We’re used to thinking of the PC market as the battleground between AMD, Intel
and nVidia. This year we finally have another chip titan joining the PC fray: Qualcomm.
Jensen Huang hosted a pre-Computex keynote, which mostly focused on data center AI
hardware and software. Though he introduced nothing that will be making its way to
personal computers, with not even a brief mention of the eagerly anticipated Blackwell
based desktop GPUs, but there was some cool stuff in his demos. He did mention that the Vera Rubin would succeed Grace Blackwell next year.
Earth 2 is a major technology tour de force, able to not only model the whole world’s
weather, but also make local weather predictions down to regions the size of just tens of meters, including detailed models of air currents around tall buildings like the ones that dominate the Taipei skyline.
The nVidia Inference Microservice (NIM) platform is growing; there’s a growing suite of
microservices available for developers to relatively easily incorporate various AI models
into their applications. For now, nVidia remains focused on the data center.
CEO Cristiano Amon made his Computex keynote debut in 2024 to promote the
Snapdragon X Elite, with 16 personal computers slated to start landing on the shelves in a few weeks and more in the works.
Of course, no keynote presentation these days can be complete without mentioning AI,
and Amon did not let us down there, either. The Snapdragon X Elite includes a Qualcomm Hexagon NPU that delivers 45 TOPS in INT8, which in addition to being quite an impressive level of AI performance is also the first NPU powerful enough to meet
Microsoft’s minimum requirements for Windows 11 Copilot+, scooping both Intel and AMD.
The Hexagon NPU is also very power efficient, as is the rest of the Snapdragon X. Amon
says that this processor will enable computers with multi-day battery life.
In addition to working closely with Microsoft and Google to ensure that their software runs well on and takes full advantage of that powerful Hexagon NPU, Qualcomm also has Blackmagic on board with DaVinci Resolve. Amon also said that there are more than 1200 games tested, optimized, and ready to entertain Snapdragon X Elite computer buyers.
Qualcomm has done an excellent job of building the ecosystem around the Snapdragon X Elite, so if it lives up to its hype, it’s going to be great for consumers.
Headlining Computex is AMD’s Lisa Hsu. Kicking the conference off. She started by unveiling the Ryzen 9000 series of CPUs based on the new Zen5 architecture, offering an average performance boost over Zen4 of 16%. The desktop models range from 6 to 16 cores, all multithreaded, and all supporting PCIe 5.0 and DDR5. They are all arriving in July.
The 3rd Generation Ryzen AI NPU, code-named Strix Point, now has the official name
Ryzen AI 300 series. These all have Zen5 cores, integrated RDNA 3.5 GPUs, and 3rd
generation NPUs boasting up to 50 TOPS.
The 3rd generation NPU developed with Xilinx has a new trick up its sleeve to boost
performance, a new data type for AI called Block FP16. This offers the performance of
INT8 without sacrificing precision and without incurring any penalties for quantization.
Unsurprisingly, AMD also has a partnership with Microsoft Copilot+ along with a long list of upcoming laptops coming later this year using the Ryzen AI 300 series processors.
Since launching the EPYC line of datacenter CPUs, AMD has reached a market share of
33% and growing. It’s especially popular with hyperscalers due to its high compute density as well as its performance.
AMD unveiled its 5th generation EPYC processor, code named “Turin,” based on Zen5.
Packing 13 chiplets into a single package for a total of 192 cores and 384 threads, Turin is a monster. It will be available in the 2nd half of 2024.
To introduce the Instinct MI300X Lisa Hsu invited Christian Laforte, CEO and CTO of
Stability.ai onto the stage to show off what Instinct and Stable Diffusion can do. Stability.ai wants more bandwidth and memory, so Dr. Lisa Hsu unveiled the Instinct MI235X along with its successor the CDNA4-based MI350 due in 2025. The MI325X is an MI300X with more goodies, namely more bandwidth and memory in the form of up to 288 GB of HBM3E and 6 TB/s of bandwidth. CDNA4 is going to be 35x faster, a huge generational performance jump.
Powerful processors also consume vast quantities of data, so scaling up from a single
GPU to hundreds requires vast quantities of bandwidth. To address that need, AMD is
developing the Ultra Accelerator Link which will be available later this year. Since AMD is
part of the Ultra Ethernet Consortium that includes industry titans HP/Enterprise, Intel,
Meta and Microsoft, UA Link will interoperate with Ultra Ethernet, part of AMD’s plan to
scale out its datacenter hardware.
Intel officially unveiled Lunar Lake in Taipei. Continuing to evolve its big/little architecture with an emphasis on efficiency and, you guessed it, AI. A rather unexpected quirk here is that Intel farmed out Lunar Lake fabrication to TSMC because Intel Foundry hasn’t caught up to TSMC yet.
Intel also took a page out of Apple’s book and integrated 32 GB of LPDDR5X memory into the chip package, making the memory configurations fixed at the point of manufacture.
Lunar Lake includes a new NPU cleverly called NPU 4, rated at 48 TOPS and therefore
also Copilot+ ready. Also integrated is an Arc Xe2-LPG GPU which also contributes AI
performance, for a total of 120 TOPS.
Even though Intel still reigns supreme in the thin and light notebook market, its emphasis on efficiency reflects the pressure it’s seeing from Apple and AMD, and now from Qualcomm. Intel claims that Lunar Lake outperforms Snapdragon X Elite and leads it in efficiency. We’ll see when Lunar Lake arrives this fall.
Intel also teased Panther Lake, due in 2025 and manufactured on Intel’s 18A (Angstrom) fabrication node. Intel is aiming to power up the first 18A wafer next week.
On the datacenter side Intel has some interesting products in the works, including Sierra
Forest, a Xeon processor with only efficiency cores. 288 efficiency cores. Meanwhile, Intel is seeking to take some AI market share from nVidia by leading on price with Gaudi 3, unveiled earlier this year.
Like AMD, Intel also discussed scaling, UALink, and Ultra Ethernet. Both Intel and AMD
endorsing Ultra Ethernet shows that it already achieved critical mass.
As far as AI has advanced thus far, it’s just getting started. The monstrous power of the
hyperscale systems are both awesome and terrifying, but that’s not even scratching
the surface of the potential of AI.
Cerebras worked with researchers at Sandia and Lawrence Livermore National Labs to
experiment with AI-accelerated molecular dynamics simulations. The researchers
compared the performance of the tuned simulation on just one WSE-2 to the performance of Frontier, the most powerful supercomputer in the world with 9,472 AMD EPYC 64-core CPUs and 34,888 Instinct GPUs.
In a manner of speaking the WSE-2 won that contest. In a more precise and accurate
manner of speaking, the WSE-2 steamrolled the Frontier by a factor of 179x. Considering the fact that the WSE-3 is currently shipping, is 2x faster, and scales up to 2048 chips, the molecular dynamics test is simply mind-boggling.
But there’s more.
A group of researchers at the European Weather Center trained an AI with the vast, rich
meteorological dataset that the European Centre for Medium-Range Weather Forecasts
has been gathering for decades and tested the model on a personal computer. The AI-trained model has been able to outperform the physics-based models running on
hyperscale supercomputers in some areas. While getting data into the system still requires an enormous amount of work, this raises the prospects of being able to run a weather forecasting system on an ultralight notebook computer with a day or more of battery life.
The sheer power of AI has two immediate benefits. One is enabling computers to
accomplish tasks that seem to be far beyond their computational limits, and another is to drastically reduce power consumption. NPUs require less power than traditional computing machinery for the same tasks, as long as they’re trained properly for those tasks.
And that makes it even harder to predict where technology will take us in the next five
years.