When Xindong/Innosilicon launched its Fenghua 1 (Fantasy 1) discrete graphics processing device (GPU) in mid-November, the organization unveiled a alternatively spectacular set of features for the GPU. This week, Innosilicon posted efficiency requirements for the chip. Though the Fenghua 1 can’t contend in opposition to the quickest offerings in our greatest graphics playing cards listing, and probably lands very a way down from the best of the GPU benchmarks hierarchy, it aims to present equivalent overall performance modern mid-vary GPUs from AMD and Nvidia.
The Xindong/Innosilicon’s Fenghua 1 (Fantasy 1) GPU is reportedly based on Creativity Technologies’ PowerVR architecture. Although we can only speculate about the correct microarchitecture utilized, the graphics processors are fairly able as they aid contemporary application programming interfaces for graphics and compute, which include DirectX, Vulkan, OpenGL, OpenCL, OpenGL ES, Caffe 1., TensorFlow 1.1.2, and ONNX. Since this is a PowerVR-centered GPU, it will come with Android, Linux, and Windows software stacks. The GPU is produced making use of a 12nm fabrication method.
There are two Fantasy 1 graphics cards at current: the single-chip Style A and the twin-chip Style B board. The one-chip Fantasy 1 boasts compute overall performance of around 5 FP32 TFLOPS for graphics and 25 INT8 TOPS for AI/ML, which is similar to the (theoretical) efficiency of Nvidia’s GeForce RTX 2060 GPU. Meanwhile, the dual-chip Fantasy 1 doubles that with FP32 compute general performance of about 10 FP32 TFLOPS and 50 INT8 TOPS for AI/ML. This is somewhat larger in comparison to functionality quantities available by Nvidia’s GeForce RTX 3060.
Take note that these figures stand for theoretical compute performance, and as we’ve seen from AMD and Nvidia in the previous, scaling by using twin-GPU styles can double compute whilst building a host of other hurdles when it comes to genuine-time graphics processing. In addition to driver help, twin GPUs typically have to comprise two copies of almost everything in memory — 1 duplicate for every single GPU and its direct attached VRAM. Syncing knowledge for frames in between the GPUs even further complicates points, normally requiring activity-certain support.
Xindong/Innosilicon’s Fantasy 1 Graphics Playing cards
|Style A||Type B|
|Number of GPUs||1||2|
|FP32 Effectiveness||5 FP32 TFLOPS||10 FP32 TFLOPS|
|INT8 Efficiency||25 TOPS||50 TOPS|
|Pixel Level||160 GPixel/s||320 GPixel/s|
|Online video Decoding||4x4Kp60, 16x1080p60, 32x720p30||8x4Kp60, 32x1080p60, 64x720p30|
|Number of consumers||16 1080p consumers||32 1080p buyers|
The Fantasy 1 Variety A card can be geared up with 4GB, 8GB or 16GB of GDDR6 or GDDR6X memory, and the Fantasy 1 Form B card should really double that with assistance for up to 32GB of memory. Both graphics playing cards support a PCIe Gen4 host interface as effectively as DisplayPort 1.4, eDP 1.4, and HDMI 2.1 interfaces.
When the raw compute functionality of the Fantasy 1 Kind B seems spectacular, it likely will never specifically translate over to video games. The developer under no circumstances mentions multi-GPU systems akin to AMD’s CrossFire or Nvidia’s SLI — neither of which are properly supported on modern-day GPUs in contemporary games. It appears like the Fantasy 1 Style B is aimed primarily at datacenters. Trying to keep in intellect that the GPU totally supports GPU virtualization as properly as PCIe SR-IOV, GPU computing in datacenters and digital desktop infrastructure (VDI) are amid the purposes it was designed for.
Probably the most astonishing component about the Fantasy 1 GPU is power intake. The normal ability use of one Fantasy 1 Form A card in ‘a multi-channel cloud environment’ is supposedly about 50W, which is noticeably reduce than the TDP of Nvidia’s GeForce RTX 2060 (TU106). Clearly, Xindong/Innosilicon’s promises have to be independently examined, but they do seem impressive.
Xindong/Innosilicon is presently sampling its Fantasy 1 graphics cards with intrigued parties. However, it is unclear when these GPUs will be accessible commercially. Other than availability and effectiveness, we would also need pricing particulars to ascertain whether they can compete with the likes of AMD Radeon, Nvidia GeForce, and Intel Arc — or additional most likely, versus the datacenter variants of people GPUs.