How would HBM modules even work? Right now HMB is on its own special fabric that is tied directly to the CPU.

I don't see how the signal integrity would ever be good enough if HBM was stuck on replaceable modules.

It would be extremely expensive and very unreliable as a replaceable type module.

The only feasible option really would be to replace the CPU or GPU module that had HBM built-in. That would be extremely expensive any time you wanted to upgrade capacity.

Small, directly adjacent to cpu, vertical 'thru package' pin header.

They already work as a stacked bga package.

I'm suggesting that the interposer just needs a compact LIF socket, instead of a BGA. Then the HBM packages can be replaced.

wierd_w wrote on 2026-07-27, 17:15:

Small, directly adjacent to cpu, vertical 'thru package' pin header. […]
Show full quote

Small, directly adjacent to cpu, vertical 'thru package' pin header.

They already work as a stacked bga package.

I'm suggesting that the interposer just needs a compact LIF socket, instead of a BGA. Then the HBM packages can be replaced.

Or they use the CAMM module socket and place it adjacent to the CPU so it has the shortest travel path, this would mean a new motherboard format to accommodate the CAMM socket. The beauty of CAMM is that it can be used for any type of memory and would allow water block cooling too. Having HBM as system memory would also allow IGPUs and on board GPUs to use it as high speed/bandwidth VRAM which could allow even better IGPUs than what we currently have. I feel we are just now reaching a point where IGPUs have become more than sufficient for the majority of PC users that it wouldn't take very much extra performance to make them worth having over a discrete GPU and VRAM is a big stumbling block as DDR5 is still too slow for graphics hardware.

You would lose some speed and bandwidth not having the HBM on the same interposer as the CPU but I doubt itll be significant enough to affect general computing.

Not sure how others here feel about CAMM but I like it, it offers a good way forward and plenty of flexibility.

Trashbytes wrote on 2026-07-27, 21:49:

Or they use the CAMM module socket and place it adjacent to the CPU so it has the shortest travel path, this would mean a new mo […]
Show full quote

wierd_w wrote on 2026-07-27, 17:15:

Small, directly adjacent to cpu, vertical 'thru package' pin header. […]
Show full quote

Small, directly adjacent to cpu, vertical 'thru package' pin header.

They already work as a stacked bga package.

I'm suggesting that the interposer just needs a compact LIF socket, instead of a BGA. Then the HBM packages can be replaced.

Or they use the CAMM module socket and place it adjacent to the CPU so it has the shortest travel path, this would mean a new motherboard format to accommodate the CAMM socket. The beauty of CAMM is that it can be used for any type of memory and would allow water block cooling too. Having HBM as system memory would also allow IGPUs and on board GPUs to use it as high speed/bandwidth VRAM which could allow even better IGPUs than what we currently have. I feel we are just now reaching a point where IGPUs have become more than sufficient for the majority of PC users that it wouldn't take very much extra performance to make them worth having over a discrete GPU and VRAM is a big stumbling block as DDR5 is still too slow for graphics hardware.

You would lose some speed and bandwidth not having the HBM on the same interposer as the CPU but I doubt itll be significant enough to affect general computing.

Not sure how others here feel about CAMM but I like it, it offers a good way forward and plenty of flexibility.

Again, I dont care about the shape. To me, CAMM is just a novel compact LIF socket, with a circuit board hanging off of it.

the main architectural advance with HBM is that total travel time of circuit traces are minimized, by having the shortest possible paths. that means having the bus go UPWARD, with stacked chips.

the stacks themselves just need to be what's modular. Something like CAMM's socket would be fine, IMO. It just needs to be a socket, and not "LULZ, We soldered it directly on! Tee Hee!"

As long as it is some kind of socket, and it is standardized, I will be happy.

wierd_w wrote on 2026-07-27, 22:14:

Again, I dont care about the shape. To me, CAMM is just a novel compact LIF socket, with a circuit board hanging off of it. […]
Show full quote

Trashbytes wrote on 2026-07-27, 21:49:

Or they use the CAMM module socket and place it adjacent to the CPU so it has the shortest travel path, this would mean a new mo […]
Show full quote

wierd_w wrote on 2026-07-27, 17:15:

Small, directly adjacent to cpu, vertical 'thru package' pin header. […]
Show full quote

Small, directly adjacent to cpu, vertical 'thru package' pin header.

