a hybrid with 64GB of HBM2 on-package and an external DDR5 bus supporting up to 4TB of socketed RAM. The HBM2 can be configured to work as L4 cache for the DDR5, or it can boot without any DDR5 at all and run entirely from the HBM2.

https://www.servethehome.com/intel-xeon-max-9480-deep-dive-i...

MI300X also has on-package RAM, but that's even further from a laptop chip than Xeon Max. :)

Intel's Lakefield also comes to mind, pretty underrated chip imo. The first x86 big/small heterogenous chip, using Intel's Foveros to stack memory, CPU, and platform chip into one package. It made for a very simple design that just needed power. LPDDR4X-4266 doesn't crazy fast but in 2020 it was pretty swift. https://www.anandtech.com/show/15877/intel-hybrid-cpu-lakefi...

Alas the 1+4 Skylake+atom were all pretty meh chips & Intel hasn't tried again. Conceptually I really.liked it, and it was doing great against the Qualcomm 8cx competitor for the tablets/2-in-1s it was in, but people wanted more.

I remember seeing this from STH as well, and my first thought was "I want one of these in my laptop!" It'd be perfect then to max ram in a laptop with 2 sodimm slots, or one of these 64gb lpcamm's. Ideally with a thinner laptop in mind, but probably not with the heatsink for THAT massive xeon.

I really hope the hbm memory on-die trickles down into pro-sumer processor chips eventually too.

> Apple's ridiculously expensive fast & wide on-package memory is a big reason why the M series chips are so fast

Increasing memory speeds have had only a marginal impact on common workloads (not AI) in the PC World - this has been a well discussed topic. Why would it be different on the Mac?

probably something to do with rounded rectangles

Memory speed usually refers to memory bandwidth, which does have a limited amount of impact on day to day apps as they don't usually require a lot of bandwidth.

Memory Latency, on the other hand, can have a huge impact on performance. That's the reason chips have had steadily increasing cache sizes over the decades along with the introduction of additional caching layers.

What Apple's M series improves by putting the memory on chip is both bandwidth and latency. The latency, however, is what will impact app performance more than anything else.

Latency improvements are the reason memory controllers were moved from motherboard northbridges into the CPU IC. Shortening that distance means a lot to shortening latency.

Amusingly the top-line latency number is higher than for X86, but I agree with you because that number hides the better performance in other latency metrics that are probably more important.

It's not about memory bandwidth. [1] shows that M1 is 66x faster at pointer chasing than an equivalent AMD. You can look at [2] and [3] to see that TLB penalty is also lower at any given test depth which is important because TLB misses aren't that infrequent and they're serial in your ability to access any memory that's not in your cache.

It's not an unreasonable hypothesis that this is because of on-package memory which means you can run with tighter thresholds since your memory traces are shorter. Their unified memory architecture also means that GPU is way more effective since you never need to copy memory from the CPU to GPU. That's probably not a huge deal for normal apps, but having rendering the screen not competing with any app resources for memory probably isn't nothing.

Neither the CPU nor the GPU can fully utilize the available memory bandwidth [4] The insane memory bandwidth is to also support the other coprocessors like the media engine & to ensure that you can do a bunch of tasks in parallel without them contending with each for resources. Those workloads do benefit greatly from memory bandwidth & the "artist" community is hugely important to Apple.

The reason M is overall faster is more because of things like better ILP (e.g. fixed-width RISC encoding vs compression-like CISC means easier to keep the pipeline fed). Apple software for the most part is better engineered for performance & the M chip is optimized for that performance profile (e.g. C# and Java vs Swift/ObjC - Rc is pretty conclusively faster & requires less memory overall not to mention that the M chip has specific optimizations to further reduce the traditional cost of Rc that they have to pay because they don't typically distinguish Rc from Arc as Rust does). The final truth is that Apple throws a lot of money at this problem as well & buy up the latest processing node to stay ahead of the competition.

[1] https://www.linkedin.com/pulse/apple-m-cpus-probably-much-fa...

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

[3] https://www.anandtech.com/show/17047/the-intel-12th-gen-core...

[4] https://www.anandtech.com/show/17024/apple-m1-max-performanc...

No. M1/2/3 has data memory-dependent prefetcher[0] which can speculatively chase pointers in contiguous memory ranges, reducing indirection cost provided the data there is homogenous, looks like pointers and prefetch attempts are successful per cache line. It also has extra good indirect load predictor.

These have little to do with memory packaging and a lot to do with cache and prefetching architecture.

It does have one of the lowest latencies for atomics[1][2] which is something Swift cares about, but the overall core design just as much benefits other, less indirection and synchronization heavy languages (side note: Swift and Java are closer with each other in defaulting to virtual dispatch, where-as C# sits in the middle of the road with static dispatch by default but some code heavily using interfaces and virtual methods which do make such calls virtual to an extent (JIT has DPGO to devirtualize, AOT can uncoinditionally devirtualize in certain scenarios too, similar to what Swift's WMO does)).

[0] https://gofetch.fail/files/gofetch.pdf

[1] https://dougallj.github.io/applecpu/measurements/firestorm/D...

[2] https://dougallj.github.io/applecpu/measurements/firestorm/C...

Yeah. But even as a NUMA area, upgradability would win a lot of brownie points here (of course modern Apple does not give a flapping bird about it)

Not sure how it works on the iMac/MacPro

Would love to see a Mac Pro with 8 of these…

Soldered? It's not soldered. It's part of the actual SoC. Could we please update our research to more recent than 20 years ago?

The switch to ARM Macs only happened 3.5 years ago, plenty of Intel models still in service that would benefit from a RAM upgrade if it weren’t soldered

The ram is soldered into an Intel Mac for the exact same reason that it's soldered into an ARM Mac.

To make upgrades incredibly expensive, allowing them to have the same product generate revenue as both an expensive mid-range machine as well as a ridiculously expensive "professional" machine. Why make different products when you can sell the same product at different prices?

In case of intel macs, the soldered RAM should have made the device and RAM options even cheaper, as the BOM is simpler. The only real argument against soldered RAM was DDR5 in SODIMM being bad - which LPCAMM2 fixes.

If they didn't use it specifically to price gouge, the soldered RAM wouldn't have been a problem. 64GB of RAM costs practically nothing at market prices, so they could honestly have had a single SKU with max RAM without notable price change over entry level - heck, maybe the BOM reduction would even sponsor it.

... But then they'd lose their mechanism to drain companies with deep pockets for necessary upgrades. Being sensible and fair is not profitable.

It's soldered to the package. The actual RAM components are the very same discrete BGA ICs that would normally be soldered to a motherboard, but they've soldered them right next to the SoC die instead for tighter signaling.

https://i.imgur.com/Y3PLp33.jpeg

Anyway results are what matters, and the module in the OP has bandwidth figures half way between the M3 and M3 Pro while retaining upgradability.

Reply on news.ycombinator.com