Qualcomm's Snapdragon 805: 2.5GHz, 128-bit Memory Interface, D3D11-class Graphics & More
by Anand Lal Shimpi on November 22, 2013 11:00 AM EST- Posted in
- SoCs
- Qualcomm
- Smartphones
- Mobile
- Tablets
- Snapdragon 800
- Snapdragon 805
Qualcomm has had an incredible year. It wasn’t too long ago that I was complaining about Qualcomm’s release cadence, the lull between Scorpion and Krait allowed competitors like NVIDIA, Samsung and TI to get a foothold in the market. Since the arrival of Krait, the move to 28nm and the launch of monolithic AP/LTE solutions, no competitor has been able to come close to touching Qualcomm. These days the choice of integrating mobile silicon really boils down to what Snapdragon variant an OEM wants to go with. TI is out of the business, NVIDIA hasn’t seen much traction with Tegra 4 and even Samsung will ship Qualcomm silicon in many of its important markets.
Qualcomm’s Snapdragon 600 was the SoC of choice at the beginning of the year, with Snapdragon 800 taking over that title more recently. Earlier this week, Qualcomm announced the successor to the 800: the Snapdragon 805.
We’re expecting to see devices based on the Snapdragon 805 to be shipping in the first half of 2014, so Snapdragon 800 will still enjoy some time at the top of the charts.
The 805 starts by integrating four Krait 450 cores. Krait 450 appears to be an evolutionary upgrade over Krait 400, with no changes to machine width, cache sizes or pipeline depth. Qualcomm claims to have improved power and thermal efficiency, as well as increased maximum frequency from 2.3GHz to 2.5GHz. I suspect the design is quite similar to Krait 400, perhaps with some bug fixes and other minor tweaks. Qualcomm is likely leveraging yield and 28nm HPM process tech improvements to get the extra 200MHz over Krait 400. Krait 450 also adds 36-bit LPAE (Large Physical Address Extensions) to enable memory support above 4GB. This is a similar path to what we saw desktop PCs take years ago, although I'd expect the transition to 64-bit ARMv8 to happen for Qualcomm next year.
The GPU sees the bigger upgrade this round. The Snapdragon 805 features Qualcomm’s Adreno 420, designed 100% in house at Qualcomm. Adreno 420 brings about a D3D11-class feature set to Qualcomm’s mobile graphics, adding support for hull, domain and geometry shaders. Adreno 420 also includes dedicated tessellation hardware. Full profile OpenCL 1.2 is now supported. Texture performance improves by over 2x per pipe, and also gains ASTC support.
Adreno 420 is more efficient at moving data around internally. The GPU has a new dedicated connection to the memory controller, whereas in previous designs the GPU shared a bus with the video decoder and ISP.
Qualcomm insists on occluding things like shader unit counts, so all we have to report today are a 40% increase in shader bound benchmarks (implying a 40% increase in shader hardware and/or more efficient hardware).
Snapdragon 805 also features hardware accelerated decode of H.265 content. Hardware encode is still limited to H.264, but this is an awesome first for Qualcomm.
The Snapdragon 805 brings a much improved ISP. Qualcomm claims more than a 50% increase in ISP throughput (1GPixel/s class) compared to 640MP/s for Snapdragon 800. The 805 leverages its Hexagon DSP to deliver this level of performance. Qualcomm lists no change in DSP architecture between the 805 and 800.
Lastly, we see Qualcomm move to a 128-bit wide LPDDR3 memory interface for Snapdragon 805. With support for LPDDR3-1600, the Snapdragon 805 features up to 25.6GB/s of peak memory bandwidth. It’s interesting to see Qualcomm go this wide just as Apple moved back down to a 64-bit wide interface. Qualcomm and Intel will be the only two shipping such a wide memory interface in the ultra mobile space come next year (although I do expect Apple to return to a wider memory bus at some point).
All of this makes for one beefy SoC, and a continuation of Qualcomm’s leadership in this space. I doubt we’ll see any slowing of Qualcomm’s roadmap after the 805 though. TSMC expects to be shipping 20nm wafers by the end of next year, and I wouldn’t be surprised to find a 20nm successor to the 805 in late ’14/early ’15. Remember that on the last process node shift we got Krait, I wonder what we’ll get this time.
105 Comments
View All Comments
ArthurG - Saturday, November 23, 2013 - link
don't assume things you don't know... this tablet being used a lot by a child, performance in games was on point, but the stylus was the deal makerTehk17 - Friday, November 22, 2013 - link
T4 isn't close to S800 in GPU performance. That's much more important than minor differences in CPU performance at this stage.tuxRoller - Friday, November 22, 2013 - link
T4 is pretty fast in single thread benchmarks but it still hasn't moved to a modern gpu and it's power hungry. They need to give up the companion core and either move to fully power planed cpus, or big.LITTLE.Impulses - Friday, November 22, 2013 - link
Ironically they had more design wins with their most crippled SoC (Tegra 2) and it's only gone downhill from there...Krysto - Sunday, November 24, 2013 - link
Nvidia just needs to stop with the "tablet chip" crap, that Intel is trying to pull, too. I hate it when they do that, because that's just code word for "our chip is not efficient enough for smartphones and maybe we'll have one that is 6 months later".So they need to start from the ground up - from the smartphone level, with their chip making process, and then increase clock speed or whatever, to make a slightly faster version for tablets, even though I think it's unnecessary, and it just means the power consumption rate will be bigger on tablets than smartphones (yes tablets have bigger batteries, but they shouldn't be used to "compensate" for a more power hungry chip like that).
Nvidia said Kepler is very efficient, but we don't know yet what that means in practice. Will they keep it at under 1W of power, or will they make it ridiculously overpowered and consuming like 3+W of energy? I'm hoping it's the former, not the latter.
I also think their delays are getting very annoying. They need to start firing people there. Kepler was supposed to arrive in Tegra 4, and Maxwell and Denver were supposed to be there in Tegra 5. They are 1 year behind their original schedule. That's just ridiculous. Now it means they won't have an ARMv8 chip in 2014 at all. They better have their Tegra 6 at 16nm FinFET shipping in Q1 2015.
melgross - Monday, November 25, 2013 - link
Nvidia is famous for overhyping their products. Their own tests. Which they happily share with the world, or used to, were always far better in results than the real world tests done here, and in other places. I never trust anything they have to say.oranos - Tuesday, November 26, 2013 - link
Buddy, Nvidia can announce a 2TFLOP mobile chipset tomorrow. It doesn't matter. The point is that Qualcomm has it's chipsets in actual phones now. You can go and purchase a myriad of options. Tegra is nothing but a epeen measuring contest for Nvidia until they can get market share in the mobile space.Suneater - Friday, November 22, 2013 - link
You can not listen to Nvidia marketing, but they never lied with their numbers. There's no doubt that Logan will be the most powerful mobile SoC in 2014. I can only agree that Tegra 1, 2, 3 were worse than the competition and tegra 4 equal but not better. Tegra 5 is a completely different story though.melgross - Monday, November 25, 2013 - link
What? Their numbers have been wrong more than they have been right.Suneater - Monday, November 25, 2013 - link
Ok, show me at least one example when they lied! They announced Tegra 4 and showed perf graphics where Tegra 4 was about 10% more powerful than SoC in iPad 4. And it actually turned out to be much more than 10% more powerful (115 GFLOPS in Xiaomi Mi3 vs 76 GFLOPS in iPad 4). So NVidia actually underrates its SoCs!