They already work as a stacked bga package.

I'm suggesting that the interposer just needs a compact LIF socket, instead of a BGA. Then the HBM packages can be replaced.

Or they use the CAMM module socket and place it adjacent to the CPU so it has the shortest travel path, this would mean a new motherboard format to accommodate the CAMM socket. The beauty of CAMM is that it can be used for any type of memory and would allow water block cooling too. Having HBM as system memory would also allow IGPUs and on board GPUs to use it as high speed/bandwidth VRAM which could allow even better IGPUs than what we currently have. I feel we are just now reaching a point where IGPUs have become more than sufficient for the majority of PC users that it wouldn't take very much extra performance to make them worth having over a discrete GPU and VRAM is a big stumbling block as DDR5 is still too slow for graphics hardware.

You would lose some speed and bandwidth not having the HBM on the same interposer as the CPU but I doubt itll be significant enough to affect general computing.

Not sure how others here feel about CAMM but I like it, it offers a good way forward and plenty of flexibility.

Again, I dont care about the shape. To me, CAMM is just a novel compact LIF socket, with a circuit board hanging off of it.

the main architectural advance with HBM is that total travel time of circuit traces are minimized, by having the shortest possible paths. that means having the bus go UPWARD, with stacked chips.

the stacks themselves just need to be what's modular. Something like CAMM's socket would be fine, IMO. It just needs to be a socket, and not "LULZ, We soldered it directly on! Tee Hee!"

As long as it is some kind of socket, and it is standardized, I will be happy.

Playing Devils Advocate here. I could see that if they standardised HBM 3 as normal system memory and it came as 64 Gig units ...I don't actually have an issue with it being soldered directly to the motherboard next to the CPU. Naturally I would still want unsoldered boards to be available for power users and DIY system builders, but for many people 64 Gig of HBM 3 on the motherboard would be more than sufficient and it would lower production costs.

Ok advocate locked away

Like you I too want to see a standardised memory socket that isn't DIMM socket based, DIMM needs to die as a socket standard already.

An alternative to a flat interposer could be a 'reimagining' of Slot1 style CPU.

CPU and cooler goes on one side, HBM stack goes on the other side, 'functionally' right on the back side of the CPU's socket thru-holes. This retains 'minimum travel distance', keeps both sides reachable, and lets your motherboard still have a user upgradable formfactor.

It would just have all the same warts slot1 had.

wierd_w wrote on 2026-07-29, 02:21:

An alternative to a flat interposer could be a 'reimagining' of Slot1 style CPU.

CPU and cooler goes on one side, HBM stack goes on the other side, 'functionally' right on the back side of the CPU's socket thru-holes. This retains 'minimum travel distance', keeps both sides reachable, and lets your motherboard still have a user upgradable formfactor.

It would just have all the same warts slot1 had.

Long ago I saw machines that had the cpu on top and on the bottom of the board opposite there was a chip soldered underneath, considered the otherwise low circuit density it was probably there to reduce travel distance and noise.

Also encountered a machine with a through hole in the motherboard under the CPU that had a heatsync contacting the bottom of the chip with the fan on the bottom of the case.

If we wanted to get creative a mechanism to install ram as close as possible wouldn’t be hard, with the new Taj Mahal case designs access to both sides of a vertical board could be made less painful.

In other news:
https://www.google.com/amp/s/www.techpo ... n-q3%3famp

Dimos wrote on 2026-08-06, 00:07:

In other news:
https://www.google.com/amp/s/www.techpo ... n-q3%3famp

Wonderful.

My X570 Aorus Elite is still going strong after 7 years.

I think I will bump up the speed of my intake fans a bit to keep things juuuuust a little bit cooler and more comfy. We're in this one for the long haul.

Ozzuneoj wrote on 2026-08-06, 15:35:

Wonderful.

My X570 Aorus Elite is still going strong after 7 years.

I think I will bump up the speed of my intake fans a bit to keep things juuuuust a little bit cooler and more comfy. We're in this one for the long haul.

That's how I'm managing my P35, X38, X58, Z77, X79, and Z97 boards. I'm stock full of DDR2 and DDR3 RAM and no overclocking!

