AI Dominates Qualcomm Snapdragon Summit With New Snapdragon Products

This last year has been a whirlwind of both hype and reality for AI, with nearly everyone in the tech industry talking about it. While there is no doubt that companies have doubled down on their AI efforts, many of their advances would simply be impossible without the efforts of chip companies like AMD, Intel, MediaTek, Nvidia and Qualcomm. While Nvidia has without a doubt taken leadership in the AI training space, the AI inference side is still very much up for grabs and will be heavily influenced based on the AI capabilities of platforms like PCs and smartphones.

Qualcomm has been investing in the client side of AI for many years. Some would even argue that the company was way ahead of its time prioritizing AI performance in its SoCs, but it seems that Qualcomm’s iterative approach may have paid off with the latest generation of mobile platforms. That brings us to this year’s Snapdragon Summit, where Qualcomm always announces its latest platforms for the coming year. This year we can expect a new smartphone platform as well as a new PC platform utilizing the company’s new Oryon CPU, which it built based on its acquisition of Nuvia in 2021.

Snapdragon 8 Gen 3

The Snapdragon 8 Gen 3 is Qualcomm’s 4nm next-generation flagship SoC, which sports the most advanced IP around. Qualcomm has upgraded its high-performance CPU core to an Arm Cortex X4. (This could be the last time we see an Arm Cortex X Series in a Snapdragon 8 series platform; more on that later.) This prime core runs at up to 3.3 gigahertz, while the remaining performance cores run at 3.2 gigahertz, and the remaining two efficiency cores run at 2.3 gigahertz. I suspect that this configuration may change next year if we get Oryon cores in mobile.

Qualcomm claims that this CPU has 30% better performance and 20% better power efficiency compared to the last generation. On the GPU side, the new Adreno GPU is claimed to have 25% better GPU performance and 25% better power efficiency, including 40% better ray tracing performance, compared to the last generation’s Snapdragon 8 Gen 2. Qualcomm is also implementing a new light reflection system that adds global illumination with ray tracing. The GPU also now supports variable refresh rate from 1 hertz to 240 hertz, which adds a new range of frames per second up to 240. Qualcomm also worked with Epic Games to bring Unreal Engine 5’s Lumen global illumination and reflections to mobile. This will once again make the Snapdragon 8 Gen 3 one of the best platforms for mobile gaming.

Given that cameras are still such a key feature on flagship phones, Qualcomm has not let off the gas pedal on improving the image signal processor in the SoC. Qualcomm’s Spectra ISP inside the Snapdragon 8 Gen 3 comes with what the company calls a cognitive triple 18-bit ISP configuration with up to 12 layers of semantic segmentation. Having that many layers of semantic segmentation allows the camera to differentiate between objects and materials with more granularity, giving it more control over modifications to the image. The improved ISP also brings support for a slew of new HDR image sensors for multi-frame HDR as well as less blanking. There is also support for a multitude of photo and video modes—too many to list here. This new ISP will also support what Qualcomm calls Night Vision video capture, which comes with a RAW AI noise reduction captured in 4K60.

Speaking of AI, the new Snapdragon 8 Gen 3 doubles down on the last generation’s huge improvements. Qualcomm’s investments in AI hardware go back many years (to 2017) so that now nearly every component in a Snapdragon SoC has some sort of integrated AI capabilities. Naturally this includes the CPU, GPU and neural processing unit, but it also extends to the ISP, modem and sensor hub. (Note that Qualcomm’s Hexagon has transformed from a DSP into an NPU.) Going as far back as 2018, the Snapdragon 845 already had 3 TOPS of AI performance and now, five years later, the Snapdragon 8 Gen 3 peaks at 45 TOPS, an increase of 15x. To support all these memory-intensive AI capabilities, the Snapdragon 8 Gen 3 will also support memory capacities of up to 24GB at up to 4800-megahertz speeds. The Snapdragon 8 Gen 3 will also work with Micron’s latest DDR5x, which runs at 9.6 Gbps in capacities up to 16GB.

The Snapdragon 8 Gen 3 also borrows many of Qualcomm’s strong suits from other business units, for example by embracing the X75 5G modem and the FastConnect 7800 to deliver Wi-Fi 7 and Bluetooth 5.3. These are less-exciting features mostly because they have already been announced and are well-known pieces of IP.

Snapdragon X Elite

The Snapdragon X Elite is Qualcomm’s strongest effort in the realm of platforms for the PC. The company seems extremely confident in the platform’s capabilities, especially since it features 12 of the company’s new custom 4nm 64-bit Oryon CPU cores. Each of these cores can clock up to 3.8 gigahertz, and a couple of them can boost up to 4.3 gigahertz, delivering some great performance. If successful, these Oryon cores should show up in the smartphone, XR and automotive platforms as well. Qualcomm’s CEO even hinted on stage that we should expect Oryon in next year’s Snapdragon 8 Gen 4. This is a departure from previous designs, in which Qualcomm would pair a big performance core with smaller efficiency cores, but because these Oryon cores are so good at being both fast and low-powered, there’s no need to have separate efficiency cores. It will be interesting to see how Qualcomm configures its future platforms using Oryon CPU cores.

