News CAMM2 benchmarks show similar performance to normal DIMMs — Kingston DDR5-7200 C38 CAMM2 with CKD compared to DDR5 DIMM

[Admin]

MSI benchmarked Kingston's DDR5-7200 C38 CAMM with CKD on the brand's Z790 Project Zero Plus motherboard to evaluate its performance against standard DDR5 modules.

CAMM2 benchmarks show similar performance to normal DIMMs — Kingston DDR5-7200 C38 CAMM2 with CKD compared to DDR5 DIMM : Read more

 

[usertests]

MSI Highlights Benefits of CAMM2 DDR5 Memory For Desktop PCs: Shorter Traces Lead To Higher Speeds & Lower Latencies, Cooler & Innovative Designs

CAMM2 memory modules will soon be seeing their debut on Desktop PCs and MSI has highlighted some of the major benefits for the platform.

wccftech.com

I'm looking at these slides, and apparently (LP?)CAMM uses more motherboard space than soldered LPDDR5, but saves space compared to SO-DIMM. Is it going to take off on Mini-ITX boards or am I missing something?

We have seen a 128 GB LPCAMM, which means two of them gets you to 256 GB. So the maximum capacity should be better than any quad-DIMM (consumer UDIMM) situation.

 

[BillyBuerger]

Was there a question about performance of CAMM2? I always saw it as primarily a form factor change. It's still the same memory standards so why wouldn't it perform the same. Other then possible trace lengths which I don't see as being that much different from standard DIMMs.

 

[thestryker]

I'm not opposed to seeing this actually out in the real world, but I'm just not sold on desktop yet. If they could show some tangible benefits and industry buy-in I'd be more interested.

We have seen a 128 GB LPCAMM, which means two of them gets you to 256 GB. So the maximum capacity should be better than any quad-DIMM (consumer UDIMM) situation.

So far none of the real world implementations of CAMM2 have been the dual module variety (LP or regular). I'm not sure how much we will see it in actual products, but if we do I'd expect it to be primarily laptops. The dual module setup stacks the two modules which would be rather limiting on desktop and I'm not totally sure how the cooling would work in server.

I'm looking at these slides, and apparently (LP?)CAMM uses more motherboard space than soldered LPDDR5, but saves space compared to SO-DIMM. Is it going to take off on Mini-ITX boards or am I missing something?

The primary advantage for it is definitely as a SODIMM replacement. I don't believe there have been any SODIMM modules released over 5600, but these should scale to whatever DDR5 does. They also bring removable LPDDR to the table which could be another differentiator in the laptop space.

In theory CAMM2 could allow for higher speeds/lower latencies than standard UDIMMs. I'm not sure when that would manifest in the real world given how good a lot of the XMP/EXPO kits have already gotten. There is a possibility it could inherently make driving speeds/latency easier on the memory controller as well (similar to how the Apex is better than every other board for memory speeds currently).

There's also the limited capacity which may turn some people off (128GB DDR5 CAMM2 is huge so don't expect to see those any time soon) and as discussed above dual module implementation seems unlikely.

 

[mras2]

MSI Highlights Benefits of CAMM2 DDR5 Memory For Desktop PCs: Shorter Traces Lead To Higher Speeds & Lower Latencies, Cooler & Innovative Designs

CAMM2 memory modules will soon be seeing their debut on Desktop PCs and MSI has highlighted some of the major benefits for the platform.

wccftech.com

I'm looking at these slides, and apparently (LP?)CAMM uses more motherboard space than soldered LPDDR5, but saves space compared to SO-DIMM. Is it going to take off on Mini-ITX boards or am I missing something?

We have seen a 128 GB LPCAMM, which means two of them gets you to 256 GB. So the maximum capacity should be better than any quad-DIMM (consumer UDIMM) situation.

LP- is Low Power, its for laptops, so.. yeah...
There has yet to be announced anything above 1 slot CAMM2 modules on motherboards. So if you get a CAMM2 motherboard, result will be it will just have 1 of these slots - Atleast for now.
Not ideal for desktop, at all.

 

[bit_user]

The two main negatives I see against CAMM2 are:

1. Larger footprint than perpendicularly-inserted DIMMs
2. Harder to cool, since some DRAM chips can go on the backside.

The main benefit is that it will probably allow for tighter timings than DIMMs, particularly against a two-DIMMs-per-channel (2DPC) setup.

Other than that, there shouldn't be any other major differences. I'm actually surprised they even found performance discrepancies as large as they did.