I'm driving an AM4 board, that I packed with ram before the madness.

Looks like i'll be driving it for a long time.

Ozzuneoj wrote on 2026-08-06, 15:35:

Wonderful. […]
Show full quote

Dimos wrote on 2026-08-06, 00:07:

In other news:
https://www.google.com/amp/s/www.techpo ... n-q3%3famp

Wonderful.

My X570 Aorus Elite is still going strong after 7 years.

I think I will bump up the speed of my intake fans a bit to keep things juuuuust a little bit cooler and more comfy. We're in this one for the long haul.

Heh, I didn't realize x570 was that old. I didn't upgrade till a decent while later when I got a pretty good deal on a 5950X when it was on sale on Amazon.

Might be time for me to grab a newer used board, CPU and RAM.

Edit... And parts ordered. Hopefully I get another 5-7 years before feeling any need to upgrade my main PC.

Current parts will go to one of my kids.

It keeps getting better:
https://www.eteknix.com/memory-production-for … -for-consumers/

Yeah but I see that as a good thing.

I mean yeah its bad if youre in the market for a new PC or new PC parts but lets be honest most of us on here are more interested in older PC parts and they are always subject to weird pricing by their nature (being old).

On the other side this bubble we know is there and growing is going to burst big when it does burst.
And we know its going to burst because the sites that offer up these pointless AI tools, of which there are now millions cant find a way to make money from them.
Theyre all free! Or prices to use them are unsustainable, the fact google had to change its service relative to its pricing shows that.
There are other sites that are getting ready to restructure their pricing/services, so we are only at the stage where these sites are starting to see the pinch, but not feeling it yet because they can adjust the prices to fix the problems.

I shell out close to £50 a month for AI services but if it goes over that I wont bother because its just not worth it to me. And Im a fully sold up member of the AI club! And I WOULDNT pay!

So when this bubble bursts its going to dump so much hardware into the market it will be painful to watch as these guys start to wonder what the bottom of the Serpentine looks like in real terms!

This bubble cant burst until the supply catches up with demand.
Because once they get all the hardware they think they need they will start paying for it, once they do they will need to rinse money out of subscribers, but the services are not worth that much so the companies wont be able to meet their debts... Because its all borrowed money.
Then and only then can the bubble burst.

Thats exactly what happed in the 90s with all the dot coms.
Dead dot com hardware was hitting the market for years after that, and it was great buying up server hardware for buttons.
And then the hardware manufacturers will pivot to consumers to fill the gap their freed up capacity gives them. Because they will want to make money and consumers will provide the cash they need.

Its going to be crap for a few years...
2028 at least. And prices will only get higher as people get desperate, but most of us soldier on with old hardware anyway.

With any luck some of these stupid hardware assemblers that make a living on youtube die off too.

wierd_w wrote on 2026-07-29, 02:21:

An alternative to a flat interposer could be a 'reimagining' of Slot1 style CPU.

CPU and cooler goes on one side, HBM stack goes on the other side, 'functionally' right on the back side of the CPU's socket thru-holes. This retains 'minimum travel distance', keeps both sides reachable, and lets your motherboard still have a user upgradable formfactor.

It would just have all the same warts slot1 had.

I don't really see how that relates to slot 1.
It seems pretty straightforward to me. Modern LGA sockets are surface mount.
Meaning the only 'pins' visible on the back side of the board are the necessary/chosen Vias.

You designate a signal pattern footprint, and a shared/separate power/earth footprint outside of the main socket area. You build a "CAMM3" around that pattern, and the next board revision/sockets have 2-4 rear accessible footprints between the screw holes.
HBM would be easiest in theory, but IIRC it's got a very wide bus.
DDR is only 64bit/stick, 128 for dual.
Cooling will be a definite factor with the traces transmitting so much heat from the cpu. It seems unlikely "naked" ram chips could function at even JDEC speeds.

Again, I can't really see the similarity to slot 1 unless you're referring to the cache?

ElectroSoldier wrote on Today, 01:52:

Yeah but I see that as a good thing. […]
Show full quote

Yeah but I see that as a good thing.

I mean yeah its bad if youre in the market for a new PC or new PC parts but lets be honest most of us on here are more interested in older PC parts and they are always subject to weird pricing by their nature (being old).

