When Xindong/Innosilicon introduced its Fenghua 1 (Fantasy 1) discrete graphics processing unit (GPU) in mid-November, the company unveiled a rather impressive set of features for the GPU. This week, Innosilicon published performance specifications for the chip. While the Fenghua 1 cannot compete against the fastest offerings in our best graphics cards list, and likely lands quite a way down from the top of the GPU benchmarks hierarchy, it aims to provide comparable performance modern mid-range GPUs from AMD and Nvidia.
The Xindong/Innosilicon's Fenghua 1 (Fantasy 1) GPU is reportedly based on Imagination Technologies' PowerVR architecture. While we can only speculate about the exact microarchitecture used, the graphics processors are quite capable as they support contemporary application programming interfaces for graphics and compute, including DirectX, Vulkan, OpenGL, OpenCL, OpenGL ES, Caffe 1.0, TensorFlow 1.1.2, and ONNX. Since this is a PowerVR-based GPU, it comes with Android, Linux, and Windows software stacks. The GPU is made using a 12nm fabrication process.
There are two Fantasy 1 graphics cards at present: the single-chip Type A and the dual-chip Type B board. The single-chip Fantasy 1 boasts compute performance of around 5 FP32 TFLOPS for graphics and 25 INT8 TOPS for AI/ML, which is comparable to the (theoretical) performance of Nvidia's GeForce RTX 2060 GPU. Meanwhile, the dual-chip Fantasy 1 doubles that with FP32 compute performance of approximately 10 FP32 TFLOPS and 50 INT8 TOPS for AI/ML. This is slightly higher compared to performance numbers offered by Nvidia's GeForce RTX 3060.
Note that those figures represent theoretical compute performance, and as we've seen from AMD and Nvidia in the past, scaling via dual-GPU designs can double compute while creating a host of other hurdles when it comes to real-time graphics processing. Besides driver support, dual GPUs typically have to contain two copies of everything in memory — one copy for each GPU and its direct attached VRAM. Syncing data for frames between the GPUs further complicates things, often requiring game-specific support.
Xindong/Innosilicon's Fantasy 1 Graphics Cards
Row 0 - Cell 0 | Type A | Type B |
Number of GPUs | 1 | 2 |
FP32 Performance | 5 FP32 TFLOPS | 10 FP32 TFLOPS |
INT8 Performance | 25 TOPS | 50 TOPS |
Pixel Rate | 160 GPixel/s | 320 GPixel/s |
Video Decoding | 4x4Kp60, 16x1080p60, 32x720p30 | 8x4Kp60, 32x1080p60, 64x720p30 |
Number of users | 16 1080p users | 32 1080p users |
The Fantasy 1 Type A card can be equipped with 4GB, 8GB or 16GB of GDDR6 or GDDR6X memory, and the Fantasy 1 Type B card should double that with support for up to 32GB of memory. Both graphics cards support a PCIe Gen4 host interface as well as DisplayPort 1.4, eDP 1.4, and HDMI 2.1 interfaces.
While the raw compute performance of the Fantasy 1 Type B looks impressive, it likely won't directly translate over to games. The developer never mentions multi-GPU technologies akin to AMD's CrossFire or Nvidia's SLI — neither of which are properly supported on modern GPUs in contemporary games. It looks like the Fantasy 1 Type B is aimed mostly at datacenters. Keeping in mind that the GPU fully supports GPU virtualization as well as PCIe SR-IOV, GPU computing in datacenters and virtual desktop infrastructure (VDI) are among the applications it was designed for.
Perhaps the most surprising part about the Fantasy 1 GPU is power consumption. The typical power consumption of one Fantasy 1 Type A card in 'a multi-channel cloud environment' is supposedly about 50W, which is considerably lower than the TDP of Nvidia's GeForce RTX 2060 (TU106). Obviously, Xindong/Innosilicon's claims have to be independently tested, but they do sound impressive.
Xindong/Innosilicon is currently sampling its Fantasy 1 graphics cards with interested parties. However, it is unclear when these GPUs will be available commercially. Besides availability and performance, we would also need pricing details to determine whether they can compete with the likes of AMD Radeon, Nvidia GeForce, and Intel Arc — or more likely, against the datacenter variants of those GPUs.