Intel's Arrow Lake may support 10,000 MT/s CUDIMM memory, leaving AMD in the dust

[zohaibahd]

Something to look forward to: Intel is poised to elevate desktop memory speeds with its next-generation Arrow Lake processors. The upcoming Core Ultra 200 desktop CPUs may support impressive DDR5 speeds of up to 10,000 MT/s, web rumblings now suggest.

Enabling these transfer rates is a new memory standard called CUDIMM (Clocked Unbuffered Dual Inline Memory Module), which incorporates an on-module clock driver chip to regenerate the signal, ensuring better stability at high frequencies. This standard represents an evolution of traditional desktop UDIMM RAM.

Currently, Intel's top Raptor Lake Refresh chips struggle to exceed 8,200 MT/s, even with the fastest RAM kits available. While 8,200 MT/s is certainly impressive it significantly outpaces AMD's Ryzen 9000 series chips, which tend to max out around 6,400 MT/s – support for 10,000 MT/s represents nearly a 20 percent increase. This could further widen Intel's lead over Team Red.

It's worth noting that the official JEDEC specification only mandates a base speed of 6,400 MT/s for CUDIMM. However, all announced CUDIMM kits have launched above 9,200 MT/s, with some reaching an impressive 9,600 MT/s right out of the box. Memory makers like Asgard have even teased future 10,000 MT/s models on the horizon.

Motherboard manufacturers are also gearing up for CUDIMM. Leaked specifications for ASRock's upcoming Z890 Taichi board indicate support for memory speeds of up to 9,200 MT/s in optimized configurations. With both CPU and motherboard support in place, 10,000 MT/s CUDIMM seems plausible for Arrow Lake.

Additionally, the leak suggests that Intel's LGA-1851 Arrow Lake socket will support CUDIMM via a "Gear 2" mode, where the memory controller operates at half the DRAM data rate. This approach allows the CPU's memory fabric to stay within specification while taking advantage of CUDIMM for ultra-high transfer rates.

While CUDIMM offers incredible bandwidth potential, there are compromises. As Tom's Hardware points out, running at DDR5-9000 speeds and above requires substantial 1.45-1.5V DRAM voltages – up to 36 percent higher than the specification. Pushing modules to these limits raises concerns about heat output and the potential impact on lifespan. Therefore, you'll likely need excellent cooling solutions to accompany these high-performance modules.

Permalink to story:

Intel's Arrow Lake may support 10,000 MT/s CUDIMM memory, leaving AMD in the dust

 

[Watzupken]

It’s fast, but it won’t be affordable. So I expect only enthusiasts or those with deep pockets will get ram this fast.

 

[Eldrach]

It’s fast, but it won’t be affordable. So I expect only enthusiasts or those with deep pockets will get ram this fast.

Not only that - but memory speed benefits has a drastic decline both at higher resolutions and based on the bottleneck from the cpu. Expect a 2-3% increase (or less)from 6000 to 10000 at 4k resolution

 

[yRaz]

Isn't the next benchmark for AMDs next boost from memory speed 12,000? I thought AMD only benefits from speeds greater than 6000 in niche applications and the ideal choice is going for lowest latency 6000 ram instead.

 

[VariableSpike]

Isn't the next benchmark for AMDs next boost from memory speed 12,000? I thought AMD only benefits from speeds greater than 6000 in niche applications and the ideal choice is going for lowest latency 6000 ram instead.

Think it would be 9000 since its to do with the infnity fabric running at 3000 and wanting a whole number multiplier so it doesn't have to mess with the infinity fabric link, though its not like 9000 would unlock that much performance anyway, especially not for anyone running an x3d cpu, with the whole comments about this being a different type of memory and needing a lot of voltage and running hot, its reminding me of a certain other type of memory of the past that was intel only and a bit of a nightmare, though this is still udimms (looking at you, RDRAM....)

 

[VitalyT]

The heat from those over-powered memory modules will melt your PC case.

 

[Rumiku]

Think it would be 9000 since its to do with the infnity fabric running at 3000 and wanting a whole number multiplier so it doesn't have to mess with the infinity fabric link, though its not like 9000 qould unlock that mich performance anyway, especially not for anyone tunning an x3d cpu, with the whole comments about this being a different type of memory and needing a lot of voltage and running hot, its reminding me of a certain other type of memory of the past that was intel only and a bit of a nightmare, though this is still udimms (looking at you, RDRAM....)

