DDR5 CUDIMM Explained & Benched

[biffzinker]

no need. Rocking solid DDR 4 and 3200


View attachment 392641

DDR4-3600 is a nice bump from 3200 especially for a Ryzen chip.

Thanks for the article about the clock driver chip @ir_cow

 

[Makaveli]

DDR4-3600 is a nice bump from 3200 especially for a Ryzen chip.

If that is 3600 at CL14 I agree

3600 at CL16 is not faster than 3200 at CL14.

 

[JimmyDoogs]

Is the 285k faster in 4k than the 9950x3D w/ CUDIMMs?

 

[Wirko]

the datalines are going into a buffer logic ic

No, there is nothing on the 64 (plus ECC) data lines between the IMC and the DRAM chips, except in certain types of server memory:

FBDIMM DDR2 (Fully Buffered)
LRDIMM DDR3/DDR4 (Load Reduced)
MRDIMM DDR5 (Multiplexed Rank)

 

[TheinsanegamerN]

What do you mean by this?

He's one of those guys who thinks that unless you test at "realistic" standards (this being the settings they themselves would personally use) then benchmarks dont matter, for example "why would you run 720p for your CPU test nobody runs that you should run settings people actually use". or "why would you test anything other then 4k ultra on the 4090 nobody will use it with any other settings in real life!"

 

[Onasi]

Is the 285k faster in 4k than the 9950x3D w/ CUDIMMs?

I don’t think that @ir_cow has a ready access to a 5090 (4090 will not be enough) to test this, but I don’t think that will be the case. The massive gap in some CPU limited scenarios that the 285k has to X3D just cannot be overcome on memory alone. The biggest pain point is BG3. Like here:

This is just an architectural diff. You aren’t getting 60 more FPS (50%!) from memory.

 

[Dr. Dro]

The CUDIMM kits work very well on AM5 in bypass mode.

That being the operative; unfortunately in bypass mode it is basically no different from any other UDIMM

Still, nice result, man.

Is the 285k faster in 4k than the 9950x3D w/ CUDIMMs?

I'm afraid there is no scenario where the U9 285K will outperform the 9800X3D or 9950X3D in gaming. It's a very poor performer in that regard, worse even than its Raptor Lake predecessors.

This gives me the shivers

I think my previous Am4 system and my current am5 system has something in the 60 nanoseconds range. Not something in 100 nanoseconds or 82 nanoseconds. Assuming it's the last number of the first row of teh AIDA benchmark

Yeah, the 285K is nasty with latency. Mine's not even optimized to the fullest potential, it's an overclocked G.Skill Z5 6800 A die kit I brought over from my initial Z690 config. Cow reviewed it, great stuff.

 

[JimmyDoogs]

From what I'm seeing the 285k and the 9950x3D are withing 2 FPS in 4k. Can the CUDIMMs push it past that extra few FPS?

 

[Dr. Dro]

From what I'm seeing the 285k and the 9950x3D are withing 2 FPS in 4k. Can the CUDIMMs push it past that extra few FPS?

The problem with this statement is that it's exceptionally generic and hinges entirely on the fact that said games are GPU bottlenecked, so no one can really answer this question based on these parameters alone. Ultimately you need to see if the application or game you intend to use respond well to the 285K's strengths. The architecture is great, Lion Cove is a very powerful P-core and Skymont indeed reaches crazy high IPC rates comparable to Golden Cove, but the chip is marred by high inter-core latencies (similar to cross CCD latencies on Ryzen even to the neighboring core) and has bad memory access latency, which really, really hurts its performance in some scenarios, one of which tends to be gaming, even at high resolutions.

If you're using, say, an RTX 4080, you'll notice it in very few games. Up that to a 4090, and you'll feel it on some more. The 5090 may paint an entirely different story, as will the cards that come to replace it over the years.

 

[mechtech]

Is JEDEC 6400 flat 52s really a very apples-to-apples comparison? Probably sky high tRFC too

What's weird is jedec 3200 was 22-22-22, strange that 6400 isn't 44-44-44?!?!

 

[forgotten05]

In supporting system components page, we have psu and ssd there besides the ram and motherboard. Ram and motherboard seems relevant but what is the point of ssd and psu? Do we need cudimm compatible psu and ssd for cudimm support, I'm confused. I thought it was related to the ram,cpu and motherboard support.

 

[Onasi]

From what I'm seeing the 285k and the 9950x3D are withing 2 FPS in 4k. Can the CUDIMMs push it past that extra few FPS?

In GPU bound scenarios, yes. The difference there is academic and irrelevant. There are obviously games where even a 5090 will be tapped out and most decent modern CPUs will be neck and neck. Such scenarios are not actually indicative of CPU performance at all. But in games that hammer the CPU in a more meaningful way the difference is far larger than just 2 FPS. It’s just that such games are comparatively rarer due to how modern engines work and the nature of most AAA games. So if you’re just looking at the averaged 4K framerates - don’t. Look at the individual games and discard those where all CPUs tested seem to be within a margin of error distance from each other.

 

[mechtech]

I can share this when I get home later tonight. Doesn't matter if its this UDIMM or CUDIMM, the JEDEC timings for each frequency tier will remain the same.

One nice thing about stock DDR4-3200 crucial is the gobs of speeds in the spd, unlike a lot of "gaming ram" which only had up to 1067 (2133) and xmp was 3200....back in the ddr4 days anyway.

