Apple Silicon and DDR5

[caribbeanblue]

If I am not mistaken, official fastest DDR5 is DDR5-6400. I don't think Apple ever used "overclocked" RAM. In the end, it's abut energy efficiency and volume. DDR5-8400 is certainly great, but it's of little use to Apple if it is only produced in super low volume (like current "gaming" RAM) and uses extreme amounts of power.

With LPDDR Apple should be able to reach the fastest speeds DDR5 can achieve at much lower power, as that is the whole point of LPDDR. LPDDR5X modules can go up to 8533MT/s per pin. Also this whole thread aged well with the M1 Pro/Max/Ultra/M2. Hopefully LPDDR5X is ready in time for M2 Pro/Max (maybe even Ultra) by the end of this year (as recent rumors are suggesting) and we'll get a 33% boost in memory bandwidth over LP5-6400 M1 Pro/Max/Ultra.

 

[caribbeanblue]

DDR5-6400 offers 51.2GB/S PER 64-bit module. Hynix mentioned about a year ago that they plan to ship DDR5-8400 (67GB/s per module). Recently ratified LPDDR5X supports similar bandwidth. I think it's unlikely that 8400 or higher DDR5 will be in production by the end of the year, but LPDDR5-6400 should be realistic.

This whole thread aged well.

 

[caribbeanblue]

DDR5-8400 is already "overclocked". Besides, Apple must deliver good power efficiency (they do it by sourcing custom-built RAM modules that consume less power than even the usual mobile RAM), and they need a supplier that can provide reasonable volume of these chips. The super-fast versions of DDR5 are still miles away I am afraid.

Not as far away as you think. Nvidia's new 'Grace' server CPU that's going to come out in 2023 is already announced to use LPDDR5X memory at 1TB/s (probably very similar to the bandwidth that the M2/M3 Ultra will have if it also uses LPDDR5X with its 1024-bit memory channel). If Grace also has a 1024-bit memory channel that means it'll drive its LP5X memory at 8533MT/s.
And then a year or two later LPDDR6 will come around and blow it out of the water with a whopping 17,000MT/s throughput. And LP6X will go to 24,000MT/s.

 

[caribbeanblue]

Overclocking the DDR5 8400 by the same multiple as the current 4266 modules in the M1 yields 11198.25, which I'll round to 11200.

Running my earlier numbers with 11200 as the speed outputs 358.4GB/s!

Imagine then, a 32 core GPU built into an M1X. If we assume linear scaling, the result is 82324 points in Geekbench 5, beating the Radeon Pro W5700XT by a respectable margin, but this is theoretical, and the M1's RAM config might hobble it to far below that.

Now two generations ahead, with a conservative ten percent increase over said generations, overclocked DDR5 and 64 cores, the result is 181112 points. For context, the Radeon Pro W6900X scores 168783.

358.4GB/s could feed such a theoretical GPU.

It would've gotten 82000 points in GB5 if GB5 let the M1 Max properly boost to its peak GPU frequency, as the M1 Max finishes the tests GB5 throws in front of it before its GPU reaches its peak frequency. This is thought to be due to a power saving 'reverse Turbo Boost' mechanism in firmware. Scores from other GPUs don't suffer from this as they don't have as aggressive power saving mechanisms and they try to just go for maximum performance.
Also, DDR5-8400 is the maximum speed DDR5 will be able to achieve by the end of its lifetime, when DDR5 matures enough, with realistic cooling and power solutions. So the 'LPDDRX' versions of DDR modules that Apple chooses to use will probably not go much above that. The fastest LPDDR5X modules out there will have a throughput of 8533MT/s per pin, which is a 33% boost over the LP5-6400 the M1 Pro/Max/Ultra/M2 uses. Though, LPDDR6 will come to market in around ~2024 or so, and that will be able to go up to 17000MT/s. And the LPDDR6X will go up to 24000MT/s.
An M3/M4 Ultra with LPDDR6-17,000 memory would have a peak theoretical memory bandwidth of 2.2TB/s, almost as fast as the UltraFusion interconnect that connects the two M1 Maxes of the M1 Ultra.

 

[leman]

Not as far away as you think. Nvidia's new 'Grace' server CPU that's going to come out in 2023 is already announced to use LPDDR5X memory at 1TB/s (probably very similar to the bandwidth that the M2/M3 Ultra will have if it also uses LPDDR5X with its 1024-bit memory channel). If Grace also has a 1024-bit memory channel that means it'll drive its LP5X memory at 8533MT/s.

Very much possible, but isn't that a custom supercomputer stuff anyway? If they are not sold on general market (or only sold in limited quantities), the volume will be low enough so that they can use all kind of stuff. Apple has to rely on what is available today in bulk. Besides, I wrote that post almost a year ago and now you say "by the end of 2023" — I surely hope that faster RAM will be available three years after I wrote that

 

[caribbeanblue]

Very much possible, but isn't that a custom supercomputer stuff anyway? If they are not sold on general market (or only sold in limited quantities), the volume will be low enough so that they can use all kind of stuff. Apple has to rely on what is available today in bulk. Besides, I wrote that post almost a year ago and now you say "by the end of 2023" — I surely hope that faster RAM will be available three years after I wrote that

I'm aware of that, but still, you said it like we won't reach DDR5-8400 until 2027 or something, when in reality OEMs are planning to go much above that by a lot earlier. Yes, it might still take a while until LP5X, LP6, etc. are ready in the quantities that Apple needs them, i.e first smartphones with LPDDR5-6400 came to market at the beginning of 2020 while the iPhone 13 Pro still doesn't have LP5, but that probably also has to do with chip shortages, so you can't extrapolate what will happen in the future from what happened in the past 2 years. Also I'm sure the higher-end Macs that will use LP5X won't be that high volume either.
The first smartphone SoC that supports LP5X-7500 (MediaTek Dimensity 9000) came out at the beginning of 2022.

 

[leman]

I'm aware of that, but still, you said it like we won't reach DDR5-8400 until 2027 or something

I wrote "by the end of the year [2021]" in the post one above you quoted. But doesn't matter, it was easy to miss.