Asgard launches world's first DDR5-9600 DIMMs | Hacker News Reader

Asgard launches world's first DDR5-9600 DIMMs

tomshardware.com

84 pointsapsec11257 comments

I am only hypothesizing here, but I do not expect these to be useful on AMDs current AM5 platform due to the design of the memory controller.

The reason is there is currently a drop off in memory subsystem performance when you go above DDR5-6000 speeds on Ryzen 7x00 (and AFAIK Ryzen 9x00), because above that, the memory controller inside the chip can't run in lock-step with the memory frequency. Much of the enthusiast memory out there that is capable of higher speeds already, like say DDR5-7200, doesn't pair well with AMD chips as a result, as in that case the memory controller is forced to work at a 1:2 instead of a 1:1 ratio, introducing latency into the memory subsystem that results in overall worse performance.

On the other hand, if these CUDIMMs can hit high enough frequencies (like 10k+ MT/s,) it's possible that the benefits will outweigh the negative related to this quirk of the memory controller. Until I see this tested or someone with more expertise clarifies this point, it's hard to say. Perhaps DDR5-9600 will result in better performance than the current DDR5-6000 "sweet spot".

What goes unsaid in the article is that this will likely only work with Intel's Arrow Lake? Arrow Lake isnt being launched until next month though, so official specs are not available yet. And the full 9600 speed will very likely require overclocking.

It doesnt appear Zen5 can get anywhere close those speeds.

These are part of a new generation of DDR5 modules that add a clock driver to the module itself, rather than relying entirely on the clock signal coming all the way from the CPU's memory controller. That is expected to significantly change the game for DDR5 overclocking, and expectations based on the limitations of configurations lacking the clock drivers may not be accurate. I expect we'll see 6+ months of firmware tuning before we have a clear picture of what each platform is ultimately capable of, but early on the focus will be on Intel platforms.

How does that work? I imagine you now have the feed to clock from ram back to the CPU? Or does it synchronize two local clocks somehow?

The clock redriver uses a phase locked loop to generate a few identical clock signals for use by the DRAM chips on the module. This same tech is also used for clock multipliers and dividers throughout digital systems.

Rambus has details on their "DDR5 client clock driver" here: https://www.rambus.com/memory-interface-chips/ddr5-client-di...

This anandtech article seems to be a decent summary of what it's doing, the JEDEC JESD82-531A.01 spec is also freely available (with registration) if you're feeling adventurous

https://www.anandtech.com/show/21455/making-desktop-ddr5-eve...

I could see this maybe reducing latency, but wouldn't you still be limited by the memory bus speed regardless of how high you could clock the DDR?

The bus clock and memory clock are the same.

Not an expert, but Zen 5 desktop feels like it is being held back by memory bandwidth. Several tests show there is no real benefit in >6000 DRAMs.

Bandwidth is currently limited by the infinity fabric implementation on consumer AMD CPUs, 6/8 core CPUs with s single CCD top out bandwidth quite early and improvements from higher clocks are minimal even from like 5600, assuming timings are overclocked similarly and beyond basic xmp/expo

See my top level post for more on this. It's not that there is no benefit to higher frequency memory on AM5, but that there is a penalty introduced above 6000 MT/s which isn't easily outweighed by further raising the memory speed.

Granite ridge can currently do 8000-8400 when overclocked at a 1:2 ratio of memory controller clock to memory clock, though with some work needed to get it working.

The 9600 may be possible with DIMMs using the clock driver even on granite ridge

> It doesnt appear Zen5 can get anywhere close those speeds.

Looks like it can get to DDR5-8400:

> Finally, we have some absolutely killer RAM. The G.SKILL Trident Z5 RGB series is one of that vendor's top brands, and this is among the top models that it offers. Clocked at a scorching 8400 MT/s with a CAS latency of just 40 cycles, this is the fastest RAM we can reasonably recommend for Ryzen 9000. It comes in the form of a pair of 24GB modules, and using it with Zen 5 could be as easy as just flicking on that XMP switch.

* https://hothardware.com/news/zen5-ddr5-memory-guide

Does that count as "anywhere close"?

I assume the 285K will be able run these at their full 9600?

I'd love to see their test fixture for binning the chips.

did they also make a compatible motherboard ? because I can't get mine to go past 4800Mhz when my RAM is rated for 6400Mhz with AM5 CPU.

for those who hadn't done this for a while, PSA: motherboards these days have a RAM compatibility list on their product page, check it before overspending on RAM.

