CPU Boards & Prices


I wish this benchmark site had an OMAP5432 thrown in the mix. I realize it would be imperfect.


I hadn't realized before now that they include both X86 and ARM in their benchmarks suite. I stumbled into this when searching for someplace that could give me an indication of how the 12nm Mediatek P70 in the Cosmo would perform.

PC has said that they will be providing the documentation needed for Debian Linux to work on the Cosmo.
Mediatek has a bad rep in regarding Linux.
It -is- a 12nm processed ARM SoC that is apparently available for niche devices - that may get some Linux support from it's use in the Cosmo.

Something to ponder?
 
The P70 looks pretty interesting. I’ve never worked with any Mediatek hardware, so not sure what their Linux support is like. If we could get some info on that that would be pretty neat. We’d probably have to first get a quote and discuss how many units were interested in. That definitely won’t be happening for a bit though since we have to launch the OMAP5 unit first.
 
The P70 looks pretty interesting. I’ve never worked with any Mediatek hardware, so not sure what their Linux support is like. If we could get some info on that that would be pretty neat. We’d probably have to first get a quote and discuss how many units were interested in. That definitely won’t be happening for a bit though since we have to launch the OMAP5 unit first.

Before I'd even go that far, maybe give the Linux folk playing with the Cosmo six months or so to see how it works for them?
 
Before I'd even go that far, maybe give the Linux folk playing with the Cosmo six months or so to see how it works for them?

Yep, although it’s not especially likely we’ll be able to build off their work as their kernel will likely be behind closed doors.

The P70 also has built in WiFi and LTE which makes things a little awkward as we already have hardware for that.
 
Branch prediction is always associated with speculative execution as far as I know.
AFAIK the more primitive implementations are purely based on statistics, i.e. they always predict the same outcome because analyzing a whole bunch of compiled code showed that one outcome of a comparison is significantly more likely to happen than the other.
 
I dotnt understand much from this Thread, its like "Mandarin" for me,
So for a noob whit a pretty poor Shools Out "Hauptschulabschluss", whats now about the CPU Board and the Prices??

Also i only need the PC Abilities of the Pyra for some Youtube, Netflix, Prime and Office Work at home, and its powers for nothing bigger than Dreamcast..,
So i think the OMAP 5 should be fine for me
 
  • Like
Reactions: rSl
If there will be any future CPU board at all, for sure it will be powered by NXP chip.

I am your clairvoyant, mark my word :cool:
 
Yep, although it’s not especially likely we’ll be able to build off their work as their kernel will likely be behind closed doors.

The P70 also has built in WiFi and LTE which makes things a little awkward as we already have hardware for that.

Their Linux platform development process on the Gemini was pretty open - not great, but pretty open. From what I could tell they created a GUI bootable example OS and the instructions to build it and an OS installer program (To install Android, Linux or Sailfish onto the Gemini from a host computer).

I'm not sure how the WiFi and LTE on the chip would be bad, but it may not be useful to us.

Again, something to ponder. This is the first SoC I have found on a sub 20nm process that has an example of being made available to a relatively niche or small volume project. Planet Computers Cosmo, 4000 units in production, Mediatek P70.

At this point it's just something to throw onto the list of potentials to ponder. There aren't really that many options on that list.
 
AFAIK the more primitive implementations are purely based on statistics, i.e. they always predict the same outcome because analyzing a whole bunch of compiled code showed that one outcome of a comparison is significantly more likely to happen than the other.
It could still leak state though. All you need for speculative leaks to occur as I understand it is a conditional branch followed closely by a memory load. Other code running on the same CPU a few milliseconds later can see what's in the cache by timing accesses to it. Actually, reading the ARM pdf concerning these speculative attacks, it seems to me that Out of order execution is needed on top of branch prediction to get most of these exploits to function. If you're only branch predicting to populate the pipeline going forward, I'd have thought a prediction miss would occur before the next instruction decided to pull values from RAM via the cache, but I guess that depends on the exact construction of the pipeline, if it fetches data a number of steps before finally executing the instruction it's possible it's got data it doesn't end up needing, and the pipeline would be flushed before it hits final execution. But with OoO execution, the processor could decide to run whole chunks of code before finally deciding to implement the conditional branch and not make use of those instructions at the end of the day.

But going back to your point, the traditional examples for many of these bugs is a long lasting loop that finally activates some loop break condition. You can use this to time when the loop ends by watching for the n+1th entry being loaded into the cache. The statistics will say the loop usually goes round again, so the time it doesn't will be predicted wrongly and it may start processing the n+1th entry even though the coder never meant it to.
 
but I guess that depends on the exact construction of the pipeline
Did you just assume my pipeline?!?

CPUs that use such a primitive branch prediction usually have a pipeline of a matching complexity. E.g., the branch prediction in Cortex-M3 CPUs only saves the instruction fetch on a success as its 3 stage pipeline is way too primitive to do more than that. That's basically the bare minimum at which it even makes sense to include any kind of branch prediction, which is why the 2-stage Cortex-M0 doesn't even do any branch prediction.
 
I assume nothing. If I recall the DABS press book that I originally learned ARM assembly from, the classical 3 stage pipeline is described as Fetch->Decode->Execute, although exactly what that means is undefined. I've has a look in the ARMv4 ARM which I happen to have a bound copy of and that doesn't say, it just says that PC is the instruction address+8, which would mean a pipeline length of 3 there too, but doesn't define the stages at all as far as I can dig.
 
Yep, although it’s not especially likely we’ll be able to build off their work as their kernel will likely be behind closed doors.

The P70 also has built in WiFi and LTE which makes things a little awkward as we already have hardware for that.
The MediaTek MT8183 is suspected to be the P70 without LTE and w/o WiFi. It seems, the GPU part isn't actually known. The MT8183's device tree is in the mainline Linux kernel. Though I don't know how complete that is. There's also the evaluation board's device tree - couldn't find the board for purchase anywhere, though.

I read somewhere, Rockchip is working on a A75 or A76 chip (forgot, which one) to be released next year.
 
I assume nothing
I'm not sure if I should be relieved or irritated that you didn't get that joke...

it just says that PC is the instruction address+8, which would mean a pipeline length of 3 there too
While ARMv4 did indeed have a 3 stage pipeline, that detail didn't change and actually became part of the whole ARM architecture. Even if the hardware implementation uses more or less stages for its pipeline, the PC will always be +8.
 
I'm not sure how the WiFi and LTE on the chip would be bad, but it may not be useful to us.
Cost. These smartphone grade processors are incredible expensive per unit (often in the $200-300 range) because of their high performance and high levels of integration. Smartphones need a SOC with everything integrated to save space, but that makes things costly. I wouldn’t be surprised it the P70 at our quantity levels was upwards of $300/unit. The price point of the Cosmo is indicative of this.

Even the iMX8QM breaches $100.
 
Back
Top