Samsung may be in talks with AMD, Nvidia to license graphics IP
Samsung and Apple are the two titans of the smartphone industry, but they’ve taken very different routes to get there. Where Apple has invested a great deal of time and money into building its own custom SoCs and designing its own microarchitecture, Samsung has historically used cores from many other companies, including Texas Instruments and Qualcomm. Last year, the company announced it would bring its first custom CPU core to market in 2016. Now, it’s reportedly in talks with both AMD and Nvidia, looking for a graphics core to match its custom CPU.
We’ve heard rumors of Samsung wanting a custom GPU core before, but nothing concrete has come of it thus far. It’s hard to find good data on recent smartphone GPU market share, but this graph by Jon Peddie Research shows the lay of the land from two years ago.
Given that Intel and Vivante are both out of the game (Vivante seems to have pivoted to focus on embedded and IoT), the major players here are Imagination Technologies (Apple’s preferred partner) and Qualcomm. Both Nvidia and AMD have graphics IP that Samsung could license — the key question is, is Samsung just looking to cut a deal that gives them access to certain patents so they can legally roll their own solution, or do they want a part that’s designed by either AMD or Nvidia?
It’s not actually clear which company would be the better fit. Nvidia manufacturers one of the best Android tablets you can buy, but recently canceled its plans to upgrade the Shield tablet. It’s been speculated that this might have been due to pressure from Nintendo and its upcoming NX. Nvidia is believed to have won that design and may have been prevented from bringing a competitor product to market. Over the last few years Nvidia has focused much more on automotive and deep learning than on launching mobile products.
AMD famously sold its mobile IP to Qualcomm, who created Adreno. There’s no reason Team Red couldn’t build a new low-power GPU, and the company’s focus on embedded and semicustom design makes it a good fit for Samsung. CEO Lisa Su has also talked about wanting to lean more on AMD’s patent portfolio while building out additional semi-custom wins. The question is whether AMD has the manpower or current resources to design something in-house. GCN isn’t known for being a low-power GPU architecture, after all, and AMD already has plenty to keep it busy with Microsoft’s Scorpio, Vega, and Zen all dropping in 2017.
Speaking of Samsung’s M1, Anandtech got a look at the architecture during Hot Chips 2016. The chip is an out-of-order design capable of decoding four instructions per cycle and with a perceptron-based branch predictor. This last isn’t a brand-new feature, but companies are often cagey about the exact details of how they implement branch prediction in silicon.
Above, we see Samsung’s M1 performance as compared with a Cortex-A57, albeit one clocked 200MHz below the Samsung chip. Performance is fairly good, even if the chip really only distinguishes itself in a few tests. As a first effort, it’s a solid core. If Samsung wants to pair it with a GPU of their own design, AMD and Nvidia are the companies to talk to. It could take several years to bring a new solution to market. But if Samsung wants Apple-like control over its own product lines, creating new GPU and CPU IP is the way to do it.
No comments:
Post a Comment