CAMM2 is available in DDR5 and LPDD5X configurations, with the latter competing against some of the fastest XMP memory kits on the market regarding bandwidth.

@TechyInAZ , LPDDR5(X) has much higher latency than regular DDR5. This shouldn't be glossed over. The other tradeoff it makes for its impressive bandwidth & power savings is that it normally must be soldered down. LPCAMM2 is a big breakthrough for enabling LPDDR5(X) to be upgradable.

Here's a good reference point on LPDDR5X latency. Through some clever tuning, ChipsAndCheese managed to get DRAM latency of Meteor Lake down below 150 ns, but regular DDR5 latency benchmarks can easily see figures under 90 ns.

Update on Meteor Lake DRAM Latency Measurements

Chip makers are always looking to save power, especially in mobile devices. Like a CPU core, Meteor Lake’s memory controller can enter lower power states with reduced clocks and voltages. I&#…

chipsandcheese.com

 

[usertests]

LP- is Low Power, its for laptops, so.. yeah...
There has yet to be announced anything above 1 slot CAMM2 modules on motherboards. So if you get a CAMM2 motherboard, result will be it will just have 1 of these slots - Atleast for now.
Not ideal for desktop, at all.

SO-DIMM was for laptops, and that ended up on Mini-ITX and mini PCs.

All of the details related to CAMM are up in the air as far as I'm concerned. It's unfinished.

 

[bit_user]

We have seen a 128 GB LPCAMM, which means two of them gets you to 256 GB.

First of all, 128 GB LPCAMM2 modules are only possible using 32 Gb chips, according to JEDEC, which I think are still quite new.

As for double-stacking, I've seen references to this, but it's not mentioned in this JEDEC whitepaper from May 2024. That makes me wonder if it's an official thing, or still at the level of an idea or proposal.

https://www.jedec.org/sites/default/files/Tom_Schnell_FINAL_ 2024-05-03.pdf
​

So the maximum capacity should be better than any quad-DIMM (consumer UDIMM) situation.

Huh? Are you assuming quad-LPCAMM will be possible? Otherwise, I don't see how it passes quad-DIMM. Quad CAMM definitely won't be a thing - at least, not on any CPU with just a 128-bit memory interface.

All of the details related to CAMM are up in the air as far as I'm concerned. It's unfinished.

WTF? CAMM2 is now an official JEDEC standard. See above whitepaper!

BTW, pay special attention to the form factor variants and capacity limits!

 

[bit_user]

There's also the limited capacity which may turn some people off (128GB DDR5 CAMM2 is huge so don't expect to see those any time soon) and as discussed above dual module implementation seems unlikely.

In the above whitepaper, they describe an AXXX form factor that's the same 40 mm in length as CXXX, but supports up to 128 GB (using dual-die packaged 32 Gb DRAM). I assume that will show up as dual rank?

Also, I find it interesting how DXXX appears to have half the contact points as BXXX. This would imply that if you're using the D variant, you're basically committing to using two of them (or else you'll get just half the bandwidth).

 

[thestryker]

Also, I find it interesting how DXXX appears to have half the contact points as BXXX. This would imply that if you're using the D variant, you're basically committing to using two of them (or else you'll get just half the bandwidth).

Since it also specifically mentions stacking in the capacity area I think that's the only logical conclusion. My assumption is that this is the design for servers since capacity vs area is the biggest problem for CAMM2 there.

In the above whitepaper, they describe an AXXX form factor that's the same 40 mm in length as CXXX, but supports up to 128 GB (using dual-die packaged 32 Gb DRAM). I assume that will show up as dual rank?

Is this packaging even a thing? Samsung doesn't sell any package over 32Gb right now for DDR5 and that's the highest capacity of the trio.

First of all, 128 GB LPCAMM2 modules are only possible using 32 Gb chips, according to JEDEC, which I think are still quite new.

I think that the printing in the presentation is wrong here since LPDDR5/X currently cap out at 128Gb packages* and LPCAMM2 modules all show 4 packages which would be 64GB. So I'm pretty sure it should say either 128Gb/192Gb/256Gb or 16GB/24GB/32GB. Those capacities would make sense for regular DDR5 though.

*Samsung has 1 part in both that go up to 144Gb || SK Hynix has 144Gb in LPDDR5 and put out a press release regarding 192Gb LPDDR5X but it's not in any public catalog.

As for double-stacking, I've seen references to this, but it's not mentioned in this JEDEC whitepaper from May 2024. That makes me wonder if it's an official thing, or still at the level of an idea or proposal.

