400 MIPS for around 5USD with XS1-L1 from XMOS

Hi all,

... the XS1-L1 from XMOS is an interesting chip if you ever want some more 'steam-power'

Viktor

As a transputer fan I have been following the XMOS processors for a bit and have an XC-1 G4 4 Core board. (4 * 400Mips)

They are impresive but limited in some strange ways (memory), the documentation and dev tools are a little new as yet too and need quite a bit more work to minimise pain.

When starting from scratch be prepared for a long haul of wading through treacle. Don't expect to be productive with them too quickly, you will die of frustration.

Other than that they have some clever ideas based around multi threading and event driven IO, some of this would be particularly relevant to machinery control.

My best advice on them though would be to wait 8 months or so and see if they survive the quality of their doc's and tools. XMOS has all the hall marks of an academic startup, phenominaly clever, but they need to understand the needs of their potential customers a little better.

aka47

---

Necessity hopefully becomes the absentee parent of successfully invented children.

... in my diploma-thesis i had to estimate the computing capacity of transputer-grids compared to the actual numbers-chrunchers (around 1987).

I did some occam-programming and simulating on a T4x4-blade (4 T400-Transputers with 20Mips) and a network made from an ATW, some Atari ST's and a MIPS-workstation.

My findings were, that with an ATW and the full capacity of T4-blades for a price of around 250000 Deutschmarks i could reach around 95% of the scalar capacity of a CRAY-XMP for around 2.5 million USD

So this isn't really new for me, but from 20MIPS then to 400MIPS now and only a fraction of the costs this should be a real good deal for starting massive parallel programming in the low-cost/DIY-area

Viktor

I designed and implemented some Ultrasonic NDT processing networks of transputers (algorithmic distribution) for what was then British Rail Research. The equipment scanned rail track at 40Mph and measured rail dimensions (wear analysis) as well as reporting on material failure (cracking and voids).

I agree the XMOS silicone is very much doable for low cost parallel experimentation. If you are prepared to put in a disproportionate amount of effort getting started. THis should improve over time though if they want to compete and survive.


The basic model of link and event based operation is the same as the transputer. The IO makes much of this. The Multithreading is round robin per instruction making it very deterministic. (Transputer was round robin on a time slot basis and could only de-schedule on certain instructions making it indeterministic). The determinism & multithreading plus event driven IO makes the processor perfect for real world machine control and perhaps engine management scenarios.

None of the IO is memory mapped though, it is all accessed through dedicated instructions using the event driving hardware. This means whilst you can do things with it that mere mortal microcontrolers would choke on (incliding writing your own MMU) you cannot expand the on chip memory. You can access off chip memory writing your own memory interface or MMU for their IO but because it is'nt memory mapped you cannot execute from it. It is purely Data Space.

Certain of the key interupts that would enable you to create your own (fast enough) paged memory scheme are not present.

All in all this leaves you restricted to the on core memory and OTP store.

The OTP store being one time programable is a touch old school and limits your ability to update/modify code you put in there. The OTP is again not memory mapped so you cannot execute in place. It is data space only and is only available via the boot process.

At boot up the contents of the OTP are copied to the RAM and execution proceeds from the RAM.

The 64K Bytes then whilst looking generous has to have everything in it (data and instructions) for that core and all of the threads. The 64k now doesn't start to look that generous, particularly as it is a 32 bit processor. Plus you cannot expand it. Create a lookup table or two and you have had it.

What originally looked to be the most revolutionary, the Transputer is back, gotta have, best ever micro is revealed as having an arm and a leg tied behind it's back.

It does have its uses, it is seriously niche though in a way that makes the Transputer look main stream. To get best value out of it you are limited to clever IO in a pairing with another processor. For memory, code and analogue functions. It is sort of competing with FPGA's. A battle it will loose.

If you can't tell from the above, having spent a disproportionate amount of time on it I am quite seriously disappointed. I hope sincerely that XMOS can turn this around but will commit no more time to them until they have addressed the issue with their docs and tools.

All in all they are a relatively young start up of very clever and talented people. They have done some thing very clever. Whether they can see beyond this to satisfy customer needs and survive commercially or not is another question.

All in all though, have a look/go and make your own mind up.

For a Transputer replacement the best option is still looking like fold your own using soft cores and FPGA's.

---

Necessity hopefully becomes the absentee parent of successfully invented children.

... as i don't have the time and ressources, to restart my previous work with the Transputers, i hope XMOS will compete and evolve ...

Massive parallel networks made from simple but fast chips, some sort of cross-linker and local memory have some interesting specs regarding online-reconfiguration of NeuralNetworkSimulation, what's remembering me on the early stages of NN, artificial learning and AI

Viktor

aka47 Wrote:
> For a Transputer replacement the best option is
> still looking like fold your own using soft cores
> and FPGA's.

