X-Silicon was x’d before X
GPU start-up brings MIMD to graphics and compute.
Dr. Jon Peddie
Jon Peddie Research | www.jonpeddie.com | August 23, 2023
X-Silicon, formed by ex-Silicon Valley experts, aims to revolutionize GPU design with a RISC-V vector-based unified graphics-compute engine (C-GPU) capable of AI, HPC, and 2D/3D graphics tasks. Their MIMD architecture enables independent CPU and GPU code execution within a single core, lowering memory usage and increasing performance. The company’s multicore layout features fast compositor fabric, enhancing data aggregation for diverse applications. X-Silicon seeks to reduce GPU latency through near-memory compute and novel hardware setups, backed by 14 patents. Targeting AR/VR, embedded devices, autos, and more, X-Silicon supports standard APIs and open programming for rapid development. Initial IP sales are aimed at OEMs and hyperscalers.
What do we think? The market is ripe for some new architectures. The big three have squeezed about all they can from the conventional SIMD and have been adding additional cores like matrix math (sometimes called tensor cores), ray-tracing interference testing engines, codecs, and audio processors. X-Silicon thinks it can satisfy many, if not all, of those specialized processor needs with a tightly-coupled tiles-communicating MIMD architecture. To do that, they are going to draw on the RISC-V ISA.
X-Silicon - a new company formed by former Silicon Valley veterans
Almost a year before Elon Musk made his company naming easy for people who can’t write or spell, a group of engineers, architects, programmers, and biz dev folks from AMD, Qualcomm, Intel, ATI Technologies, Dell, and others were working on a new GPU design and realized they should name their company. Not being newbies, they knew start-ups waste a lot of time trying to get a clever, unforgettable name and logo. So X was chosen. It could be experimental, sexy, or unknown—hopefully all would be true.
In order to reinvent the GPU shader core, X-Silicon says it is creating a new scalable RISC-V vector-based unified compute-graphics engine (C-GPU) that can efficiently compute the kinds of next-generation workloads that traditional GPUs were never designed for.
Such applications include AI, HPC, vision, geometric computing, as well as 2D and 3D graphics. The company says its MIMD architecture is uniquely capable of independently running CPU and GPU code in the same core affording capabilities such as low-memory footprint execution, bare-metal programming to the hardware registers, high-performance, low-power operation, and replacement of traditional shader programs with an open-source RISC-V ISA for CPU and GPU, using a single instruction stream.
X-Silicon’s single-core architecture
(Source: X-Silicon)
In the company’s multi-core design, multiple C-GPU cores are tiled across a chip and connected using an on-chip fast compositor fabric that can dynamically aggregate outputs from each core into a common buffer—i.e., frame buffer for graphics use cases or pipelined buffers for codec, video effects processing, and AI processing, as illustrated in the following diagram.
X-Silicon’s shader organization (Source: X-Silicon)
As such, workloads can be implemented in parallel or pipelined, and run simultaneously on a core as opposed to sequentially on traditional GPUs. It can also run an operating system on a core, says X-Silicon.
The company claims it can also reduce latency inherent in GPUs by doing near-memory compute, unified memory architectures, and other novel hardware configurations to accelerate compute. They have filed 14 patents for this.
Eric Powers, animation industry expert, former Disney/Applied Minds/Giant AI, commented, “For decades, high-end animation and effects have been unable to switch to GPUs for producing final images. The enormous complexity and scale of specialized rendering pipeline software, combined with the significant cost of crossing memory and platform barriers due to dedicated GPU designs, simply prevents mass adoption of GPUs in our most advanced technologies. An integrated design that gives HPC developers in-situ access to direct hardware acceleration like X-Silicon’s C-GPU architecture is the only future that carries us across that boundary.”
X-Silicon’s C-GPU allows market opportunities for new and emerging use cases for OEMs that wish to control their computing and GPU destiny. It allows new APIs, including custom APIs, and an ecosystem that is customized for an application, rather than bootstrapping an API to do a task it was never intended to do. It no longer requires OEMs and manufacturers to be subservient to the big five providers that are entrenched in a stagnant ecosystem. The company sees its market opportunities being in new and emerging markets such as entertainment and enterprise AR/VR with longer battery life, smart embedded devices with display needs, low-cost automotive displays and modules requiring predictable dedicated processing, wearables, custom animation processing, and others.
The company plans to support direct hardware and pixel access, so for low-memory applications, a bulky driver is not needed. X-Silicon’s approach towards open standards as well as its own open programming model will, says the company, facilitate quick and easy development for new use cases as well as improvements for existing products. Of course, X-Silicon also plans to support the gamut of APIs in the traditional software ecosystem, including OpenGL ES, Vulkan, Mesa, and OpenCL, but the company will also provide a hardware abstraction layer (HAL) that will allow direct access for others to optimize open source or create their own drivers and custom APIs. This is particularly interesting since the architecture supports new and emerging technology including new rendering models not supported by traditional architectures such as neural radiance fields (NeRFs) and non-triangular primitives.
X-Silicon’s single-core concept
(Source: X-Silicon)
The company plans to sell IP initially to OEMs, hyperscalers, and other system integrators. Due dates were not given for first silicon.
X-Silicon’s integration of a CPU and GPU (C-GPU) with a single ISA and an open Graphics Operating System (GOS) platform can have a profound impact on the overall software development, support, and maintenance of next-generation graphical rendering. This can enable an exciting era of innovation in the future of graphics, providing both emerging and legacy market segments an approach for new graphics algorithms, performance, power, flexibility, and cost. This approach turns the graphics world on its head, unlocking OEMs from the traditional GPU vendors offering black box drivers with complex APIs and costly legacy support.
Legacy GPU SIMD architectures, slaved to a host CPU and operating system and graphics services, limit innovation and help preserve an incumbent’s control of their markets. New, emerging, and smaller vertical markets, often underserved by these traditional graphics vendors, can develop and support compelling graphical solutions that can be upgraded and maintained for the life of the product. Next-generation products using new display technologies, new formats, and use paradigms (VR/AR, 360, depth, stereo, multiplane holography) often need a new approach for rendering. New development and deployment paradigms also require a consistent architecture, such as this, that is scalable from the edge to the cloud. Much more than an open-source alternative to traditional GPU vendors, X-Silicon is poised to provide a new technological graphical processing framework imbued with the latest AI and beyond triangle-based rendering to provide a platform for innovation not seen since the introduction of 3D graphics on mobile devices at the beginning of this century.
Dr. Jon Peddie | www.jonpeddie.com