Analyst: Intel's discrete GPU business is losing money, and the entire AXG division may be cut off
Aug 13,2022
Jon Peddie, who was a guest speaker at the 6th Micro Semiconductor Summit Analyst Conference, recently wrote an analysis of Intel's GPU business. He believes Intel has invested about $3.5 billion in its discrete GPU development -- investments that have yet to pay off.
Intel's AXG (Accelerated Computing and Graphics division) has lost $2.1 billion since its official launch in the first quarter of 2021. Given Intel CEO Pat Gelsinger's track record of canceling six businesses since early 2021, JPR thinks AXG could be the next business to be canceled.
Peddie wrote: "Pat Gelsinger isn't afraid to make tough decisions, and if key projects don't pan out, he'll give up, even if it's a project he might personally like, and there have been rumors that AXG will be the next division to be jettisoned. "
When Intel made public its plans to develop discrete graphics solutions in 2017, it announced plans to leverage its GPUs to address functions such as computing, graphics and machine intelligence for client and data center applications. The Core and Visual Computing business unit is an additional supporting enabler to address the emerging edge computing market.
In its five years of discrete GPU development, the company has released two low-end discrete GPUs for processing inexpensive PCs and some data center applications; introduced a low-power graphics architecture for integrated GPUs; delivered oneAPI, which can be used to program CPUs, GPUs, FPGAs, and other computing units; canceled Xe-HP GPU architecture for data center GPUs; delayed shipments of Ponte Vecchio compute GPUs for AI and high-performance computing applications multiple times (recently Once due to Intel 4 node delays) and delayed the release of the Xe HPG ACM-G11 gaming GPU by about a year.
Jon Peddie believes Intel's Arc Alchemist 500 and 700-series GPUs are likely to fail, given their late arrivals and the fact that they will have to compete with AMD and Nvidia's next-gen Radeon RX 7000 and GeForce RTX 40-series offerings. It will obviously increase Intel's losses.
Given Intel's AXG track record, the company has spent $3.5 billion so far without any substantial success, Jon Peddie said. Discrete GPUs are an entirely new market for Intel and require a lot of investment, so it's not surprising that there are losses. Meanwhile, Intel's own Habana Gaudi2 deep-learning processor shows a clearer performance advantage than the Nvidia A100 in AI. "It's only half-probable whether Intel will stop and get out, and if they don't, the company will face years of losses as it tries to enter an unfriendly, tough market," Peddie said.
While it may be reasonable for Intel to cancel the AXG division and forgo discrete GPU development to cut losses, it should be noted that Intel's AXG division is pursuing several strategically important directions in general, particularly for discrete GPUs. development.
Peddie argues that discrete GPU development itself has really cost Intel so far, but it should be noted that without a competing GPU-like architecture that serves everything from low-end notebooks to supercomputers, Intel won't be able to solve it Many new growth opportunities.
Habana Gaudi2 looks like a competitive deep learning solution, but it cannot be used for supercomputing applications. Furthermore, without further improvements to the Intel Xe-HPC data center GPU architecture, the company will not be able to build hybrid processing units for AI/ML and HPC applications. Without such an XPU, Intel's 2027 plan of 1021 floating-point operations per second will become increasingly unrealistic.
While Intel's discrete GPU efforts fell short of expectations, Intel needed a well-defined parallel computing architecture to handle upcoming application loads. If Intel stops discrete GPU development, it will have to completely redesign its roadmap in terms of products and architecture. For example, it had to find a competitive GPU architecture supplier for its client processors, because Intel's small iGPU development team inside would have a hard time competing with solutions from AMD and Apple for its client SoCs.
The elimination of the AXG division appears to be an increasingly attractive management decision. However, GPUs and derivative hybrid architectures are strategically important to many of the markets Intel serves and applications to serve in the coming years, so eliminating the AXG division seems counterproductive.
Peddie wrote: "Pat Gelsinger isn't afraid to make tough decisions, and if key projects don't pan out, he'll give up, even if it's a project he might personally like, and there have been rumors that AXG will be the next division to be jettisoned. "
When Intel made public its plans to develop discrete graphics solutions in 2017, it announced plans to leverage its GPUs to address functions such as computing, graphics and machine intelligence for client and data center applications. The Core and Visual Computing business unit is an additional supporting enabler to address the emerging edge computing market.
In its five years of discrete GPU development, the company has released two low-end discrete GPUs for processing inexpensive PCs and some data center applications; introduced a low-power graphics architecture for integrated GPUs; delivered oneAPI, which can be used to program CPUs, GPUs, FPGAs, and other computing units; canceled Xe-HP GPU architecture for data center GPUs; delayed shipments of Ponte Vecchio compute GPUs for AI and high-performance computing applications multiple times (recently Once due to Intel 4 node delays) and delayed the release of the Xe HPG ACM-G11 gaming GPU by about a year.
Jon Peddie believes Intel's Arc Alchemist 500 and 700-series GPUs are likely to fail, given their late arrivals and the fact that they will have to compete with AMD and Nvidia's next-gen Radeon RX 7000 and GeForce RTX 40-series offerings. It will obviously increase Intel's losses.
Given Intel's AXG track record, the company has spent $3.5 billion so far without any substantial success, Jon Peddie said. Discrete GPUs are an entirely new market for Intel and require a lot of investment, so it's not surprising that there are losses. Meanwhile, Intel's own Habana Gaudi2 deep-learning processor shows a clearer performance advantage than the Nvidia A100 in AI. "It's only half-probable whether Intel will stop and get out, and if they don't, the company will face years of losses as it tries to enter an unfriendly, tough market," Peddie said.
While it may be reasonable for Intel to cancel the AXG division and forgo discrete GPU development to cut losses, it should be noted that Intel's AXG division is pursuing several strategically important directions in general, particularly for discrete GPUs. development.
Peddie argues that discrete GPU development itself has really cost Intel so far, but it should be noted that without a competing GPU-like architecture that serves everything from low-end notebooks to supercomputers, Intel won't be able to solve it Many new growth opportunities.
Habana Gaudi2 looks like a competitive deep learning solution, but it cannot be used for supercomputing applications. Furthermore, without further improvements to the Intel Xe-HPC data center GPU architecture, the company will not be able to build hybrid processing units for AI/ML and HPC applications. Without such an XPU, Intel's 2027 plan of 1021 floating-point operations per second will become increasingly unrealistic.
While Intel's discrete GPU efforts fell short of expectations, Intel needed a well-defined parallel computing architecture to handle upcoming application loads. If Intel stops discrete GPU development, it will have to completely redesign its roadmap in terms of products and architecture. For example, it had to find a competitive GPU architecture supplier for its client processors, because Intel's small iGPU development team inside would have a hard time competing with solutions from AMD and Apple for its client SoCs.
The elimination of the AXG division appears to be an increasingly attractive management decision. However, GPUs and derivative hybrid architectures are strategically important to many of the markets Intel serves and applications to serve in the coming years, so eliminating the AXG division seems counterproductive.