Monday, February 2nd 2026

Intel Mandates 7,467 MT/s+ Memory for "Panther Lake" Arc B390/B370 Integrated Graphics

Intel is reportedly mandating OEM integrators of the latest "Panther Lake" SoCs to use LPDDR5X memory starting at 7,467 MT/s and beyond, with an interesting software differentiator between adequate speeds or those below. According to Golden Pig Upgrade, LDDR5X memory running below the 7,467 MT/s threshold will force the software to display a generic mark of "Intel (R) Graphics," while configurations with that exact memory speed or higher will display the full name of "Intel (R) Arc (TM) Graphics B390" or "Intel (R) Arc (TM) Graphics B370" marking in Windows 11 Task Manager. Reportedly, Intel is doing this to stop OEMs from cutting corners with their "Panther Lake" laptop configurations, where they could bundle a lower-speed LDDR5X memory out of the Intel specification.

Without the required memory, Intel's Arc B390/B370 iGPUs would likely be left starving for bandwidth, as the whole SoC is being powered by that memory. At higher speeds, memory can do the data transfer faster, resulting in a boost of the overall system performance, and most importantly, the frame rate that these chips can push out. Especially for single-package systems like "Panther Lake" is, faster memory is a great way to get extra performance boost. Intel's flagship SKUs are advertised as capable of running with LPDDR5X memory at speeds up to 9,600 MT/s, which is just below the point where LPDDR5X technology tops out at 10,700 MT/s. Interestingly, there are lots of options from Intel's OEM partners that integrate top-end memory, which is a positive sign for the ecosystem.
Additionally, Intel is developing a separate "Panther Lake" SoC focused on handheld devices, likely featuring a new Arc B SKU rumored to be named Arc B380. This version is expected to be more power-efficient, possibly through selective die binning, maintaining the full 12 Xe cores configuration at slightly reduced clock speeds to deliver similar performance with a lower TDP. With a minimum memory speed requirement, these SoCs are expected to perform well, and OEMs might even use the fastest LPDDR5X running at 9,600 MT/s to achieve optimal handheld performance.
Sources: Golden Pig Upgrade, via VideoCardz
Add your own comment

23 Comments on Intel Mandates 7,467 MT/s+ Memory for "Panther Lake" Arc B390/B370 Integrated Graphics

#1
ixi
Where is B770? -,..,-

Back to topic. Performance from what i have seen - decent for iGPU. Only question, how much does the cpu cost... mini pc with this would be great.
Posted on Reply
#2
vbq7qK68eyYAH4iR
Mere Mortal
I like this. Given that the iGPU seems to have some real horsepower, it makes sense to "force" OEMs to hold up their side of the bargain and pair a device with the right hardware, so consumers don't getting screwed. I do wish they'd have put the frequency a little higher, but it's a good base line.

Off Topic. I was watching The Phawx's review of the Intel X9 388H Pantherlake and in it he mentioned that it includes a memory compression feature, that seems to make its limited memory bus far less of an issue. This might be a good area for TPU to do an article on, exploring what this compression is exactly, as I hadn't heard of it in previous Panther Lake content, what work loads will benefit from it and a prediction/napkin math on how it helps Panther Lake compete with Strix Point and Strix Halo. Maybe an interview with Tom Peterson but it and how Intel came up with it, how they selected the specific compression algorithm and why, do they see it playing a part in their desktop environment, like AMD tried to with SAM and Zen 3.
Posted on Reply
#3
TheinsanegamerN
Good. We all know OEMs would cheap out and use 5200mt SODIMMs if allowed. AMD should have done the same with the 300 series, the HX 370 loses upwards of 40% of its performance when not using LPDDR5x.
Posted on Reply
#4
soder
Yeah, agree to #1: Where.Is.B770?
Posted on Reply
#5
Chaitanya
ixiWhere is B770? -,..,-

Back to topic. Performance from what i have seen - decent for iGPU. Only question, how much does the cpu cost... mini pc with this would be great.
Performance we have seen is only for the "flagship" class of GPU and most CPUs wont be using that iGPU so better wait for the reviews of other SKUs before making a buying decision.
Posted on Reply
#6
tommo1982
AMD and Intel should include RAM with their SoC's. HBM is out of the question due to cost, however LPDDR5 should be fine. The GPU is integrated, it's target is not games, so 4GB or 8GB of dedicated memory could do wonders. Possibly replace low end dGPU's in some cases.
Posted on Reply
#7
ncrs
tommo1982AMD and Intel should include RAM with their SoC's. HBM is out of the question due to cost, however LPDDR5 should be fine. The GPU is integrated, it's target is not games, so 4GB or 8GB of dedicated memory could do wonders. Possibly replace low end dGPU's in some cases.
Intel did this recently with Lunar Lake.
Posted on Reply
#8
R0H1T
AleksandarK9,600 MT/s, which is where the LPDDR5X technology tops out.
It tops out at ~10.6 GT/s right now.
Posted on Reply
#9
sam_86314
Hopefully they'll allow LPCAMM to be used unlike AMD with Strix Halo.

