Wednesday, May 20th 2026

Leaked Intel Crescent Island PCB Surfaces with Xe3P GPU, 160 GB LPDDR5X and 16-Pin Power

A PCB leak from YuuKi_AnS has given us the first look at Intel's Crescent Island PCIe accelerator and a few things stand out immediately. The Xe3P GPU die is visibly larger than the Xe2-based BMG-G31 in the current Intel Arc lineup. Xe3P is the architecture following the current Xe3. While Xe3 is heading to client products under the Arc C-Series branding, Xe3P is being positioned as a more scalable design that covers client iGPUs all the way up to data center inference. Memory is the other talking point confirming previous rumors by leaker Jaykihn from April. Rather than HBM, Intel is going with LPDDR5X, the PCB showing 20 memory sites in total, 12 on the front and 8 on the back, each carrying 8 GB for a total of up to 160 GB. This is a deliberate choice made by Intel to sidestep the HBM supply issues while keeping costs down, though LPDDR5X trades bandwidth for lower pricing compared to HBM standards.

The rest of the PCB shows 13 populated VRM stages out of 18 total positions, and power comes in via a single 16-pin connector on the back of the board. A USB Type-C port on the side appears to be for engineering and testing purposes rather than a production feature. Intel Crescent Island GPUs are still a few months out, but the leaked PCB looks close to a final design according to the source. However, no gaming variant is expected based on current rumors. The company recently introduced the Arc Pro B70 and B65 professional cards based on Battlemage Xe2 silicon, while rumors surrounding a higher-end Arc B770 gaming model have largely gone quiet ahead of Computex.
Sources: YuuKi_AnS on X, via Videocardz
Add your own comment

15 Comments on Leaked Intel Crescent Island PCB Surfaces with Xe3P GPU, 160 GB LPDDR5X and 16-Pin Power

#1
wNotyarD
Nomad76While Xe3 is heading to client products under the Arc C-Series branding
Is it really?
Posted on Reply
#2
TheDeeGee
Bit hard to see, but i think it's all 6 12V pins into a single plane.

aka

Set sail for fail.
Posted on Reply
#3
Mrgravia
Arc C series consumer cards? I'll believe it when i see it.

Still waiting for an A310/380 replacement to drop.
Posted on Reply
#4
Nomad76
News Editor
wNotyarDIs it really?
MrgraviaArc C series consumer cards? I'll believe it when i see it.

Still waiting for an A310/380 replacement to drop.
Since nobody's rushing with new consumer GPUs anytime soon... we might have a long wait ahead. As for the Xe3 on the Arc C series, that seems the most logical step, but how and when is another story.
Posted on Reply
#5
Six_Times
Impressive. The question is, when will it truly launch?
Posted on Reply
#6
igormp
Assuming all of those 20 slots are independent channels, each being 32-bit wide, we would be looking at a 640-bit wide device.
Depending on how they clock it, on the higher end we could see 8533~9600MT/s being used, which would mean 683~768GB/s of memory bandwidth.

Far from great, but with that capacity it would be a pretty nice device for MoE inferencing, as long as its priced correctly ofc.
Posted on Reply
#7
trsttte


This is not for any of us, fully data center oriented, which makes me wonder why did they bother with 12VHPWR 16pin connector, why not just go with a simple EPS connector?

160gb of LPDDR5X sure does sound fun
Posted on Reply
#8
yfn_ratchet
trsttteThis is not for any of us, fully data center oriented, which makes me wonder why did they bother with 12VHPWR 16pin connector, why not just go with a simple EPS connector?
Might still be targeted towards desktop usage, where ATX 3.x PSUs are already highly prevalent. Kind of like how Arc Pro Battlemage fits nicely into a desktop system. I imagine they only foresee you using one or two of these, unless you've got one of those real beefster full towers with two PSU bays in the bottom.
Posted on Reply
#9
Wirko
igormpAssuming all of those 20 slots are independent channels, each being 32-bit wide, we would be looking at a 640-bit wide device.
Depending on how they clock it, on the higher end we could see 8533~9600MT/s being used, which would mean 683~768GB/s of memory bandwidth.

Far from great, but with that capacity it would be a pretty nice device for MoE inferencing, as long as its priced correctly ofc.
Sure, the LPDDR on this card is not a replacement for HBM but rather for GDDR. Fast LPDDR6X might allow a lower system cost (PCB space, substrate, GPU and memory chip packaging) compared to GDDR at similar bandwidth. And Intel may know a bit more than us about future memory shortages.
Posted on Reply
#10
igormp
Extra sweet rumors:
videocardz.com/newz/intel-crescent-island-gpu-to-support-lpddr5x-9600-memory-and-1-5-tb-s-bandwidth

They are apparently going to use 64-bit modules, so 1280-bit in total with 9600MT/s, which doubles the estimate I had done before to 1.5TB/s of mem bw. This gets it close to a RTX PRO 6000's mem bw with almost double the memory, so it'd allow Intel to price it around $10k without much issues.

