Intel’s Xe-HP compute GPUs were the company’s first significant-efficiency discrete GPUs in years to electrical power on and were the initial standalone Xe GPUs that Intel shown to the community. But these GPUs will by no means make it to the market as the enterprise made the decision to abandon this lineup. Rather, in accordance to a tweet from Raja Koduri, senior vice president and basic supervisor of the Accelerated Computing Units and Graphics (AXG) Group at Intel, the business will concentrate on its Xe-HPG and Xe-HPC GPU people.
We deployed Xe HP in our oneAPI devcloud and leveraged it as a SW enhancement auto for oneAPI and Aurora. We at this time do not intend to productize Xe HP commercially, it evolved into HPG and HPC that are on normal industry manufacturing routeOct 28, 2021
The Xe-HP Family Axed
“We deployed Xe-HP in our oneAPI devcloud and leveraged it as a software development car for oneAPI and Aurora,” explained Koduri. “We at this time will not intend to productize Xe-HP commercially, it progressed into HPG and HPC that are on normal sector creation path.
When Intel in-depth its Xe spouse and children of GPU architectures in August 2020, it unveiled 4 architectures that had been intended to electricity four item services: the Xe-LP for built-in and lower-finish discrete GPUs, the Xe-HPG for superior-efficiency gaming graphics playing cards, the Xe-HP for professional data centre GPUs, and the Xe-HPC for supercomputers.
Xe-HP GPUs have been intended to scale from 1 GPU tile to 4 GPU tiles and were supposed to electrical power all the things from cloud gaming to online video streaming platforms and from virtual desktop infrastructure (VDI) to commercial accelerated computing.
To exhibit off the likely of its Xe architecture in substantial-functionality implementations, Intel demonstrated a quad-tile Xe-HP GPU in August 2020. That early Xe-HP GPU outperformed Nvidia’s A100 compute GPU by double and offered performance of all-around 42 FP32 TFLOPS. The silicon was built internally using the company’s 10nm Improved SuperFin (now termed Intel 7) fabrication procedure, the exact node that is utilized to make Intel’s Alder Lake and impending Sapphire Rapids processors.
Intel’s Xe-HP architecture relies on specialized execution models (EUs), significant-bandwidth inside interconnects, and an HBM2E-primarily based memory subsystem considering the fact that these GPUs are (or fairly had been) built to course of action vast amounts of information. These EUs assistance info formats and instructions frequent for details facilities and high-general performance computing (HPC), which contains FP64 for technological computing, bfloat16 for AI/ML workloads, DP4A convolution instruction for DL, and Intel’s new XMX guidelines. In addition, Xe-HP has innovative media encoding/decoding capabilities to handle video clip streaming companies.
Missed Window of Option?
Specified the fact that Intel obtained its very first Xe-HP GPUs from the fab in April 2020 and showcased them publicly in August, it most likely planned to commercialize them by now. The company made its Xe-HP GPUs available to application designers again in November 2020 and before this year it shipped samples of these compute GPUs to builders.
Really normally, the corporation did not disclose why it decided not to roll out its Xe-HP commercially, but we can speculate each about technological and strategical causes.
Configurations of Intel’s Arctic Audio 2T samples for builders only included playing cards with up to 960 EUs about two tiles that were rated for 300W, which did not appear to be like the greatest combination of general performance and electrical power use.
Whilst the FP32 functionality of Intel’s quad-tile Xe-HP appears to be like extraordinary currently, we do not know anything about its efficiency in other workloads and how it compares to the efficiency of current components. Meanwhile, this merchandise need to be aggressive for at least a few of decades and this is a thing that can’t be assured offered the truth that AMD is about to roll out its next-technology compute GPUs in November, whilst Nvidia is rumored to release its upcoming-gen compute GPUs in 2022.
Timing may perhaps not be Intel’s only problem with its Xe-HP GPUs. Back in the day, Intel tackled movie streaming platforms like YouTube with Xeon processors or even special Visual Compute Accelerators (dependent on Xeons). Google’s YouTube is changing Intel’s Xeon-based mostly equipment with servers working its homegrown visible (trans)coding models (VCUs), which could significantly shrink the addressable sector for Intel’s Xe-HP GPUs meant to change Xeons and VCAs. If other important movie streaming platforms abide by the match, the media transcoding capabilities of Xe-HP GPUs will hardly be appreciated. Meanwhile, media transcoding has always been a incredibly important Xe-HPC component.
Or a Rational Final decision?
On the other hand, Intel has its Xe-HPC-based mostly Ponte Vecchio ‘PetaFLOPS-course AI’ compute GPU that seems to be performing better than anticipated. Becoming aimed at supercomputers, it are unable to actually deal with all styles of datacenter workloads, but it can handle loads of them, including technical computing that requires double precision as effectively as AI/ML. Even so, the caveat is that Ponte Vecchio is expensive to make. Nevertheless, assuming that Intel can ship ample Ponte Vecchio compute GPUs to interested get-togethers, launching a decreased-expense rival for its individual product is not rational.
Intel is also about to introduce its Xe-HPG GPUs optimized for gaming. These parts can deal with desktops and laptops, but they can also be made use of to electricity cloud gaming companies and presented that they assistance GPU visualization (skill to provide far more than a single client), they can be even better than Xe-HP. To provide avid gamers, Intel requires to provide a complete software program stack that is continuously up-to-date to present higher efficiency in the hottest game titles. Supporting a independent stack for Xe-HPG and Xe-HP GPUs is a hard and highly-priced task. Furthermore, Xe-HP GPUs are a lot more advanced than Xe-HPG GPUs considering that they have to support all the aforementioned HPC details formats and their mix of overall performance and energy consumption may not be much too appealing for cloud gaming expert services.
For the exact same purpose, Xe-HP might not be much too captivating to digital desktop apps that need lower energy and reasonably reduced price tag.
But while Intel may have excellent causes to ax Xe-HP spouse and children for now, it is however a bit disappointing to see this happening. Continue to, as Intel’s discrete GPU initiative evolves, we may perhaps well see Intel’s GPUs built exclusively for information centers.