[ale] Really cool new hardware

Solomon Peachy pizza at shaftnet.org
Tue Jan 19 17:05:59 EST 2021


On Tue, Jan 19, 2021 at 03:29:04PM -0500, jonhall80 at comcast.net wrote:
> All true, but if the core is not designed to run at the lower voltage 
> than turning up the clock boosts the power input and the core runs 
> hotter.  To say voltage has "nothing to do with the CPU core" is a bit 
> cavalier.

Licensable CPU cores (ie what we're talking about here, as opposed to a 
specific CPU/SoC model) have been supplied in synthesizable 
verilog/vhdl/etc form for the past two decades.

While there is usually a target physical process in mind (especially 
with respect to specific power/performance/area targets) there is 
nothing inherent to the HDL core design itself that _requires_ a given 
process.

Where things get highly process-specific are the analog components (eg 
the pad ring, clocks/PLLs, ADC/DACs, SERDES, GPIOs, etc).  These 
components don't tend to benefit much from process improvements.

Granted, with non-licensed or in-house cores the CPU designers tend to 
hand-tune the design for the best possible performance/power/area on the 
given target node, but with each new process/tooling generation and 
exponential transistor count growth this has grown both less necessary 
and increasingly more difficult.

> Also the fact that ARM does not compete with the ARM chip licensees.  
> There is no "ARM ARM" that you can buy on the marketplace as you can 
> buy an "Intel x86".

NVIDIA's attempt to acquire Arm may threaten this.

> Your comments here are also true, but as memory models change you may 
> find that the cores/GigaByte-power also changes.  As memory is 
> redesigned to use less and less power then reducing the amount of 
> power that the cores use is more significant.

While power-per-bit keeps going down, memory and data set sizes are 
growing ever larger, and the more cores you have on-die the greater the 
load and contention on the memory and I/O subsystems.

I'm not feeling terribly optimistic on this front; while algorithm 
design has certianly improved in the past few decades, we're hitting the 
point where mainstream stuff (ie not just classical HPC) has started to 
feel the memory bandwidth/data locality pinch.

> There is at least a third path, and that is the one provided by 
> vendors such as SiFive where they develop the tools and allow you to 
> use them (granted, for a fee) but the output is yours and they also 
> have the expertise to help you do it.

How is this different than directly licensing the EDA tools from (eg) 
Cadence?  The only difference I see is who you're renting the tools 
from.

> I did not say that "Extended capabilities" were at the ISA level, but 
> the openness of the architecture and the process could allow a tighter 
> integration of extended capabilities than you would ever get from 
> Intel, AMD or any of the "chip makers", perhaps even ARM (unless you 
> were willing to pay a lot just to get their attention).

That's not as much as a panecea as it first sounds; it turns out that 
extending the ISA without inadvertantly breaking something else takes a 
lot of specialized skill (and a metric f-ton of verification testing), 
and there are no shortcuts to be had.

> Finally, as you keep (rightly) pointing out that chips have yet to 
> appear in the high(er) performance end, the same could be (and is) 
> said of ARM architecture.  However, the world has been hit heavy with 
> the performance of the new Apple chip, which can indicate that the 
> performance is possible when the market demand brings about the 
> investment.

Yeah, Apple is one of the few organizations out there that has the 
volume (and margin) to design a high-performance CPU+GPU solely for 
their own products. This tight vertical integration allows them to make 
some trade-offs that others cannot (eg the relatively paltry maximum 
RAM, poor external I/O, and making it borderline illegal to tinker..)

It's funny how we've now gone full circle from vertically-integrated 
companies divesting themselves of their chipmaking and chip-designing 
arms in the name of competitiveness, only to see the opposite happening 
for that same reason.

It will be interesting to see where Apple takes this next, but there's 
nothing particularly magical about what they did.  

> But I am also very aware of the cost of starting a business when the 
> first thing you have to do is talk to lawyers and license managers.

...ain't that the truth.

 - Solomon
-- 
Solomon Peachy			      pizza at shaftnet dot org (email&xmpp)
                                      @pizza:shaftnet dot org   (matrix)
High Springs, FL                      speachy (freenode)
-------------- next part --------------
A non-text attachment was scrubbed...
Name: signature.asc
Type: application/pgp-signature
Size: 833 bytes
Desc: not available
URL: <https://mail.ale.org/pipermail/ale/attachments/20210119/afcacf8d/attachment.sig>


More information about the Ale mailing list