If higher density options become available, then 320~480GB total VRAM leaves it at a capacity higher than the current top of the line products from Nvidia and AMD, ofc at a way slower mem bw and compute perf, but hopefully at a fraction of the price as well. It'd be PERFECT for inference.
WirkoSure, the LPDDR on this card is not a replacement for HBM but rather for GDDR. Fast LPDDR6X might allow a lower system cost (PCB space, substrate, GPU and memory chip packaging) compared to GDDR at similar same bandwidth. And Intel may know a bit more than us about future memory shortages.
With the above rumor such setup becomes directly competitive with the likes of GDDR7, but at higher memory density.
I can't comment on cost since I don't know how much cheaper/expensive GDDR modules and controllers are when compared to LPDDR ones.
Posted on Reply
#11
Wirko
igormpThey are apparently going to use 64-bit modules, so 1280-bit in total with 9600MT/s, which doubles the estimate I had done before to 1.5TB/s of mem bw.
64 bits may be a wrong assumption - or it may be correct. A chip package contains many 32-bit LPDDR5X dice, so it's not a big stretch to make packages with a 64-bit data bus.
Posted on Reply
#12
igormp
Wirko64 bits may be a wrong assumption - or it may be correct. A chip package contains many 32-bit LPDDR5X dice, so it's not a big stretch to make packages with a 64-bit data bus.
64-bit packages are already a thing, micron has many LPDDR5X options ranging from 16Gb all the way up to 128Gb (used a lot in phones and even PC devices like LNL or the Mac Neo). The rumor points to Intel starting with 64Gb packages (8x20=160GB), but they should be able to easily double that soon after.
I'm not sure if 192/256Gb are already available, but those would be nice to see.

EDIT: 192Gb dies are available, the Strix Halo refresh is going to be using those to bump its capacity to 192GB:
www.servethehome.com/amd-reveals-ryzen-ai-max-pro-400-series-192gb-ram-for-ai-systems/
With the recent advent of 24GB(192Gbit) LPDDR5X memory chips, AMD is quickly taking advantage of the tech to boost the Ryzen AI Max family’s already spacious memory capacity to a new level, while boosting memory bandwidth by about 7% to 273GB/sec.
With a minor caveat that Strix Halo uses 32-bit modules instead of 64-bit ones, but 64-bit modules should not be hard to become a thing.
WirkoA chip package contains many 32-bit LPDDR5X dice
Aren't LPDDR5X dies usually 8 or 16-bit wide? I know that Micron does have some parts which apparently use 32-bit dies, but that doesn't seem to be the norm.
Anyhow, internal organization of a package is almost irrelevant, what matters most is the width of your individual package. Micron lists availability for 16/32/64-bit wide parts, I'm not sure if there are options outside of this range from other manufacturers.
Posted on Reply
#13
Mrgravia
Nomad76Since nobody's rushing with new consumer GPUs anytime soon... we might have a long wait ahead. As for the Xe3 on the Arc C series, that seems the most logical step, but how and when is another story.
I'd settle for a 8gb version of the Arc B50.

Just anything for the lower end, especially for lower end LP.
Posted on Reply
#14
Visible Noise
igormpExtra sweet rumors:
videocardz.com/newz/intel-crescent-island-gpu-to-support-lpddr5x-9600-memory-and-1-5-tb-s-bandwidth

They are apparently going to use 64-bit modules, so 1280-bit in total with 9600MT/s, which doubles the estimate I had done before to 1.5TB/s of mem bw. This gets it close to a RTX PRO 6000's mem bw with almost double the memory, so it'd allow Intel to price it around $10k without much issues.

If higher density options become available, then 320~480GB total VRAM leaves it at a capacity higher than the current top of the line products from Nvidia and AMD, ofc at a way slower mem bw and compute perf, but hopefully at a fraction of the price as well. It'd be PERFECT for inference.


With the above rumor such setup becomes directly competitive with the likes of GDDR7, but at higher memory density.
I can't comment on cost since I don't know how much cheaper/expensive GDDR modules and controllers are when compared to LPDDR ones.
And at a third of the latency of GDDR, half of HBM.
Posted on Reply
#15
igormp
Visible NoiseAnd at a third of the latency of GDDR, half of HBM.
Not like this matters much for its target market, but yeah, it's an improvement nonetheless.
Posted on Reply
Add your own comment
Nov 11th, 2026 04:11 CST change timezone

New Forum Posts

Popular Reviews

TPU on YouTube

Controversial News Posts