Of course, even if it is allowed, no OEMs are going to use it...
Posted on Reply
#10
AntiX3DGlazer
tommo1982AMD and Intel should include RAM with their SoC's. HBM is out of the question due to cost, however LPDDR5 should be fine. The GPU is integrated, it's target is not games, so 4GB or 8GB of dedicated memory could do wonders. Possibly replace low end dGPU's in some cases.
Why? The only benefit is power savings. Literally what elese would having on-package LPDDR do? It's not any faster.
Posted on Reply
#11
hsew
7467MT/s is a far cry from 9600. It's more than a 20% deficit, or nearly a 30% boost, depending on how you look at it. Intel should have mandated 8533 MT/s at the minimum.
Posted on Reply
#12
AntiX3DGlazer
sam_86314Hopefully they'll allow LPCAMM to be used unlike AMD with Strix Halo.

Of course, even if it is allowed, no OEMs are going to use it...
Kinda a dead standard. It doesn't support fast enough LPDDR to make sense for the high end SKU. Highest speed supported is 7467 iirc.
Posted on Reply
#13
jesdals
I did not know that 7,467 MT/s modules was a thing?
Posted on Reply
#15
TheinsanegamerN
AntiX3DGlazerWhy? The only benefit is power savings. Literally what elese would having on-package LPDDR do?
Wow, I cant imagine why power savings would be important on a mobile chip

Hmmmmmm........
AntiX3DGlazerIt's not any faster.
One of the biggest factors in LPDDR5x managing significantly higher speeds is the proximity of the memory to the memory controller on the CPU. So theoretically on package DDR would be able to maintain faster speeds then LPDDR5x could manage.
sam_86314Current max speed is 8533MT/s, and until the DRAM-pocalypse happened, Micron was saying it would hit 9600MT/s in 2026.

semiconductor.samsung.com/dram/module/lpcamm2/
assets.micron.com/adobe/assets/urn:aaid:aem:4e076108-df95-4c2d-8785-06c30049afb5/original/as/lpddr5x-camm2-technical-brief.pdf
We heard the same promises with CAMM and to my knowledge we have yet to see these speeds demonstrated in a retail product.
Posted on Reply
#16
AntiX3DGlazer
TheinsanegamerNOne of the biggest factors in LPDDR5x managing significantly higher speeds is the proximity of the memory to the memory controller on the CPU. So theoretically on package DDR would be able to maintain faster speeds then LPDDR5x could manage.
This is objectively wrong within reason. The speed of the DRAM is not affected significantly by its proximity to the SoC. I submit for your consideration the fact that Panther Lake supports LPDDR5X-9600 while not having on-package memory.
TheinsanegamerNWow, I cant imagine why power savings would be important on a mobile chip

Hmmmmmm........
You're right! However, you know what else it is?
very expensive.
Posted on Reply
#17
TheinsanegamerN
AntiX3DGlazerThis is objectively wrong within reason. The speed of the DRAM is not affected significantly by its proximity to the SoC. I submit for your consideration the fact that Panther Lake supports LPDDR5X-9600 while not having on-package memory.
This is objectively wrong within reason. (as opposed to all the subjective wrongs outside reason I guess?)

www.technalysisresearch.com/downloads/TECHnalysis%20Research%20White%20Paper%20for%20Rambus.pdf

Shorter trace lengths allow greater signal integrity, which allows higher memory speeds. Long trace length leading to poor memory speed is why the industry has been looking into replacements for SODIMMs, its why CUDIMMS can clock higher (the clock generator is on the memory stick instead of the CPU) and it was the biggest curdle for CAMM2 memory speeds. Panther lake supporting LPDDR5x 9600 does not disprove the idea that getting the memory even closer could allow higher speeds. Pro overclocking boards come with only 2 DIMM slots to reduce the number and length of traces to improve overclocking results. I could go on.