The motherboard QVLs aren't particularly useful, the higher tier kits can often just not work as they're currently more often than not limited by the CPU memory controller quality instead of purely motherboard or RAM ICs.

Did you try 6200/6000 for the kit? 6400 is up there for speeds that may not work on AM5 with synced controller and memory clocks without tuning it yourself, or working at all. Or is that 4 DIMMs

I had similar issue with my 3950x I had stability issues with 3600MT, so I optimized clicks instead (I loaded XMP profile and then changed clock to lower value before tightening clocks with memtest runs in between) What people often don't realize that dimm quality allows you to get lower absolute time before data is available for system. So when you lower clocks, you don't need to wait that many cycles for the data. In that case you might not get that much bandwidth that you would have originally, but you can get the same or similar latency also changing order of dimms on board can somehow alter how low you can get (I have 4 dimms and swapping 2 on the same channel improved significantly clocks that I was able to achieve - no idea why)

Make sure you're on the latest BIOS as many improvements have been made since early AM5.

Anecdata: I just bought a used AM5 mobo and threw in this memory kit CMK64GX5M2B6400C32; hitting 6400 was as simple as turning on XMP and that's it.

I haven't done this for a while, but every time I come back around the new specs with even basic hardware blow the doors off my old rig.

More than the specific list of which RAM is compatible check the table of how many slots/GB is supported at which speed. 1x 8 GB DIMM is a lot easier to drive than 4x 48 for example. Also keep in mind a lot of it depends more on the CPU than the motherboard itself, though both play a part, and the QVL should be a "safe bet". Particularly if you're willing to throw a bit of extra voltage at the RAM.

CUDIMs should help a bit with those scaling tables due to having the clock redrivers, but we'll have to see to be certain how much.

And they’re often not stable even with the given XMP profiles. All of my XMP profiles throw errors and require backing off speed by a few hundred hertz to be stable.

Worth though. Much cheaper way to get a few extra percentage points than custom liquid cooling. And since every game where you would care about max FPS is bound by single threaded performance fast RAM is it.

> 1x 8 GB DIMM is a lot easier to drive than 4x 48 for example.

That's a bit of an exaggeration. What matters are DIMMs per channel, and ranks per DIMM. Consumer systems have at most two memory slots per channel, so 2x 8GB modules installed correctly is just two separate instances of one DIMM per channel, and 4x48GB is two separate instances of two DIMMs per channel, with each module being dual-rank.

The best configurations for overclocking are supposed to be motherboards with only one memory slot per channel, so that when operating with one DIMM per channel there are no empty slots providing stubs of wiring that degrade signal integrity. But very few motherboards restrict themselves like this for the sake of memory overclocking.

> But very few motherboards restrict themselves like this for the sake of memory overclocking.

Though I'd hope they'd finally start doing it considering the per DIMM capacities DDR5 brings (32 GB, with 64 GB planned), and most consumers really not going for capacities that'd require 2 DPC (DIMM per channel).

1 DPC should've been the default by now with the amount of 1 DPC and 2 DPC boards swapped, as it's just making things worse in the most common use case

A lot of consumers care more about either of total capacity or the concept of "I can just double my cheap RAM later" than care about ~10% differences in RAM clock speed. The latter is for those that really know what they're doing to eek out a couple more percentage points in benchmarks or non-consumer use cases, the former appeal to both budget users and prosumers alike (just not OC enthusiasts).

From that angle I don't necessarily disagree with 4 channel boards being more common than 2 channel boards, even though I personally lean more towards a 2 channel board myself (well, depending how well CUDIMMS scale I may change my opinion).

> I can just double my cheap RAM later

The thing is that this is quickly becoming impossible, even with fairly low speeds like 6000 2 DPC is not achieveable in some cases, and consumers that are buying the 4 DIMM boards aren't particularly aware of it

Even 6000 MHz isn't necessarily "cheap low speed" RAM though, it's still a 400 MHz OC on Zen 5/14th Gen. ~$44 gets you 2x8 GB 4800 but it's ~$63 to start with 2x8 GB 6000.

Simplifying it to one vendor for the sake of conversations: for boards that explicitly target 9800X OCing users it's a bit silly to see 4 slots all the time but for the majority of boards that target the whole lineup it's not nearly as silly. Same for the 9900/9950X class where a lot of the users might be intentionally looking for larger amounts of slow RAM for heavier applications, not necessarily the absolute best bandwidth or latency the processor can manage.

