WP_Term Object
(
    [term_id] => 24
    [name] => TSMC
    [slug] => tsmc
    [term_group] => 0
    [term_taxonomy_id] => 24
    [taxonomy] => category
    [description] => 
    [parent] => 158
    [count] => 414
    [filter] => raw
    [cat_ID] => 24
    [category_count] => 414
    [category_description] => 
    [cat_name] => TSMC
    [category_nicename] => tsmc
    [category_parent] => 158
    [is_post] => 1
)

ARM tests out TSMC 10FinFET – with two cores

ARM tests out TSMC 10FinFET – with two cores
by Don Dingee on 05-25-2016 at 4:00 pm

About 13 months ago, the leak blogs posted news of “Artemis” on an alleged ARM roadmap slide, supposedly a new 16FF ARM core positioned as the presumptive successor to the Cortex-A57. Now, we’re finding out what “Artemis” may actually be, inside a multi-core PPA test chip on TSMC 10FinFET.

That same leaked slide had two other code names, “Ares” and “Prometheus”, both at 10nm. “Ares” purportedly was in the Cortex-A72 range, in line as the next ARMv8-A server core. “Prometheus” was the 10nm effort in the Cortex-A57 range for mobile. If you read Chapter 10 of “Mobile Unleashed”, you’ll note I didn’t mention any of that even though it was available before publication. Instinct said to wait.

One thing years of managing product marketing teams has taught me: roadmaps change, and it’s dangerous to buy or sell off just one snapshot of a roadmap. I sat watching a PowerPC processor roadmap where the box in the upper right corner was always two years away, for at least five years of seeing the updated presentation quarterly. In fact, I once went so far as to create a project code name for our board-level roadmap that was difficult to pronounce in some Asian geographies simply to protect the project until the actual announcement.

Putting out a roadmap that is complete fiction is really bad, but adapting a roadmap to keep pace with competition and technological advancements is just plain smart. There are likely a lot of project names running around inside of ARM, and they have just identified “Artemis” as the CPU core inside a test chip taped out last December in TSMC 10FF and about to arrive in silicon.


The big takeaway here is TSMC 10FF shows dynamic power cut pretty much in half, albeit at a slightly lower clock in this test chip. ARM points out the 10FF process and their POP are still unoptimized, and that is exactly why they do these types of test chips with TSMC before turning a new core loose on their customers.

Architecturally however, there is a significant wrinkle in this test chip. It’s not in the quad-core “Artemis” cluster, or in the Mali GPU, but in a Cortex-M0 core on the peripheral side.

The Cortex-M0 has ventured down to 28nm in real-world microcontrollers. I can’t find a mention of Cortex-M0 on 16FF, but I did find one mention of Cortex-M0 on a Samsung 20nm test chip. (Corrections welcome, please cite your source with a link.) 10nm is a huge leap, one no MCU vendor could afford under present conditions. It’s possible someone is looking at integration of a low end MCU-class core on an SoC for tasks like security, however I think there’s more of a fire on the other side of the smoke.

Bottom line: I suspect ARM is toying with something on the roadmap in between the current Cortex-M family and the Cortex-A5, targeting wearables. There is a violent conversation on whether an MMU is actually needed for wearables – see this Samsung presentation at the 2014 Tizen Developer Conference. ARM may have just used the Cortex-M0 in this test chip for simplicity’s sake to get a read on 10FF PPA for a smaller core.

There was a box on that leaked roadmap slide from 13 months ago named “Mercury” that resembles those remarks. It would be stunning to see if ARM tries to move the wearable ball from 28nm down to 10nm in an attempt to up the node volumes and get TSMC’s costs down for everyone. (I was at a recent conference where someone quipped that they thought the number in front of the “nm” was the number of customers – probably not far from the truth, if the conversation is mobile phone SoCs only.)

In any case, “Artemis” isn’t yet a product, and ARM’s press release on the 10FF TSMC test chip is vague. I had to dig around to find more details in AnandTech and HotHardware. The real 10FF core announcement is probably coming at ARM TechCon 2016, and according to TSMC the first actual chips are due around this time next year.

The thought that somebody might figure out how to drive serious volumes with wearables at 10FF is intriguing – you read it here first if things go that way. Now that I’ve thrown my wild speculation into the mix, what do you think?


0 Replies to “ARM tests out TSMC 10FinFET – with two cores”

You must register or log in to view/post comments.