You know how VRAM on GPUs has gotten progressively closer to the GPU itself as speeds increase? Do you think they did this for funsies? No, it was critical in maintaining the higher speeds these chips run at. Its also made cooling them and building the PCB a PITA, so if they could maintain those speeds with GDDR5 or GDDR4 spacing they would absolutely do so.
AntiX3DGlazerYou're right! However, you know what else it is?
very expensive.
You know whats fun? Evidence.

Dell offers the pro plus lineup with both intel lunar lake CPUs and AMD Hx series CPUs. Lunar lake has that super expensive on package memory. Guess what? With the same screen, memory, battery, ece, the only difference being the CPU, the highest end lunar lake chip, the Ultra 7 268V is $100 CHEAPER then the HX 370 system it competes with.

Posted on Reply
#18
Cheeseball
Not a Potato
TheinsanegamerNThis is objectively wrong within reason. (as opposed to all the subjective wrongs outside reason I guess?)

www.technalysisresearch.com/downloads/TECHnalysis%20Research%20White%20Paper%20for%20Rambus.pdf

Shorter trace lengths allow greater signal integrity, which allows higher memory speeds. Long trace length leading to poor memory speed is why the industry has been looking into replacements for SODIMMs, its why CUDIMMS can clock higher (the clock generator is on the memory stick instead of the CPU) and it was the biggest curdle for CAMM2 memory speeds. Panther lake supporting LPDDR5x 9600 does not disprove the idea that getting the memory even closer could allow higher speeds. Pro overclocking boards come with only 2 DIMM slots to reduce the number and length of traces to improve overclocking results. I could go on.

You know how VRAM on GPUs has gotten progressively closer to the GPU itself as speeds increase? Do you think they did this for funsies? No, it was critical in maintaining the higher speeds these chips run at. Its also made cooling them and building the PCB a PITA, so if they could maintain those speeds with GDDR5 or GDDR4 spacing they would absolutely do so.

You know whats fun? Evidence.

Dell offers the pro plus lineup with both intel lunar lake CPUs and AMD Hx series CPUs. Lunar lake has that super expensive on package memory. Guess what? With the same screen, memory, battery, ece, the only difference being the CPU, the highest end lunar lake chip, the Ultra 7 268V is $100 CHEAPER then the HX 370 system it competes with.

258V/268V should be cheaper, as it's only 8C/8T at a cap of 37W PL2 with on-package 32GB of RAM. It's good for lightweight ultrabooks and gaming handhelds but it will end up losing to the HX 370 (12C/24T, up to 55W sPPT) on heavier workload.

That being said, that Dell is overpriced for what its offering. For a 16" laptop it should at least be able to handle 55W+ TDP and have two USB4 ports that are separately routed (e.g. both ports have their own dedicated PCI-E lanes for tunneling).
Posted on Reply
#19
Steevo
TheinsanegamerNThis is objectively wrong within reason. (as opposed to all the subjective wrongs outside reason I guess?)

www.technalysisresearch.com/downloads/TECHnalysis%20Research%20White%20Paper%20for%20Rambus.pdf

Shorter trace lengths allow greater signal integrity, which allows higher memory speeds. Long trace length leading to poor memory speed is why the industry has been looking into replacements for SODIMMs, its why CUDIMMS can clock higher (the clock generator is on the memory stick instead of the CPU) and it was the biggest curdle for CAMM2 memory speeds. Panther lake supporting LPDDR5x 9600 does not disprove the idea that getting the memory even closer could allow higher speeds. Pro overclocking boards come with only 2 DIMM slots to reduce the number and length of traces to improve overclocking results. I could go on.

You know how VRAM on GPUs has gotten progressively closer to the GPU itself as speeds increase? Do you think they did this for funsies? No, it was critical in maintaining the higher speeds these chips run at. Its also made cooling them and building the PCB a PITA, so if they could maintain those speeds with GDDR5 or GDDR4 spacing they would absolutely do so.

You know whats fun? Evidence.

Dell offers the pro plus lineup with both intel lunar lake CPUs and AMD Hx series CPUs. Lunar lake has that super expensive on package memory. Guess what? With the same screen, memory, battery, ece, the only difference being the CPU, the highest end lunar lake chip, the Ultra 7 268V is $100 CHEAPER then the HX 370 system it competes with.

8 thread VS 24 threads, single thread Intel wins by a little, but AMD kills them at multithreaded.

AMD kills the it at 35W, for very low watts Apple simply slaughters them both at 6W for a comparable score in single threaded cinebench.