> Even 6000 MHz isn't necessarily "cheap low speed" RAM though, it's still a 400 MHz OC on Zen 5/14th Gen. ~$44 gets you 2x8 GB 4800 but it's ~$63 to start with 2x8 GB 6000.

I'd argue that whatever JEDEC specs intel/AMD puts up are mostly irrelevant to the reality of things, and the upgrade minded people won't be limiting speeds because of their speculative future scenario. This only leaves people who are on a very constrained budget and can't buy all the RAM at once. And they could buy a single 16GB stick and then buy another one later on when necessary. The losses from only using one channel aren't going to be as pronounced as on DDR4 either because of the 8GB's halved bangrkups and DDR5's internal channel split.

> Same for the 9900/9950X class where a lot of the users might be intentionally looking for larger amounts of slow RAM for heavier applications, not necessarily the absolute best bandwidth or latency the processor can manage.

I think DDR5 changes things up a bit here as 64GB is currently easily used and is enough for most, even heavier, workloads. Like for example for HN's audience some programming without worrying about memory being eaten up outside some more extreme cases. I agree that there's still a lot of people who'd need the ram even accounting for that, but they don't exactly need 80% of boards to have 4 slots

According to PCPartPicker, 2x48GB kits at DDR5-5200 are currently the cheapest DDR5 on a $/GB basis, so you really need to have a workload requiring more than 96GB of RAM to justify having more than two DIMM slots. That's pretty solidly into workstation territory.

Is it far off in terms of $/GB in a normal user starting with 2x24 GB 5200 at ~half the up front cost while still being able to upgrade 4 years later, if needed, when memory is even cheaper and RAM requirements are higher versus all that pro/con difference against being able to run 2 high end DIMMs at 6400 instead of 6000 or similar difference by limiting the number of slots on the board. Again, I still think this makes plenty of sense for boards focused on OC enthusiast but paying more for the 400 MHz on a tradeoff of not being able to expand just doesn't seem particularly worth it on normal boards for normal users.

Or, perhaps stated differently, other than enthusiasts purchasing the best binned low density RAM who else has what practical gains by general consumer boards having fewer slots?

People on HN can probably be assumed to want to go past JEDEC 2x48 GB RAM every time, I, again, wouldn't argue what HN should do as what the average consumer should be expected to do though. Much like I don't expect to recommend the average HN reader get the 6 core variant but it doesn't mean it's not as popular to general consumers due to price.

Same with "64 GB being easily used", true or not it's still not the capacities most are buying in new builds (32 seems to be the new common target for gamers, with 16 being the new minimum there, for example)

4x 8 GB is still a lot easier to drive at high speed than 4x 48 GB, but yes, one factor in the above comparison is indeed the number of DIMMs per channel. Ranks per DIMM is definitely a factor as well, though on large capacity DIMMs that tends to be a one sided story. Another factor in the rabbit hole is banks, tying it all back.

The other part of 2 DIMM boards is the trace lengths can be that much shorter (they need to be the same length so fewer slots means less max length to match to).

Your issue is more with AMD CPUs having poor quality memory controllers; not the motherboard.

> PSA: motherboards these days have a RAM compatibility list on their product page, check it before overspending on RAM.

That's a valuable PSA, but FYI it's been that way since the 80s. I've been bitten by compatibility issues at least once a decade since the 90s myself, and the answer from forums or support is always "did you check the compatibility table tho?".

Some memory manufacturers have lists of motherboards compatible with a given memory kit. E.g. I checked that my motherboard was listed in [1] before buying the memory, and everything worked effortlessly together with EXPO on.

[1] https://www.gskill.com/qvl/165/393/1665020366/F5-6000J3040G3...

People should be much more sceptical than they are being here. Setting the xmp profile of Hynix or Samsung memory chips is right up there with formatting usb keys way over spec. Did you know Hynix or Samsung already binned these themselves and set them to a lower spec?

The number of performance memory chips I’ve had that had to be downclocked from their xmp profile for stability is ridiculous. It’s gotten to the stage where I’d recommend only buying memory from the first party suppliers (eg. only buy Samsung or Hynix memory). The gaming companies like the above are simply taking that memory and overwriting the xmp profile while adding some flashy but useless heat spreaders.