 

[JimmyDoogs]

Thanks for the great answers to my questions guys. I do want to pick up the 9800x3D or 9950x3D but I was curious about CUDIMMS in relation to 4k. Very interesting stuff. Much obliged!

 

[freeagent]

Dammit AMD pull your balls together and fix the damn latency lol..

I don't even want to show you what mine is right now.. you would all laugh..

But it "feels" no different than before. The bump to 2R is what I can "feel" in game..

Ugh. Dammit.

 

[Dr. Dro]

Thanks for the great answers to my questions guys. I do want to pick up the 9800x3D or 9950x3D but I was curious about CUDIMMS in relation to 4k. Very interesting stuff. Much obliged!

CUDIMMs are not currently supported on AMD platforms. They work, as you see in the thread, but in bypass mode - this means the CKD chip which is what amplifies the signal and makes the CUDIMM what it is remains disabled. So unless you're paying the same or less against what an equivalent UDIMM kit costs, it is not worth that premium.

If you're going with a 9800X3D for gaming, grab a low latency 6000 MT/s kit rated between CAS 26 to 30, that is the sweet spot. Fast memory is only really worth it on Intel, since you need to reduce fabric and interconnect multipliers when using fast RAM on AMD, largely negating the boon from the bandwidth increase in real world scenarios.

 

[tfp]

That being the operative; unfortunately in bypass mode it is basically no different from any other UDIMM

Still, nice result, man.



I'm afraid there is no scenario where the U9 285K will outperform the 9800X3D or 9950X3D in gaming. It's a very poor performer in that regard, worse even than its Raptor Lake predecessors.



Yeah, the 285K is nasty with latency. Mine's not even optimized to the fullest potential, it's an overclocked G.Skill Z5 6800 A die kit I brought over from my initial Z690 config. Cow reviewed it, great stuff.

View attachment 392644

265k isn't great either but memory latency with 8200 CUDIMM, Ring at 4Ghz, D2D at 3.2Ghz, and NGU at 3.2Ghz is similar to Makaveli's 9800X3D latency.

AID64 isn't as accurate with Hypervisor enabled, I might try booting in safe mode just to see if it changes.

After booting to safemode, looks like CPU clock speed was down impacting Cache scores but memory latency dropped as well, nice work Windows.

 

[Crackong]

does it matter for X3D CPUs ?

 

[tfp]

does it matter for X3D CPUs ?

Yes if a game or application uses more than L3 cache size it will still matter but not to the same extent as intel with smaller caches?

 

[Makaveli]

does it matter for X3D CPUs ?

Yes but not as much the vanilla chips get impacted by this more.

 

[IMABEARLOL]

I'm excited for CUDIMM for the future. This article was very informative in explaining the technology.

However, the benchmarks are misleading and almost disingenuous. You are comparing apples to oranges in a config that almost no one will use (6400MT JEDEC) in what seems like an effort to make CUDIMMs look absolutely amazing.

It would be more... unbiased ... to compare the extreme clocked CUDIMM in XMP to an extreme clocked UDIMM with XMP timings. To show what maximizing memory clocks offers, as that is the intent of CUDIMMs, that is what they offer us right now over UDIMMs.

It would instill trust in your readers to present a more fair and unbiased comparison if your intent was to inform them and not influence them. If the technology has merit it will speak for itself, but by presenting it this way you are misleading those who might not fully understand and insult those with more understanding.

Truly complimenting your writeup though. Even if I would have done it differently, I couldn't have done it so eloquently. Please don't mistake criticism as a lack of recognition of achievement.

 

[biffzinker]

I'm excited for CUDIMM for the future. This article was very informative in explaining the technology.

However, the benchmarks are misleading and almost disingenuous. You are comparing apples to oranges in a config that almost no one will use (6400MT JEDEC) in what seems like an effort to make CUDIMMs look absolutely amazing.

It would be more... unbiased ... to compare the extreme clocked CUDIMM in XMP to an extreme clocked UDIMM with XMP timings. To show what maximizing memory clocks offers, as that is the intent of CUDIMMs, that is what they offer us right now over UDIMMs.

It would instill trust in your readers to present a more fair and unbiased comparison if your intent was to inform them and not influence them. If the technology has merit it will speak for itself, but by presenting it this way you are misleading those who might not fully understand and insult those with more understanding.

Truly complimenting your writeup though. Even if I would have done it differently, I couldn't have done it so eloquently. Please don't mistake criticism as a lack of recognition of achievement.

Except an OEM prebuilt is likely to have the JEDEC 6400 installed.

 

[InVasMani]

I feel like CKD's will end up integrated on motherboards eventually as well midway between CPU and system memory. I imagine possibly on the rear rather than front because it could help with RF given CPU and memory are on the front side alternating a CKD chip on the rear side in between is a bit akin to like twisted wires for Ethernet or audio to help improve signal integrity. Essentially a CKD is weight average improvement of signal integrity to boost up MT/s.

 

[T1beriu]

Did you use DDR5-6400 with dogshit CL 52? Seriously? Did you actively searched for the worst kit available on the market?? 90% of available kits are in the CL 30-36 range, at that speed. Useless test and a waste of your time and the time of your readers. Do better! Much better!

LE: 6400 CL52 was CUDIMM, not regular DDR5 as I thought. The comparison makes sense.

 

[bug]

If these kits don't clearly state the need for motherboard support, I foresee a surge in posts asking why people can't get their systems to boot with RAM at max advertised speeds