Intel has made a small stride, mostly due to back side power delivery making their nodes more efficient at low power.
Posted on Reply
#20
hsew
Steevo8 thread VS 24 threads, single thread Intel wins by a little, but AMD kills them at multithreaded.
Not anymore, honestly. Lunar Lake is effectively defunct/EOL since Panther Lake Launched.
SteevoApple simply slaughters them both at 6W for a comparable score in single threaded cinebench.
Apple wins in this context if you can live without x86, professionally.
Posted on Reply
#21
DavidC1
sam_86314Hopefully they'll allow LPCAMM to be used unlike AMD with Strix Halo.

Of course, even if it is allowed, no OEMs are going to use it...
Even LPCAMM isn't good enough for Strix Halo. Framework wanted to use it. LPCAMM isn't as close to memory as soldered LPDDR is:
tommo1982AMD and Intel should include RAM with their SoC's. HBM is out of the question due to cost, however LPDDR5 should be fine. The GPU is integrated, it's target is not games, so 4GB or 8GB of dedicated memory could do wonders. Possibly replace low end dGPU's in some cases.
This is a bad idea. Intel stated multiple times in their earnings call that Lunarlake's on package memory is what caused their margin hit and they won't make that mistake again.
hsew7467MT/s is a far cry from 9600. It's more than a 20% deficit, or nearly a 30% boost, depending on how you look at it. Intel should have mandated 8533 MT/s at the minimum.
It's fast enough. Asus Expertbook with LPDDR5-8533 is on par with Zenbook Duo with 9600. Xe3 is well balanced architecture. With 35% scaling 30% difference is only about 10-15% difference in gaming performance. It's still faster than anything else out there. Pantherlake supports DDR5-7200 SO-DIMMs.

Also, the 4 Xe core version is competitive with 8 Xe core version in Lunarlake.
Posted on Reply
#22
AntiX3DGlazer
TheinsanegamerNThis is objectively wrong within reason. (as opposed to all the subjective wrongs outside reason I guess?)

Shorter trace lengths allow greater signal integrity, which allows higher memory speeds. Long trace length leading to poor memory speed is why the industry has been looking into replacements for SODIMMs, its why CUDIMMS can clock higher (the clock generator is on the memory stick instead of the CPU) and it was the biggest curdle for CAMM2 memory speeds. Panther lake supporting LPDDR5x 9600 does not disprove the idea that getting the memory even closer could allow higher speeds. Pro overclocking boards come with only 2 DIMM slots to reduce the number and length of traces to improve overclocking results. I could go on.
I suppose I will be more specific. Signal integrity is a problem that is solved by amplifying the signal so that it is stronger when it reaches destination. This is done by increasing the power given to the memory controller to maintain the same speeds. Thus, iso-speed, a memory controller must use more power to communicate with its DRAM chips if it has to go further along a PCB (or in extreme examples through a retimer/amplifier circuit along the board). Techincally, having the memory closer does mean that you can clock the DRAM higher while maintaining the same power level, but that relies on your memory controller having the headroom to support such clocks (and why wouldn't you just clock the DRAM as high as physically possible regardless?). Speaking to your final example, those pro overclocking boards don't magically improve out-of-the-box memory support, becuase the memory controller becomes the limiting performance factor far before the traces in the PCB or 2DPC slot configuration do. This requires overlcocking which, guess what, increases power! I think we're both right. On-package memory allows for shorter traces and lower power, which, depending on the configuration of the memory controller, can allow for increased DRAM clocks.
TheinsanegamerNDell offers the pro plus lineup with both intel lunar lake CPUs and AMD Hx series CPUs. Lunar lake has that super expensive on package memory. Guess what? With the same screen, memory, battery, ece, the only difference being the CPU, the highest end lunar lake chip, the Ultra 7 268V is $100 CHEAPER then the HX 370 system it competes with.

Intel themselves have said that MoP is a "a one-off with Lunar Lake" due to its impact on Intel's margins and how expensive it is for the company. Macro conditions can change how much Intel is actually charging the OEM for the completed SoCs, but the cost to Intel is greater and the margin is lower than its other chips (which is impressive given how poor Intel 7's cost structure is- the main driver of their client volume through 2025). Thus, the individual laptop being marginally cheaper is not a good representation of how much it costs Intel to produce the chip.
Posted on Reply
#23
DavidC1
Intel laptops can also be cheaper because they have higher marketshare, thus more volume discounts for every other component on the laptop.
Posted on Reply
Add your own comment
Nov 18th, 2026 09:03 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts