Computex 2026: Intel launches Crescent Island GPU with up to 480GB VRAM

Posted by ANR2ME@reddit | LocalLLaMA | View on Reddit | 24 comments

https://www.neowin.net/news/computex-2026-intel-launches-crescent-island-gpu-with-up-to-480gb-vram/

>Crescent Island is based on the company"s Arc Xe 3P architecture which lies inside current Panther Lake iGPs as well. This is Intel"s latest, most powerful card and it packs up to 480 GB of VRAM capacity. Unlike typical high-end professional GPUs which rely on HBM for improving power efficiency, the Intel GPU here has LPDDR5X.

>Cooling on the unit is handled by air cooler that can handle a TDP of 350 watts. Intel says that these cards can deal with next generation AI workloads and come with support for a wide range of datatypes and microscaling formats, from native FP4/MXFP4 to FP64, and more.

[-]

JaredsBored@reddit

I don't really get this. Lots of memory but it's going to be slow. Will they even have more bandwidth than a current gen 12 channel epyc system? Seems dubious

[-]

brown2green@reddit

It seems it will have a 1280-bit bus width and 1.5 TB/s total memory bandwidth.

https://videocardz.com/newz/intel-crescent-island-gpu-to-support-lpddr5x-9600-memory-and-1-5-tb-s-bandwidth

[-]

EbbNorth7735@reddit

That's plenty

[-]

ANR2ME@reddit (OP)

Based on this article https://www.tomshardware.com/pc-components/gpus/intel-details-long-awaited-crescent-island-ai-gpu-at-computex-boasts-up-to-480-gb-of-lpddr5x-to-combat-memory-shortages-company-shares-more-details-of-its-xe3p-inference-accelerator-at-computex

With 10.7 Gbps LPDDR5X, Crescent Island would offer 684 GB/s of memory bandwidth.

[-]

a_beautiful_rhind@reddit

Which is decent for sure.. but the price won't be and then what about software support.

[-]

ANR2ME@reddit (OP)

From what i read, they promised to have 0-day software support. So it can be used on the day it released.

[-]

Thin_Pollution8843@reddit

480GB with 34gbs bandwidth

[-]

ANR2ME@reddit (OP)

According to https://www.tomshardware.com/pc-components/gpus/intel-details-long-awaited-crescent-island-ai-gpu-at-computex-boasts-up-to-480-gb-of-lpddr5x-to-combat-memory-shortages-company-shares-more-details-of-its-xe3p-inference-accelerator-at-computex

With 10.7 Gbps LPDDR5X, Crescent Island would offer 684 GB/s of memory bandwidth.

[-]

Few_Painter_5588@reddit

I think the bigger news here is that it can be air cooled and has a TDP of 350 watts. A node of 4 could handle one instance of deepseek V4 Pro. For comparison, one B200 requires 1200ish watts and has 192GB of VRAM, so a cluster of 8 would barely run Deepseek V4 Pro.

[-]

LagOps91@reddit

thx, i'd rather buy a house or two.

[-]

sn2006gy@reddit

Is this the first hw native support for MXFP4?

[-]

ANR2ME@reddit (OP)

I think so, for Intel GPU that is. 🤔 Nvidia’s Blackwell and AMD CDNA 4 support it too isn't?

[-]

JaredsBored@reddit

Not Intel's first. Gaudi 3, which is an H200 competitor, has native MXFP4 support.

Source https://www.tomshardware.com/pc-components/cpus/intel-details-guadi-3-at-vision-2024-new-ai-accelerator-sampling-to-partners-now-volume-production-in-q3

[-]

ANR2ME@reddit (OP)

Hmm...🤔

Intel also announced new branding for its data center CPU portfolio, with the Granite Rapids and Sierra Forest chips now branded as the new ‘Xeon 6’ family. Those chips are on track to come to market this year and add support for the new performance-boosting standardized MXFP4 data format, which you can read about here.

And at https://www.tomshardware.com/pc-components/cpus/intel-unveils-new-xeon-6-branding-for-granite-rapids-and-sierra-forest-processors-efficiency-core-models-launch-this-quarter-performance-core-models-come-soon-after

Intel also provided an update on the launch schedule, indicating the Sierra Forest and Granite Rapids chips remain on track. The company also briefly mentioned new software support for the MXFP4 data format

Software support for the MXFP4. That is not native hw support isn't 🤔

[-]

JaredsBored@reddit

I should've found a better source, this one says it more clearly:

The Gaudi 3 AI chip comes with 128 GB of HBM2E with 3.7 TB/s of bandwidth and 24 200 Gbps Ethernet NICs, with dual 400 Gbps NICs used for scale-out. All of that is laid out on 10 tiles that make up the Gaudi 3 accelerator, which you can see pictured below. There is 96 MB of SRAM split between two compute tiles, which acts as a low-level cache that bridges data communication between Tensor Cores and HBM memory. Intel also announced support for the new performance-boosting standardized MXFP4 data format and is developing an AI NIC ASIC for Ultra Ethernet Consortium-compliant networking.

https://www.techpowerup.com/321342/intel-launches-gaudi-3-ai-accelerator-70-faster-training-50-faster-inference-compared-to-nvidia-h100-promises-better-efficiency-too

[-]

ANR2ME@reddit (OP)

It didn't mentioned hardware/native support 🤔 just support.

Meanwhile, At https://community.intel.com/t5/Blogs/Tech-Innovation/Artificial-Intelligence-AI/Deploying-Llama-4-Scout-and-Maverick-Models-on-Intel-Gaudi-3/post/1699547 The native data types supported on Intel Gaudi 3 Architecture are: - Matrix Multiplication Engines (MMEs) Precision support: Native FP8 (both E4M3 and E5M2 formats), BF16, FP16, and FP32 - Tensor Processing Cores (TPCs) Precision support: FP32, BF16, INT32, INT16, INT8 and other typical formats

[-]

JaredsBored@reddit

It's incredible how far downhill Intel's documentation has gone. Intel Ark doesn't clarify it, the white paper doesn't clarify it, the media announcements make it sound like it's supported, but an Intel forum post is the definitive source that it's software emulated. Ffs Intel.

[-]

ANR2ME@reddit (OP)

You will need 20 to get 480GB VRAM 😅

[-]

Borkato@reddit

:p maybe one day!