TSMC: Supercomputing, AI Will Drive Semiconductor Business, Not Phones

TSMC: Supercomputing, AI Will Drive Semiconductor Business, Not Phones

For the last few years, phone and tablet SoCs have been the driving force behind product cycles and foundry technology improvements. Over the last decade, yearly product introduction cycles for phones became the norm, and the entire semiconductor industry retooled to match this requirement. Now, TSMC thinks the bloom is off this particular rose.

HPC — high-performance computing — has replaced phones as the major driver of revenue, according to TSMC. While this includes some hardware common to the PC, like GPUs, it also includes accelerators, new specialized types of microprocessors designed for inference and machine learning workloads, AI processors, and a wealth of other types of products. There’s probably even some overlap between categories; some companies like Qualcomm have played up the idea of leveraging its SoC capabilities of CPU, GPU, and DSP in the same package as a heterogeneous compute platform.

TSMC headquarters in Hsinchu, Taiwan.
TSMC headquarters in Hsinchu, Taiwan.

TSMC, meanwhile, is expecting the profits for those wins to wind up in its own coffers. The company projects annual revenue for 2018 could be up as much as 15 percent, with the semiconductor industry as a whole growing roughly 8 percent. This graph from Statista shows market share data for TSMC versus everybody else.

TSMC: Supercomputing, AI Will Drive Semiconductor Business, Not Phones

Whoof. GlobalFoundries and UMC check out alright, but everybody else is buried back in the minor leagues (and UMC really isn’t a leading-edge foundry). Samsung, were it on this graph, would presumably account for as much market share as GF, but Samsung isn’t a pure-play foundry (it builds hardware for its own use as well).

Shifting over to HPC as a revenue driver, as opposed to the phone market, implies some interesting things about the future of phones. It was companies like Apple and Samsung that pushed foundries to roll out yearly cadence updates and indirectly encouraged a reduction in capability between foundry nodes. The reason we now hear so much about second- or third-generation process nodes is because everyone now expects each new SoC to offer improved performance on yearly cadences, even when those cadences don’t align with foundry upgrade cycles.

Overall, we don’t expect introduction speeds to slow down much. While shifting away from phones with their yearly cadences might seem attractive in certain businesses, there’s an army of marketing teams and a keeping-up-with-the-jones’ mentality that push for yearly product cycles, even when they don’t necessarily make much sense. TSMC’s sales growth may also be the result of winning back Qualcomm, which is rumored to be moving back to TSMC with the Snapdragon 845.

Top image credit: Facebook data center

Continue reading

Protect Your Online Privacy With the 5 Best VPNs
Protect Your Online Privacy With the 5 Best VPNs

Investing in a VPN is a smart choice right now, but the options are vast. To help narrow things down a bit, we've rounded up five of our very favorite consumer services.

RISC-V Tiptoes Towards Mainstream With SiFive Dev Board, High-Performance CPU
RISC-V Tiptoes Towards Mainstream With SiFive Dev Board, High-Performance CPU

RISC V continues to make inroads across the market, this time with a cheaper and more fully-featured test motherboard.

The PlayStation 5 Will Only Be Available Online for Launch Day
The PlayStation 5 Will Only Be Available Online for Launch Day

The PlayStation 5 isn't going to be available in stores on launch day, and if you want to pick up an M.2 SSD to expand its storage, you'll have some time to figure out that purchase.

ARMing for War: New Cortex-A78C Will Challenge x86 in the Laptop Market
ARMing for War: New Cortex-A78C Will Challenge x86 in the Laptop Market

ARM took another step towards challenging x86 in its own right with the debut of the Cortex-A78C this week. The new chip packs up to eight "big" CPU cores and up to an 8MB L3 cache.