I definitely wouldn’t mind soldered RAM if there’s still an expansion socket. Solder in at least a reasonable minimum (16G?) and not the cheap stuff but memory that can actually use the signal integrity advantage, I may want more RAM but it’s fine if it’s a bit slower. You can leave out the DIMM slot but then have at least one PCIe x16 expansion slot. A free one, one in addition to the GPU slot. PCIe latency isn’t stellar but on the upside, expansion boards would come with their own memory controllers, and push come to shove you can configure the faster RAM as cache / the expansion RAM as swap.
Heck, throw the memory into the CPU package. It’s not like there’s ever a situation where you don’t need RAM.
All your RAM needs to be the same speed unless you want to open up a rabbit hole. All attempts at that thus far have kinda flopped. You can make very good use of such systems, but I’ve only seen it succeed with software specifically tailored for that use case (say databases or simulations).
The way I see it, RAM in the future will be on package and non-expandable. CXL might get some traction, but naah.
Couldn’t you just treat the socketed ram like another layer of memory effectively meaning that L1-3 are on the CPU “L4” would be soldered RAM and then L5 would be extra socketed RAM? Alternatively couldn’t you just treat it like really fast swap?
Wrote a longer reply to someone else, but briefly, yes, you are correct. Kinda.
Caches won’t help with bandwidth-bound compute (read: ”AI”) it the streamed dataset is significantly larger than the cache. A cache will only speed up repeated access to a limited set of data.
Using it as cache would reduce total capacity as cache implies coherence, and treating it as ordinary swap would mean copying to main memory before you access it which is silly when you can access it directly. That is you’d want to write a couple of lines of kernel code to use it effectively but it’s nowhere close to rocket science. Nowhere near as complicated as making proper use of NUMA architectures.
The cache hierarchy has flopped? People aren’t using swap?
NUMA also hasn’t flopped, it’s just that most systems aren’t multi socket, or clusters. Different memory speeds connected to the same CPU is not ideal and you don’t build a system like that but among upgraded systems that’s not rare at all and software-wise worst thing that’ll happen is you get the lower memory speed. Which you’d get anyway if you only had socketed RAM.
Yeah, the cache hierarchy is behaving kinda wonky lately. Many AI workloads (and that’s what’s driving development lately) are constrained by bandwidth, and cache will only help you with a part of that. Cache will help with repeated access, not as much with streaming access to datasets much larger than the cache (i.e. many current AI models).
Intel already tried selling CPUs with both on-package HBM and slotted DDR-RAM. No one wanted it, as the performance gains of the expensive HBM evaporated completely as soon as you touched memory out-of-package. (Assuming workloads bound by memory bandwidth, which currently dominate the compute market)
To get good performance out of that, you may need to explicitly code the memory transfers to enable prefetch (preferably asynchronous) from the slower memory into the faster, á la classic GPU programming. YMMW.
In systems where memory speed are mismatched, the system runs at the slowest module’s speed. So literally making the soldered, faster memory slower. Why even have soldered memory at that point?
I’d assume the soldered memory to have a dedicated memory controller. There’s also no hard requirement that a single controller can’t drive different channels at different speeds. The only hard requirement is that one channel needs to run at one speed.
…and the whole thing becomes completely irrelevant when we’re talking about PCIe expansion cards the memory controller doesn’t care.
I definitely wouldn’t mind soldered RAM if there’s still an expansion socket. Solder in at least a reasonable minimum (16G?) and not the cheap stuff but memory that can actually use the signal integrity advantage, I may want more RAM but it’s fine if it’s a bit slower. You can leave out the DIMM slot but then have at least one PCIe x16 expansion slot. A free one, one in addition to the GPU slot. PCIe latency isn’t stellar but on the upside, expansion boards would come with their own memory controllers, and push come to shove you can configure the faster RAM as cache / the expansion RAM as swap.
Heck, throw the memory into the CPU package. It’s not like there’s ever a situation where you don’t need RAM.
All your RAM needs to be the same speed unless you want to open up a rabbit hole. All attempts at that thus far have kinda flopped. You can make very good use of such systems, but I’ve only seen it succeed with software specifically tailored for that use case (say databases or simulations).
The way I see it, RAM in the future will be on package and non-expandable. CXL might get some traction, but naah.
Couldn’t you just treat the socketed ram like another layer of memory effectively meaning that L1-3 are on the CPU “L4” would be soldered RAM and then L5 would be extra socketed RAM? Alternatively couldn’t you just treat it like really fast swap?
Could it work?
Yes, but it would require:
Right now, the easiest solution for fast, high-bandwidth RAM is just to solder all of it.
Wrote a longer reply to someone else, but briefly, yes, you are correct. Kinda.
Caches won’t help with bandwidth-bound compute (read: ”AI”) it the streamed dataset is significantly larger than the cache. A cache will only speed up repeated access to a limited set of data.
Using it as cache would reduce total capacity as cache implies coherence, and treating it as ordinary swap would mean copying to main memory before you access it which is silly when you can access it directly. That is you’d want to write a couple of lines of kernel code to use it effectively but it’s nowhere close to rocket science. Nowhere near as complicated as making proper use of NUMA architectures.
The cache hierarchy has flopped? People aren’t using swap?
NUMA also hasn’t flopped, it’s just that most systems aren’t multi socket, or clusters. Different memory speeds connected to the same CPU is not ideal and you don’t build a system like that but among upgraded systems that’s not rare at all and software-wise worst thing that’ll happen is you get the lower memory speed. Which you’d get anyway if you only had socketed RAM.
Yeah, the cache hierarchy is behaving kinda wonky lately. Many AI workloads (and that’s what’s driving development lately) are constrained by bandwidth, and cache will only help you with a part of that. Cache will help with repeated access, not as much with streaming access to datasets much larger than the cache (i.e. many current AI models).
Intel already tried selling CPUs with both on-package HBM and slotted DDR-RAM. No one wanted it, as the performance gains of the expensive HBM evaporated completely as soon as you touched memory out-of-package. (Assuming workloads bound by memory bandwidth, which currently dominate the compute market)
To get good performance out of that, you may need to explicitly code the memory transfers to enable prefetch (preferably asynchronous) from the slower memory into the faster, á la classic GPU programming. YMMW.
In systems where memory speed are mismatched, the system runs at the slowest module’s speed. So literally making the soldered, faster memory slower. Why even have soldered memory at that point?
I’d assume the soldered memory to have a dedicated memory controller. There’s also no hard requirement that a single controller can’t drive different channels at different speeds. The only hard requirement is that one channel needs to run at one speed.
…and the whole thing becomes completely irrelevant when we’re talking about PCIe expansion cards the memory controller doesn’t care.