2014 SoCs: A Win for Nvidia?
Let me preface this by saying that I have really disliked Nvidia SoCs and GPUs over the past few years. I have been an unabashed critic of the Tegra 2-4, and Nvidia's move to Kepler. However, after today's events, I really like Nvidia's new approach and I would like to outline the reasons why I think that Nvidia has won this generation, and maybe the next few!
Nvidia's past efforts with their Tegra SoCs have always seemed a day late and a dollar short. The Tegra 2 may have been the first dual core SoC, but its famous lack of decent video decoding and NEON extensions doomed it from a start.
The Tegra 3 famously trumpeted its 5 (4 + 1) cores and its new and improved GPU. It was using slow A9s and was outshone by even the old A5 and the more efficient Snapdragon Pro. Its GPU improved, but was still horrible inefficient for mobile. Its bandwidth was also severely limited and caused many issues, notably in the Transformer.
The Tegra 4 was better, and many a Nvidia fan heralded it as Nvidia's savior. It destroyed benchmarks and seemed more powerful than the Snapdragon 800 that was soon released. However, I saw the reference platform and I knew the issue: power consumption. Rightly so, Nvidia couldn't get a single phone win until late 2013, its Icera soft modem was just not up to snuff, and Qualcomm and Intel were splitting tablet SoCs. The only notable win for Nvidia in 2014, IMO, was MS Surface.
Recently announced was the K1, aka the Tegra 5. This really blew me out of the water, and was a complete 180 from the direction Nvidia was headed. Well, I need to slow down for a second. There are really two K1s. One is completely evolutionary, and the other is completely revolutionary.
The first SoC in the K1 family is merely a slightly changed T4. It is the same 4 x A15 + 1 LP formula, just at 25% higher frequencies and minor manufacturing improvements with HPM. Nvidia says that this will occupy the same power envelope as the T4, and on paper, it looks like it will while gaining about 30% in performance, however, it is still prohibitively high for smartphones and
Wow. That is all I can say to this CPU. This is the Project Denver we have all been waiting for. One year early! However, it is not x86, but it is instead ARM. You can read about this and the licensing issues in Demerjian's piece over at SemiAccurate. In short, this CPU has a special software layer that abstracts regular ARM code and recompiles for the specific architecture that Nvidia is using to gain major efficiency benefits. As a result, this CPU is all but two extremely fat cores. This is a complete rejection of Nvidia's moar cores approach. This is very much like Apple's A7 and Intel's more recent Atoms. I applaud Nvidia's new path, and the new CPU should be much faster and efficient.
Fast. Really fast. Think about ½ to ⅓ of a 740M. Think much faster than Playstation 3 and Xbox 360.
Here is a chart courtesy of Nvidia, and Anandtech (Sorry, editor is not allowing images ATM):
192 Cuda Cores! This has me worried though. What is the power consumption going to be like? IMO, they are going to have scale down further to meet sensible power limits. However, if they can succeed in efficiency, much like they had with Kepler on laptops, they could be golden. All the APIs like OpenGL are up to date, and other features like ASTC really shine.
To be brutally honest, Nvidia is going to have a very rough first two quarters. I predict that both HTC and Samsung will go with Qualcomm with either the 805, or a new processor. The first gen K1 is too little of an improvement over the T4 power consumption wise to be worth using. And that brings us to another point, where is Qualcomm's new tech? Is it only the 805? I doubt it, when the 410 is out and 64 bit. I think this brings me to my point; Qualcomm was caught with its pants down on a 64-bit Arm8-based chip. They were blindsided by Apple, and no one thought that Nvidia would have Denver cores ready.
I have no doubts, the 810 will be out this year, but the question is when and how good will it be. Qualcomm has great engineers, but if their original product cycle is one year out, they will be in a tough position to rebound. Especially in tablets, I believe that Qualcomm will be beat to design wins by Nvidia.
With this new architecture of Nvidia's, their most analagous competition from an architecture stand point would be Intel and Apple. In is good and bad in way where neither of them are competing in the same markets, Intel will most penetrate the Windows market, and Apple is Apple and the iPhone and iPad will sell regardless. Apple's A8 will be interesting because they don't have much more room to grow up. They have no new architecture, so their gains will be silicon-based. I believe that they will move to 25nm or 22nm, which will give them either performance leeway, or efficiency gains. Intel is much livelier competition, and the new Atom's will be tough to beat. However, Nvidia has the GPU edge with its new Kepler cores. Even Apple, usually the forefront of GPU tech, doesn't have much room to grow with the released PowerVR GX 6 series.
WHAT WORRIES ME:
- Discrete Icera Modems for the time being, prohibitive for smartphones of connected tablets
- Power Consumption, always seems to be Nvidia's weak spot.
- Software, will Denver be done right?
- Underlying architecture only seems to be A15-based, A57 could offer many benefits
- Ship Date!!!
The ball is in Nvidia's court. On paper, they have done everything right except integrating a modem and using an A57 base, but those can be nullified with a good discrete soft modem, and good software respectively. 2014 looks to be a bright year for Nvidia, and Android in general. I am looking forward for healthy SoC competition, and maybe a Nvidia powered Nexus device!