Chinese GPU Maker, Innosilicon, Debuts Its First Domestically Produced Graphics Cards: First Demo, Up To 10 TFLOPs, 32 GB GDDR6X Memory

Hassan Mujtaba

Wccftech


Chinese GPU Maker, Innosilicon, Debuts Its First Domestically Produced Graphics Cards: First Demo, Up To 10 TFLOPs, 32 GB GDDR6X Memory

Chinese GPU Maker, Innosilicon, Debuts Its First Domestically Produced Graphics Cards: First Demo, Up To 10 TFLOPs, 32 GB GDDR6X Memory

Chinese GPU maker, Innosilicon, has officially debuted its domestically produced graphics cards based on the Fantasy One GPU. The GPU maker introduced four new cards that are designed in collaboration with the GPU designer, Imagination Technologies.

Chinese Domestic GPUs Make Official Debut, First Demo of Innosilicon's Fantasy One GPU Powered Graphics Card Showcased

While Innosilcon already announced its Fantasy One GPU and the respective graphics card lineup at a press conference last month, they have revealed more information on their website, giving a detailed look at the specifications of each variant. During a recent, event, the company demoed its Fantasy One GPUs in GFXBench too which can be seen below:

Innosilicon graphics card. pic.twitter.com/lyMpQAo5Km

— Stewie (@Stewrandall) December 23, 2021

No details or performance measurements for the graphics cards are mentioned but from the quality of the demo, it looks very dated and isn't as impressive of a showcase considering what modern GPUs are capable of but it's something for the domestic Chinese GPU market. It looks like Chinese manufacturers are still lagging a couple of years behind AMD and NVIDIA.

InnoSilicon Fantasy One Type A Graphics Card

The first product to talk about is the Fantasy One GPU-powered Type A graphics card. Type A is the name of the graphics card and is not to be mistaken with USB Type-A ports. The Type A card comes in a dual-slot and dual-fan form factor and is said to offer up to 5 TFLOPs of FP32 and 160 GPixel/s fill rate. It seems like that this is a slightly cut-down variant of the Fantasy One GPU as the BXT GPU IP that it is based upon offers up to 6 TFLOPs of FP32, 24 TOPs AI and 192 GPixel/s fill rate. The GPU also relies on a multi-chip design so we are talking chiplets here but we don't have more details on that right now.

Fenghua No. 1 is a 15-year masterpiece of Inspur Technology. It fills the two major gaps in domestic 4K-level desktop graphics and server-level graphics, and supports domestic new infrastructure 5G data centers, Xinchuang Desktop, Meta Universe, cloud games, cloud desktops, Multi-billion-level industries such as graphics workstations are the beginning of SingDong's empowerment of the domestic GPU ecological chain.

via InnoSilicon

In terms of memory, the single Fantasy One GPU offers up to 16 GB of GDDR6X memory that runs across a 128-bit bus interface. There are also 4 GB & 8 GB variants of the same card. The memory chips operate at 19 Gbps transfer speeds, offering up to 304 GB/s of bandwidth to the GPU. In addition to that, the card offers DisplayPort 1.4, HDMI 2.1, and VGA outputs. The card is powered by a single 6-pin connector. It is stated that the typical TDP for 4K rendering is around 20W while cloud rendering power is rated at 50W. There's also a server-specific model of Type A which utilizes a passive heatsink.

Fenghua No. 1 Type A Card

  • FP32 floating point: 5T FLOPS

  • Rendering capability: 160G Pixel/s

  • Codec capability: Simultaneously 4 channels of 4K60 frames, 16 channels of 1080P60, and 32 channels of 720P30

  • Number of users: 16 1080P users, 32 720P users

  • AI calculation: 25TOPS (INT8)

  • innosilicon-fantasy-one-gpu-discrete-graphics-cards-type-a-_5

  • innosilicon-fantasy-one-gpu-discrete-graphics-cards-type-a-_6

InnoSilicon Fantasy One Type A SFF Graphics Card

There is also an entry-level/small form factor solution based on the Type A Fantasy One GPU. This graphics card definitely rocks reduced specifications and clocks over the full variant considering its limited size and power. The graphics card PCB shown has a 4+1 phase design and rocks a small fan heatsink. It features HDMI 2.1 and DisplayPort 1.4 outputs. There are no power connectors on the card and it looks like you will also get lower memory capacity given the size constraints of the PCB.

  • innosilicon-fantasy-one-gpu-discrete-graphics-cards-type-a-_3
  • innosilicon-fantasy-one-gpu-discrete-graphics-cards-type-a-_4

InnoSilicon Fantasy One Type B Graphics Card

On the high-end side, there's the Innosilicon Type-B graphics card featuring dual Fantasy One GPUs on the same PCB. The card uses an interconnect known as Innolink which might be a PLX chip that interconnects the two GPUs on the same board. The graphics card offers up to 10 TFLOPs FP32, 50 TOPs INT8 and up to 320 GPixel/s fill rate. It can also run up to 32 1080p/60 FPS or 64 720/30 FPS streams. Type-B rocks 32 GB of GDDR6X memory across a 128-bit bus interface. Power is provided through dual 8-pin connectors. All cards are PCIe Gen 4.0 compliant.

The Fantasy One Type A graphics cards are aimed at the consumer, workstation, and data center markets while the Type B graphics cards are aimed at the data center, cloud computing & AI use cases. These graphics cards will pack support for a range of APIs including DirectX & were demoed within Unigine Heaven, a fairly old benchmark for today's standards but shows that the intended DirectX technology does work on the Fantasy One silicon.

Fenghua No. 1 Type B Card

  • FP32 floating point: 10T FLOPS
  • Rendering capability: 320G Pixel/s
  • Codec capability: simultaneously 8 channels of 4K60 frames, 32 channels of 1080P60, 64 channels of 720P30
  • Number of users: 32 1080P users, 64 720P users
  • AI calculation: 50TOPS (INT8)

Innosilicon also isn't stopping at Fantasy One and already talking about Fantasy 2 and Fantasy 3 Chinese domestically produced GPUs which are said to be unveiled in 2022 and will utilize 5nm process technology.

The post Chinese GPU Maker, Innosilicon, Debuts Its First Domestically Produced Graphics Cards: First Demo, Up To 10 TFLOPs, 32 GB GDDR6X Memory by Hassan Mujtaba appeared first on Wccftech.

Continue Reading

Loading data