Indeed... use a common cheap FPGA, softcore programming, and generic boards with a few high power lines. Wonderful idea for getting modular parallel cards that work together. Something along the line of what fpga4fun.com offers... Who's gonna volunteer to do it...

Who ever most wants it....

---

Necessity hopefully becomes the absentee parent of successfully invented children.

Neural Networks

Wow theres a blast from the past. Did'nt Minsky and Papert screw that pooch, purely to arrange preferential funding for their own kinkiness for natural language processing and expert systems (equally dead, but still surprise, surprise, discredited, though more explored, whilst paying their pensions).

Who ever said that (computing) science was egalitarian..........

For me they rank alongside, George W Bush, Erich Honecker, Tony Blair, Margaret Thatcher, Joseph Stalin, Adolf Hitler, Marshal Tito, The Pope, Mary Whitehouse, Generalissimo Franco & Benito Mussolini. Purely for their contribution to the humanitarian cause.

aka47

---

Necessity hopefully becomes the absentee parent of successfully invented children.

... there were/are some nice NN-concepts and working samplec worth reactivating

I remember some NN-/fuzzy-logic-approaches for optimized CNC-controlling.

With one article in a german scientific journal i programmed a NN for optimizing the paths between high counts of drill-holes like - it was the solving of the 'trading salesman'-problem where i had 3 times the count of drillholes as Neurons and they started with random coordinates and 'crystalized' to a connection-diagramm with nearly 98% to 99% correctness in ten minutes with my GFA-Basic on an 8-MHz-Atari ST compared to a comercial programm on a 33MHz-PC-AT with 30 minutes calculating time

Another clever approach was an automatic training network for 'learning' controlling an unknown CNC-machine without any predefs, only with position-feedback ...

Imagine a situation, where you construct a CNC-robot, atach the controller, insert a pen as tool and by image recognition with a web-cam your NN trains/learns to steer the tooltip around your working area.

When learned, you can output any coordinates, either cartesian, rotary, delta or hexapod ... and if you change something in the mechanical or dynamical setup, simply restart the training ...

Viktor

Neural Networks have a special fascination for me.

I somehow feel that they are about as good at repeatability and precision as we are though. Machines can be much better at this than us.

There is always the question of what exactly has the pattern matching converged upon ????

aka47

---

Necessity hopefully becomes the absentee parent of successfully invented children.

> There is always the question of what exactly has
> the pattern matching converged upon ????

... scanning the internet for 'illegal' contents ... fast heuristic search engines ... scanning the road traffic for number plates or human faces

Viktor

There was a note able case in point.

DARPA were using NN to do aerial reconnaissance analysis. Basically looking at piccys of enemy territory to visually look for bunkers, armaments, military activity etc.

They trained the network up to an acceptably high degree of recognition in fairly short order.

But when presented with live data the degree of recognition was terrible.

Eventually they were able to work out what it was. The training set used photos that were all taken on sunny days for pictures containing items of interest and pictures that were of poorer image quality where there was nothing.

NN trains to the patterns that are most obvious to it. Not what you want it to recognize. It is after all a self programming pattern matcher.

Like most things, rubish in rubish out.....

---

Necessity hopefully becomes the absentee parent of successfully invented children.

... this let me remember a similar 'failure' where a NN trained a reconfigurable FPGA(?) for a perfect FFT-filter ...

When it runs with good results, the developer made a presentation in the meeting-room, where the FFT crashed ...

The conclusion was that in the computer-room the ambient temperature were nearly constant around 25 centigrades, in the meeting-room around 22 ... and the NN used the thermal noise of some of the FPGA-resistors for creating 'white noise'.

The next approach was then with three identical setups at different temperatures, which should run unisono/synchrone, what resulted in a successfull FFT running in a temp-range between -20 and +70 centigrades ...

Viktor

Yeah that is pretty much the point, I was working around.

Folk are used to deterministic computing bot in terms of timings and results.

Self organising code, does what it says on the can, it self organises.

To be capable of teaching it we have to be prepared to see the world that the NN is experiencing.

People are very good at anthropormophism, unfortunately it is a one way street.

In both of the examples NN's were very successful at detecting and training itself to what it perceived the difference to be. The fact that what the trainers were testing for and what the NN presented coincided, did'nt guarantee the outcome was what was desired.

So NN is very successful, But at what ???

I would argue NN's are very successful at displaying traits that would be considered intelligence were they found in a biological systems. (people don't expect logic from biology)

---

Necessity hopefully becomes the absentee parent of successfully invented children.