On the other side this bubble we know is there and growing is going to burst big when it does burst.
And we know its going to burst because the sites that offer up these pointless AI tools, of which there are now millions cant find a way to make money from them.
Theyre all free! Or prices to use them are unsustainable, the fact google had to change its service relative to its pricing shows that.
There are other sites that are getting ready to restructure their pricing/services, so we are only at the stage where these sites are starting to see the pinch, but not feeling it yet because they can adjust the prices to fix the problems.

I shell out close to £50 a month for AI services but if it goes over that I wont bother because its just not worth it to me. And Im a fully sold up member of the AI club! And I WOULDNT pay!

So when this bubble bursts its going to dump so much hardware into the market it will be painful to watch as these guys start to wonder what the bottom of the Serpentine looks like in real terms!

This bubble cant burst until the supply catches up with demand.
Because once they get all the hardware they think they need they will start paying for it, once they do they will need to rinse money out of subscribers, but the services are not worth that much so the companies wont be able to meet their debts... Because its all borrowed money.
Then and only then can the bubble burst.

Thats exactly what happed in the 90s with all the dot coms.
Dead dot com hardware was hitting the market for years after that, and it was great buying up server hardware for buttons.
And then the hardware manufacturers will pivot to consumers to fill the gap their freed up capacity gives them. Because they will want to make money and consumers will provide the cash they need.

Its going to be crap for a few years...
2028 at least. And prices will only get higher as people get desperate, but most of us soldier on with old hardware anyway.

With any luck some of these stupid hardware assemblers that make a living on youtube die off too.

The market will self correct but it wont be in the direction we want, itll be a correction where they will simply sell the spare compute rather than sell the hardware, the Data Centers are not going anywhere even if AI dies and they will need to find a way to sell that compute. So the profits from selling hardware to the DCs will still be there, it wont be as lucrative as it is now but it will still be far more substantial than selling it to the consumer plebs. nVidia will simply pivot fully into selling compute via subscription and streaming services, Amazon and Microsoft will follow suit and the DIY space will ...well it simply wont exist as it does now and it will become a high roller luxury space where only the people with large disposable incomes will be able to afford the hardware.

People will try their best to sell me on .. "that's bullshit itll never happen" ...but we have been watching it happen the last few years like a frog in a pot of boiling water as more and more services and features have been moved to the compute cloud and chained to a monthly/annual subscription fee. So even if you own the hardware you will still be forced into their online compute system for a monthly fee of $19.99 for 100 hours of compute.

So the bubble bursting wont be the "FIX" you think it is, the fabs making the core hardware don't want lower pricing and nothing will force them to return to pre AI bubble pricing either.

The totality of it all is, THEY DONT WANT US OWNING COMPUTE, or a level of compute that can run local AI that can compete with theirs and they will 100% make sure that even if it all crashes and burns DIY PCs are going with them.

myne wrote on Today, 03:07:

I don't really see how that relates to slot 1. It seems pretty straightforward to me. Modern LGA sockets are surface mount. Mea […]
Show full quote

wierd_w wrote on 2026-07-29, 02:21:

An alternative to a flat interposer could be a 'reimagining' of Slot1 style CPU.

CPU and cooler goes on one side, HBM stack goes on the other side, 'functionally' right on the back side of the CPU's socket thru-holes. This retains 'minimum travel distance', keeps both sides reachable, and lets your motherboard still have a user upgradable formfactor.

It would just have all the same warts slot1 had.

I don't really see how that relates to slot 1.
It seems pretty straightforward to me. Modern LGA sockets are surface mount.
Meaning the only 'pins' visible on the back side of the board are the necessary/chosen Vias.

You designate a signal pattern footprint, and a shared/separate power/earth footprint outside of the main socket area. You build a "CAMM3" around that pattern, and the next board revision/sockets have 2-4 rear accessible footprints between the screw holes.
HBM would be easiest in theory, but IIRC it's got a very wide bus.
DDR is only 64bit/stick, 128 for dual.
Cooling will be a definite factor with the traces transmitting so much heat from the cpu. It seems unlikely "naked" ram chips could function at even JDEC speeds.

Again, I can't really see the similarity to slot 1 unless you're referring to the cache?

