Chinese Vendor Designs PCIe 4.0 GPU, Targets GTX 1080 Performance
The high-performance GPU industry has been a two-horse race for very nearly two decades. After the collapse of 3dfx, no new company emerged to seriously challenge the ATI/Nvidia split. While Intel holds a substantial stake of the total GPU market, its integrated business has only focused on 2D, video, and basic gaming 3D. Intel’s upcoming Xe architecture, expected in 2020, will take a serious shot at breaking into the consumer space. Now, there’s a word of a potential fourth player in the field, albeit it possibly in a more specialized area.
According to THG, Jingjia Micro is a military-civilian integrated company that’s primarily focused on developing GPUs for the military market thus far. The company began by building China’s first homegrown GPU, the JM5400, built on 65nm. The success of the JM5400 allowed the company to expand and move to newer manufacturing nodes. Its next products, the JM7000 and 7200, were built on 28nm. Now, Jingjia Micro wants to expand its reach further and target the performance of the GTX 1050 and 1080 with a pair of new designs — the JM9231 and JM9271.
A post at cnbeta has additional information. Currently, the JM7200 is said to offer performance equivalent to the GeForce GT 640, albeit in a much lower power envelope — 10W, supposedly, compared with the 50W Nvidia specced for that card. We’d like to see that claim independently verified. The OEM variant of the GT 640 was a Fermi-based part built on 40nm, but that chip had a 65W TDP. The 50W variant was a Kepler-derived part built on 28nm — the same process node Jingjia Micro uses. The JM part also supposedly has 4GB of RAM, while the GT 640 50W version had just 1GB of GDDR5.
The JM9231 and JM9271 are supposedly the first fully programmable GPUs that Jingjia Micro has developed; there are references to the previous JM5400 and JM7200 families being based on fixed-function rendering pipelines. These limitations wouldn’t fly under modern APIs for Windows, but the company started life as a military GPU vendor, and such applications obviously have very different requirements for APIs and product certification.
The new JM parts obviously aren’t going to gun for the highest-end cards from Nvidia or AMD, but even approaching high-end performance from 2016 – 2017 would allow them to contend for the midrange and budget markets. Bringing up the software stack and winning developer support would obviously be critical to any market play, and there doesn’t seem to be any information about whether the JM9231 or JM9271 include any performance improvements or ideas that we haven’t seen before from the major vendors. Such events are rare, but not unheard of. PowerVR once attempted to establish itself as a third player in PC graphics with the Kyro and Kyro II, which won some market share for itself as a unique solution with higher memory bandwidth efficiency than either ATI or Nvidia.
The use of HBM memory in a product of this sort is rather interesting, as is the comparatively low memory bandwidth (by HBM standards). Given that both products lack modern API support, it’s possible they’re intended strictly for military use — though in that case, referencing the GTX 1080 would be a bit odd. Either way, China clearly has its eye on competing more aggressively in terms of overall silicon performance. A few more years, and we might see new products from vendors we haven’t seen before challenging ‘homegrown’ alternatives like AMD, Nvidia, and (if its Xe launch goes well), Intel.
- AMD, Nvidia High-End GPUs Are Much Better Deals Now Than 6 Months Ago
- No, AMD Hasn’t Quit Making Reference 5700 and 5700 XT GPUs
- How Are Process Nodes Defined?