In addition to the CPU, Qualcomm also included a powerful GPU with 4.6 TFLOPs of compute performance and an equally powerful NPU with 45 TOPS of performance. Taken together, these enable up to 75 TOPS of system AI performance across all the cores, even though there is unlikely to be a workload that can utilize all 75 TOPS. The Snapdragon X Elite also supports up to 64GB of LPDDR5x at 8533 MT/s (mega transfers per second), which is considerably more bandwidth than the Snapdragon 8 Gen 3. Qualcomm claims 135GB per second of memory bandwidth with this memory subsystem, which is important for more memory-intensive workloads such as video editing and AI.

In terms of storage, the Snapdragon X Elite supports both NVMe PCIe Gen 4 and UFS 4.0, which should deliver enough bandwidth to the internal SSD, but I’d like to see Qualcomm support PCIe 5 storage sometime soon. The Snapdragon X Elite also embraces the Snapdragon X65 5G modem, which gives the Snapdragon 8 Gen 3 an edge on 5G connectivity. That said, the X65 is still a great modem, and Qualcomm has also paired it with the FastConnect 7800 in the Snapdragon X Elite, so any device with this pairing will still have great Wi-Fi 7 and Bluetooth connectivity.

Qualcomm has made a lot of competitive claims against both Intel and AMD, for example stating that it delivers 2x faster CPU performance in multi-threaded applications versus the Intel i7 1360P. Qualcomm also claims to match Intel’s power in this class at 68% lower power consumption. Something to consider is that this processor from Intel features only four performance cores and eight efficiency cores, making it unlikely to be competitive with the Qualcomm product in terms of performance, but it still should be more competitive on power. That said, Qualcomm did compare the Snapdragon X Elite against the Intel i7 13800H processor, which I believe to be a better comparison; in that case, Qualcomm claims 60% better performance at ISO power and 65% less power consumption at the same performance. Qualcomm even dared to roll out a comparison against Apple’s M2, claiming 50% faster peak multi-threaded performance. I cannot wait to validate these claims myself.

Qualcomm continued the comparisons on the GPU, claiming a 2x faster GPU than the i7 13800H from Intel while also using 74% less power at equal performance. Qualcomm also compared the Snapdragon X Elite against AMD’s Ryzen 9 7940HS, claiming 80% faster GPU performance at ISO power and 80% less power consumption versus the competition at the same performance. Qualcomm even touted the AI performance of the NPU on the Snapdragon X Elite, claiming a 100x increase in NPU performance since 2017 on its own platforms while placing its competition in a much lower performance bracket. Qualcomm is clearly going for the x86 jugular, and it will be really interesting to see how this manifests itself in PCs next year.

Snapdragon S7 Gen 1 And S7 Pro Gen 1

At the Snapdragon Summit, Qualcomm also introduced an entirely new tier of its wireless audio products with the Snapdragon S7 Series, which sits atop its current flagship S5 series. This new S7 series includes two products, an S7 and an S7 Pro. Both of these come with considerable improvements across the board to give wearable audio product manufacturers more flexibility. The S7 series comes with a 3x increase in DSP performance, a whopping 100x increase in on-device AI performance, 6x faster compute and 3x more memory compared to the previous generation. The S7 series updates to Qualcomm’s fourth generation of ANC and adaptive transparence mode.

The S7 series also comes with Bluetooth 5.4 and Auracast broadcast audio, while the S7 Pro adds Qualcomm’s XPAN technology, which takes advantage of Wi-Fi to enable a wearable to maintain an audio connection beyond the range of Bluetooth. I expect that we will see a lot of Snapdragon S5 Gen 2 devices transition to the S7 Gen 1 in their next generation; some might even adopt the XPAN technology, which I believe will remain a niche offering but also be used as differentiation from other products on the market.

Wrapping Up

Qualcomm continues to lean into AI performance on both its smartphone and PC platforms with the Snapdragon 8 Gen 3 and the Snapdragon X Elite. There is no doubt that Qualcomm has established itself as the leader in mobile AI performance. With inference becoming increasingly costly in the cloud as well as latency-sensitive, there will be an increased shift towards client-side processing. Once that happens, we’ll see more competition among the different smartphone and PC vendors for inference leadership. On the mobile side, it remains to be seen if anyone can catch Qualcomm, especially considering its tight partnership with Meta and support of Llama 2 and other cutting-edge LLMs.

Qualcomm also seems to be on the cutting edge for the PC market when it comes to NPU performance and overall system AI performance. I think this lends to the Snapdragon X Elite being a popular platform for next year’s AI PC push, which seems to be happening across the industry thanks to Microsoft’s deep investment in AI, including its partnership with OpenAI. I believe that Qualcomm’s biggest hurdle will continue to be compatibility with 64-bit Arm apps, unless Microsoft makes a considerable effort in the next generation of Windows to push developers to support native 64-bit Arm applications. That said, the relationship between Microsoft and Qualcomm seems to be growing, and I wouldn’t be surprised to see the Snapdragon X Elite leading the industry into the AI PC era.

Qualcomm has spent considerable time and resources to enable high-performance AI on both the smartphone and PC with its iterative approach to performance and its AI software stack. The real question will come down to how much OEMs and developers will be able to take advantage of those capabilities, or whether we will have to wait another cycle to see these AI applications really take hold.

 

Source link