DDR5 is only that, DDR6 is 4x24bit sub channel architecture widening the entire bus from 64bit to 96bit, DDR6 also fully adopts the CAMM2 platform so it makes sense that DIMM slots will soon be a thing of the past. DDR5 AFAIK doesn't support CAMM2 without extra control circuits on the CAMM2 board which does complicate things but I cant see them bothering on releasing CAMM2 DDR5 to consumers outside of laptops and specific OEMs.

HBM2e/HBM3 is 1024bits wide but CAMM2 can handle that with little issue .. its the CPUs that would need to change to use it.

HBM4/4e is also coming with a 2048bit bus and 4e doubles the bandwidth to 4096GB/s so ..I really want to see them just drop DDR in favour of HBM .. but itll never happen. (HBM4e is an utter monster and I want all GPUs to use it over the GDDR shit)

I think it's moderately likely 4/8/16gb of HBM will be integrated into the CPU package in the future.
Not as a cached copy of main ram. As RAM.
External sticks will be a kind of "level2" ram for bigger, background or less "important" processes.

Trash tablets and laptops will just use the big HBM SOC and maybe one slot.
"proper desktops" will have external slots as well as whatever's on the SOC.

The trend is always toward unifying laptop/desktop/server architectures. They branch for a while, then one is killed once the compromises are adequately addressed.

So a huge last level cache (effectively) and then a single addon stick.

Hahahahahahahaha. No thanks.

Yes, I would like to buy a 2TB stick of RAM.

The minimum I would accept as HBM would be 32GB (right now). And then additional RAM on slots would have to be at least dual channel, more likely quad channel.

I've had 64GB in my main desktop for at least 5 years. One of my laptops has 128GB and another has 96GB.

My wife's desktop has 96GB.

I am interested in old hardware very much, but also in contemporary hardware.

Been a nerd since I was around 12. My family got our first computer when I was 14 but I had been messing with computers at school and friends houses for a couple years before that.

It was only a few months before I started installing upgrades in our computer.

First computer I overclocked was a friend's 486 25. Saw a jumper to switch between 25 and 33Mhz on the motherboard and was hooked on overclocking instantly.

cyclone3d wrote on Today, 04:33:

So a huge last level cache (effectively) and then a single addon stick. […]
Show full quote

So a huge last level cache (effectively) and then a single addon stick.

Hahahahahahahaha. No thanks.

Yes, I would like to buy a 2TB stick of RAM.

The minimum I would accept as HBM would be 32GB (right now). And then additional RAM on slots would have to be at least dual channel, more likely quad channel.

I've had 64GB in my main desktop for at least 5 years. One of my laptops has 128GB and another has 96GB.

My wife's desktop has 96GB.

I am interested in old hardware very much, but also in contemporary hardware.

Been a nerd since I was around 12. My family got our first computer when I was 14 but I had been messing with computers at school and friends houses for a couple years before that.

It was only a few months before I started installing upgrades in our computer.

First computer I overclocked was a friend's 486 25. Saw a jumper to switch between 25 and 33Mhz on the motherboard and was hooked on overclocking instantly.

No the on SOC HBM would be essentially super fast wide bus high bandwidth memory that the CPU can use for anything that needs absolute speed (Like the iGPU, or providing a massive on die cache that the CPU and GPU can use that bypasses system memory, think HUGE X3D cache), the motherboard memory behaves as it has always done as your main system memory that everything including the dog can use. DDR6 is essentially quad channel memory so I see little benefit over using more than one slab of it if it comes in 32/64 or 24/48 gig modules. (Likely they will drop the 32/64 ones as DDR6 is 4x24bit channels, so 24/48/96/192 sizing makes more sense here)

X3D wont stay as it is for much longer and whatever Intel is cooking wont either and HBM would unify that with an off the shelf solution. HBM4e is a monster memory and would run circles around DDR6 even if it was just 32Gb on SOC.

https://www.rambus.com/blogs/hbm3-everything- … now/#generation Take a gander at the table listing bandwidth and speeds between the versions ...4e is the guy your wife tells you not to worry about. Youll notice that HBM has a device limit so needs to be used sparingly, 32 on SOC and DDR6 making up the system memory would be a good solution.