foax wrote on 2024-09-21, 20:40:

This is so cool, I'm definitely going to get my hands on a few of these to do some oldschool LAN gaming!

Maybe I've missed it somewhere but how big are these boards? Apologies if you've already answered this! I have an cute little Moomin biscuit tin that I reckon I can use an an enclosure for one.

Haha, that is a very weird size requirement! 😆
I'm currently working on the layout of the board with the MXM connector. It's currently about 96x90mm. Height will depend very much on what kind of cooling/fans are used.

Given the first one was seemingly crazy enough that it might work, perhaps it's time to drop some more dumb ideas? Definitely think MXM opens up a lot of possibilities btw. Just on some of the points raised in the last page or two:

1 - Cooling
Easily the best performance for its size, for things with a relatively limited power budget like this, is the AirJet: https://www.froresystems.com/
It's not particularly energy efficient, but given this isn't being powered by batteries, a couple of extra watts of waste energy is no great concern. What is, however, is availability. You can't pluck one off the shelf like a Noctua, or even something more exotic/niche like a Dynatron. As I understand it, they're now selling them in dev kits with 20 units (https://www.froresystems.com/products/snowgoo … development-kit) and that's about it, unless you're buying in volume. How much that dev kit costs and whether that figure is realistic on a per system basis (you'd probably need 1-2x of the mini tiles for the CPU, and 2x for the GPU), I have no idea. But the whole solid state/piezoelectric cooling concept is neat, and would definitely keep things compact.

2: Clock controls
Although something as elegant as Throttle Blaster - which is a super cool project btw - isn't doable here without a ton of work, going on a quick scan of this (http://www.bitsavers.org/components/dmp_elect … ence_091216.pdf), there should be a way to intercept the A0h value and inject changes on the fly based on some kind of hardware trigger/switch - or at worst through setmul-like software. That would give 37.5; 50; 60; 75; 100 and 150. Disabling caches and the various other tricks setmul and the like could give some additional granularity from there.

jerms wrote on 2024-09-22, 12:24:

Given the first one was seemingly crazy enough that it might work, perhaps it's time to drop some more dumb ideas? Definitely th […]
Show full quote

Given the first one was seemingly crazy enough that it might work, perhaps it's time to drop some more dumb ideas? Definitely think MXM opens up a lot of possibilities btw. Just on some of the points raised in the last page or two:

1 - Cooling
Easily the best performance for its size, for things with a relatively limited power budget like this, is the AirJet: https://www.froresystems.com/
It's not particularly energy efficient, but given this isn't being powered by batteries, a couple of extra watts of waste energy is no great concern. What is, however, is availability. You can't pluck one off the shelf like a Noctua, or even something more exotic/niche like a Dynatron. As I understand it, they're now selling them in dev kits with 20 units (https://www.froresystems.com/products/snowgoo … development-kit) and that's about it, unless you're buying in volume. How much that dev kit costs and whether that figure is realistic on a per system basis (you'd probably need 1-2x of the mini tiles for the CPU, and 2x for the GPU), I have no idea. But the whole solid state/piezoelectric cooling concept is neat, and would definitely keep things compact.

2: Clock controls
Although something as elegant as Throttle Blaster - which is a super cool project btw - isn't doable here without a ton of work, going on a quick scan of this (http://www.bitsavers.org/components/dmp_elect … ence_091216.pdf), there should be a way to intercept the A0h value and inject changes on the fly based on some kind of hardware trigger/switch - or at worst through setmul-like software. That would give 37.5; 50; 60; 75; 100 and 150. Disabling caches and the various other tricks setmul and the like could give some additional granularity from there.

Cooling: Yeah those Frore units look nice - not a good sign when there's no price given though... :p Will be keeping an eye on those for later!
Clock controls: Disabling caches is already implemented - disabling the L2 doesn't do much, a few percent slower perhaps. The L1 on the other hand definitely helps! Not quite sure what you have in mind with those A0h shenanigans, feel free to elaborate!

Update on the 3.3 (MXM) revision:
- Extended board to be able to fit the MXM connector and mounting studs.
- Bigger and more usable power- and reset buttons.
- VGA- and HDMI connectors (signals coming from the MXM card).
- MicroSD card slot for the Raspberry Pi CM4. Much more handy than relying on the small eMMC and USB bootloader programming.
- Bigger and better PC speaker.
- Two PWM fan connectors instead of one (EMC2303 vs EMC2301 controller).
- Revised power system. It now requires a 45W (15V @ 3A) USB PD charger instead of a 2-3A 5V one. Needed to provide higher-voltage power to the MXM board. The 15V is stepped to down 5V using a switched buck converter, and 5V is further stepped down to 3.3V with linear converters (as before).
- 6-layer PCB for better routing and power delivery.
- The camera module connector is gone (sorry, Ahmad! 😬). It was just too niche, better to use a standalone Pi Zero2 or something for KVM.
- A single SATA connector, which we know works great from my ITX board.