Rambus...anyone of computer building age and interest during that time remembers vividly learning, especially if for the first time, how a company can be so evil and vile, and yet get away with it on a technicality.

People will scorch Intel, AMD, Nvidia, Facebook et all, but Rambus earned a special place in Technology hell for their 'contribution'.

 

[opckieran]

Will CUDIMM stave off the adoption of CAMM2? It may prolong the DIMM form factor for an additional generation on desktops, but laptops will still have to contend with Z-height and power constraints. Perhaps the technology will be implemented in (LP)CAMM2 offerings or even within a future DDR standard at some point in the near future.

 

[letsgoiowa]

The heat from those over-powered memory modules will melt your PC case.

Wat? Memory uses so little power it is passively cooled in most cases lol. Even if you had to strap a fan to it it's literally like 1/20th of a 4090

 

[Cabuzzi]

Think it would be 9000 since its to do with the infnity fabric running at 3000 and wanting a whole number multiplier so it doesn't have to mess with the infinity fabric link, though its not like 9000 would unlock that much performance anyway, especially not for anyone running an x3d cpu, with the whole comments about this being a different type of memory and needing a lot of voltage and running hot, its reminding me of a certain other type of memory of the past that was intel only and a bit of a nightmare, though this is still udimms (looking at you, RDRAM....)

Funny how many people remember that fiasco. I still remember when the leaked Intel slides came out showing PC133 being faster.

 

[Cabuzzi]

Rambus...anyone of computer building age and interest during that time remembers vividly learning, especially if for the first time, how a company can be so evil and vile, and yet get away with it on a technicality.

People will scorch Intel, AMD, Nvidia, Facebook et all, but Rambus earned a special place in Technology hell for their 'contribution'.

Agreed. That may be one of the filthiest things that ever came out of that era, and Intel was right there with them, holding hands, skipping along, and laughing about all the money they thought they were going to make.

 

[YouShallNotPass]

Wat? Memory uses so little power it is passively cooled in most cases lol. Even if you had to strap a fan to it it's literally like 1/20th of a 4090

"running at DDR5-9000 speeds and above requires substantial 1.45-1.5V DRAM voltages – up to 36 percent higher than the specification"


Technology Typical Voltage
DDR 2.5 V
DDR2 1.8 V
DDR3 1.5 V

We were there and 1.5v definitely won't melt our PC cases.

 

[hahahanoobs]

August we hype CAMM2; September we hype CUDIMM.

 

[ThomasTECH]

Way to go Intel! I think this advancement is just the first to come. The technology they are using to hit these targets for mega transfers has additional applications, I expect to see much more coming from them in the next 5 years.

 

[TooQik]

Think it would be 9000 since its to do with the infnity fabric running at 3000 and wanting a whole number multiplier so it doesn't have to mess with the infinity fabric link...

If multiples of 3000 is the driver, then 12000 would be fine. As yRaz pointed out, it's effectively doubling the current 6000 "sweet-spot".

 

[Fanboiism101]

Sure, I'll spend an exorbitant amount of money to buy high speed memory instead of saving for other and better hardware to get ~15% improvement in my performance. What a sweet deal it will be.

 

[Subset6]

My friends we've reached the end of the classic GPU as we know it. We had a good run. Synthetic CPU/GPU models are where the industry is headed. Unified packaging is the only cost effective way forward, you simply need a memory pipeline to the end user and that's it. No local rendering, no user control..

 

[bandit8623]

Will CUDIMM stave off the adoption of CAMM2? It may prolong the DIMM form factor for an additional generation on desktops, but laptops will still have to contend with Z-height and power constraints. Perhaps the technology will be implemented in (LP)CAMM2 offerings or even within a future DDR standard at some point in the near future.

Camm is just the form factor. The technology will be intertwined

 

[seeprime]

Agreed. That may be one of the filthiest things that ever came out of that era, and Intel was right there with them, holding hands, skipping along, and laughing about all the money they thought they were going to make.

Socket 423 Pentium 4 was so pathetic with its unnecessary long pipeline that my 1GHz P3 outperformed a fresh 1.3 GHz P4 with Rambus. Intel took a dump on users wirh P4 systems, to n general. Some genius at Intel decided that everything was going to be video, where the large pipeline would have been of use. Sometimes large corporations just do stupid things.