> Did you know Hynix or Samsung already binned these themselves and set them to a lower spec?

Hynix, Samsung, and Micron bin the chips for what speed they can operate at with a 1.1V supply. These enthusiast modules are using chips binned for the speed they can hit with 1.4V. The XMP/EXPO profiles don't come from Hynix, Samsung, or Micron.

Also, the vendors announcing these things so far are ones I've literally never heard of before. I'd still be a bit skeptical, but I think it would actually mean something if Kingston, G.Skill, Corsair, or Teamgroup announced this product. As it stands, I have no reason to think that I'll ever be able to actually buy this product.

> flashy but useless heat spreaders

The funny thing about heat spreaders on SDRAM is that they originated as a skeuomorph of RDRAM modules from the early Pentium 4 days, but everyone's pretty much forgotten that by now, so the market is essentially flooded with "retro" memory for no obvious reason except "everybody else does it".

Stability problems might very well come from incorrect voltages. Since XMP profiles do not contain settings for most voltages, they're set by the motherboard firmware, and those values are often too high. I use XMP/EXPO profiles as a starting point, they're almost always terrible by themselves.

Do higher frequency RAM provide an actual noticeable improvement in performance?

Genuinely curious.

I'm only aware of the differences with manual timings but if you're not running the bottom of the (1 DIMM per channel) barell frequencies like 5200-5600 differences from going higher should be minimal. Not that many things actually scale with bandwidth so the higher frequencies are mostly helping because ther'll be shorter delays between operations where the registers for the timings bottom out at a higher number.

AMD also needs to change the sync of memory controller clock to memory clock to 1:2 instead of 1:1 above ~6000-6400 so as some anecdata 7800 performed roughly the same as 6000

I think Asgard is VCuller.

China based firm named Asgard with product names like Thor? Whats going on here?

This is happening a lot in many product categories as a way to market effectively for international audiences, without some of the mistakes Chinese firms have historically made. Borrowing from other cultures is easier. For example in audio, Hifiman is a Chinese audio company that previously used alphanumeric names for their headphones but now uses various Hindu or Sanskrit (?) words to name their headphones and give them more of a “theme”.

From a quick glance at the website, the only two such products I can see are the "Susvara" (meaning "with notes") headphones and "Deva" (meaning "god") series headphones. Both of those words are Sanskrit. Their hindi equivalents are quite similar sounding (Susvar and Dev).

They also sell/have sold the Arya, the Sundara, and the Ananda.

I assumed it's to attach to the "as" naming prefix, asus, asrock, asmedia, etc.

What does that prefix mean?

Nothing. It's Pegasus with the first two letters removed so they'd be at the top of the phone book when we used phone books way back in 1989.

It's a big step up from names like FDSOOJVX.

I can always trust POYICCOT to sell weird USB-C adapters. It's a thing

Heck, If I look real hard I might even find some California based companies that don’t pull exclusively from ancient Californian lore.

The US can trace its lineage directly back to Europe. And many of the companies in the US are founded by recent immigrants (0-2 generations, worldwide), so it makes sense that it would have a varied history to pull from that includes European mythology. The same can't really be said of China.

That being said, ancient mythology isn't something worth gatekeeping. And anyone who gets real uppity about (specifically) Norse or Celtic mythology instantly starts triggering "closet/academic racist" vibes.

Stargates and wormholes.

Marketing... These not-so-famous-yet Chinese firms have picked up what the West has been doing and are trying to cut out the middle-man. Essentially the West's MO has been to buy cheap no-name generic stuff from china, have them colored/branded/packaged, and then slap on all the necessary product packaging stuff on the packaging (including a "Name"). Essentially acting as a middle-man that doesn't actually produce the product, barely does any QA and just adds a thin veneer of this being an actual "company" that manufactures and sells this product. Sometimes they use a fancy term like "white labeling".

Here is such a company in the South African context that's super obvious about it: https://www.loot.co.za/search?cat=b&brandcode/leisurequip

they are sg1 fans

Maybe something about hammers ? Or rainbow. They said they binned that korean chips, wonder how ? ;)

Anyway, we are better not buying pure american chips (they burns) and soon to be on terrorists linked list if we own one - x86 goes military ! I'm somewhat worry about them now - x86 means Windows everywhere ?

And that "AI PC", whatever it is, I already hate it...

Reply on news.ycombinator.com