I don't think it's a thing for LPDDR since the bus width per package tends to be minimum 32-bit for any capacity that would be relevant and CAMM2 stacking is based on splitting the memory bus in half.

 

[dehjomz]

MSI benchmarked Kingston's DDR5-7200 C38 CAMM with CKD on the brand's Z790 Project Zero Plus motherboard to evaluate its performance against standard DDR5 modules.

CAMM2 benchmarks show similar performance to normal DIMMs — Kingston DDR5-7200 C38 CAMM2 with CKD compared to DDR5 DIMM : Read more

I wonder if CAMM2 on Desktop might enable lower latency SKUs? CL28 or lower ?

 

[bit_user]

Is this packaging even a thing? Samsung doesn't sell any package over 32Gb right now for DDR5 and that's the highest capacity of the trio.

I don't know if you can buy it today, but die-stacked DDR4 was a thing and apparently JEDEC specified it for DDR5.

I think that the printing in the presentation is wrong here since LPDDR5/X currently cap out at 128Gb packages* and LPCAMM2 modules all show 4 packages which would be 64GB.

Again, the discrepancy could be between what exists today vs. what JEDEC specs exist for. To achieve it with 32 Gb dies, you'd need 8-high stacks. According to this whitepaper, Samsung is planning to go up to 8-high stacks of 32 Gb, which would provide the necessary capacity to enable 128 GB LPCAMM2 modules.

"As demand for high-performance, high-density mobile memory solutions in smaller package sizes continues to grow, the company plans to develop 6-layer 24GB and 8-layer 32GB modules into the thinnest LPDDR DRAM packages for future devices."

Source: https://news.samsung.com/global/sam...innest-lpddr5x-dram-packages-for-on-device-ai
​

 

[bit_user]

I wonder if CAMM2 on Desktop might enable lower latency SKUs? CL28 or lower ?

Yeah, by improving electrical characteristics, it might be possible.

Keep in mind that CL is always relative to the frequency, so it doesn't make sense to talk about it in the abstract. CL28 at DDR5-6400 is equivalent to CL21 at DDR5-4800.

 

[DavidC1]

LPCAMM2 is a big breakthrough for enabling LPDDR5(X) to be upgradable.

It won't fully replace soldered LPDDR setups. Physical distance determines signal integrity and BGA soldered is always better. CAMM allows better than before, but most portable and most power saving implementations will be soldered.

It's just normal tech advancement. Fills a gap that was getting bigger all the time.

 

[Amdlova]

After this will keep my ddr4 system a little longer.
Will update the asrock 670 pg to a asus h770 plus D4.
May try fry the IMC from the 14600T

 

[NinoPino]

Yeah, by improving electrical characteristics, it might be possible.

Keep in mind that CL is always relative to the frequency, so it doesn't make sense to talk about it in the abstract. CL28 at DDR5-4800 is equivalent to CL21 at DDR5-6400.

May be you meant CL28 @ 6400 equivalent to CL21 @ 4800 ?

 

[bit_user]

May be you meant CL28 @ 6400 equivalent to CL21 @ 4800 ?

Thank you. Fixed.
: )

 

[Rob1C]

Thanks for the article.

Now I know that I'm more excited about MR DIMMs and even MCR DIMMs, over CAMM2.

CAMM2 is fine for flat devices, like laptops and cellphones, but for desktop/server it just takes up more motherboard space; and eventually will be just ~50% faster than what it replaces, while MR DIMMs will be multiple times faster and with much greater capacity.

 

[bit_user]

Now I know that I'm more excited about MR DIMMs and even MCR DIMMs, over CAMM2.

Those aren't likely to be supported on desktop CPUs, I'm pretty sure. IIRC, they start with registered memory as a baseline requirement, which desktop CPUs don't support.

 

[thestryker]

Now I know that I'm more excited about MR DIMMs and even MCR DIMMs, over CAMM2.

Even if these were going to be available for client devices the benefits aren't guaranteed. The main benefits they bring are based on being able to eliminate 2DPC at the server level and increased bandwidth with minimal latency penalty. The key here is though that this is compared to JEDEC specifications. As an example my W680 board has ECC DDR5 5600, but the latency is CL46 and on a regular desktop client you'd be running at least 6000 at CL30.

 

[Camo5]

Please at least set the CPU core multipliers to be equal when you run your tests...
The DIMM computer was running at a 55X multiplier while the CAMM system was only at 51X, so of course the DIMM computer is going to come out on top...Not to mention the huge disparity with the North Bridge Clock