I'll be making an MXM carrier board for Mini-PCIe cards as well, but hopefully we'll be able to use @sdz's wonderful Voodoo4 boards and maybe other DIY cards as well, down the road! 👍

Eivind wrote on 2024-09-23, 21:47:

Update on the 3.3 (MXM) revision: - Extended board to be able to fit the MXM connector and mounting studs. - Bigger and more usa […]
Show full quote

Update on the 3.3 (MXM) revision:
- Extended board to be able to fit the MXM connector and mounting studs.
- Bigger and more usable power- and reset buttons.
- VGA- and HDMI connectors (signals coming from the MXM card).
- MicroSD card slot for the Raspberry Pi CM4. Much more handy than relying on the small eMMC and USB bootloader programming.
- Bigger and better PC speaker.
- Two PWM fan connectors instead of one (EMC2303 vs EMC2301 controller).
- Revised power system. It now requires a 45W (15V @ 3A) USB PD charger instead of a 2-3A 5V one. Needed to provide higher-voltage power to the MXM board. The 15V is stepped to down 5V using a switched buck converter, and 5V is further stepped down to 3.3V with linear converters (as before).
- 6-layer PCB for better routing and power delivery.
- The camera module connector is gone (sorry, Ahmad! 😬). It was just too niche, better to use a standalone Pi Zero2 or something for KVM.
- A single SATA connector, which we know works great from my ITX board.

I'll be making an MXM carrier board for Mini-PCIe cards as well, but hopefully we'll be able to use @sdz's wonderful Voodoo4 boards and maybe other DIY cards as well, down the road! 👍

The attachment Screenshot 2024-09-23 at 23.14.15.png is no longer available

The attachment tinyllama3.3_render1.jpg is no longer available

The attachment tinyllama3.3_render2.jpg is no longer available

The attachment tinyllama3.3_render3.jpg is no longer available

Amazing. Thank you for removing the camera connector as I personally found no value in it. Also, thank you for the SATA connector.
One more thing, what do you think about adding a temperature sensor under the MXM module to control the fan speed?

@Eivind, just wondering, what version of MXM are you implementing. Is this design going to solely work for the Voodoo MXM card or it can support some other MXM cards too?

The V4 MXM has a temperature sensor right underneath the VSA, and a FAN connector. The card is an MXM 3.0/3.1 Type A.
Other cards could work in such a slot, from GTX285M up to some RTX cards, but personally I don't think such a pairing would make sense. Also, even though in the standard there is support for VGA (and LVDS for that matter), the vast majority of these cards do not have this output. All they have is a bunch of eDP/DP++ outputs, and a separate DP/eDP to VGA converter IC is used on the motherboard.
There is also the problem that most of these cards have an MXM VBIOS, and won't play nicely on a regular system.

Thanks for chiming in, sdz!

sdz wrote on 2024-09-24, 07:50:

The V4 MXM has a temperature sensor right underneath the VSA, and a FAN connector.

Is it the VSA doing the fan controlling itself, or something custom you've implemented? Variable-voltage DC or PWM controlled?

sdz wrote on 2024-09-24, 07:50:

Other cards could work in such a slot, from GTX285M up to some RTX cards, but personally I don't think such a pairing would make sense.

Agreed. While a modern MXM card could work in theory (barring VBIOS issues?), there would almost certainly be driver issues for use in a retro system such as this.
Reverse-engineering and recreating older PCI cards is definitely the way to go, I say.

There's a microcontroller on the board that measures the VSA temperature and controls the fan with PWM. The PWM signal is also routed to the MXM card edge connector as well as an overtemperature signal for the system. It also:
-handles the power sequencing, monitors PMICs status and can shut off the whole board depending on VSA temperature, internal temperature and PMICs state.
-switches the VSA straps to configure 32MB or 64MB framebuffer
-controls a digipot connected to the VSA VCORE PMIC, and can set the VSA VCORE voltage between 2.5V and 3.1V in 0.1V steps
-implements bidirectional communication with the host OS via the PCIe-PCI bridge IC so that various things can be set and read
-and some other things that are not useful for this project

sdz wrote on 2024-09-24, 08:46:

There's a microcontroller on the board that measures the VSA temperature and controls the fan with PWM. The PWM signal is also r […]
Show full quote

There's a microcontroller on the board that measures the VSA temperature and controls the fan with PWM. The PWM signal is also routed to the MXM card edge connector as well as an overtemperature signal for the system. It also:
-handles the power sequencing, monitors PMICs status and can shut off the whole board depending on VSA temperature, internal temperature and PMICs state.
-switches the VSA straps to configure 32MB or 64MB framebuffer
-controls a digipot connected to the VSA VCORE PMIC, and can set the VSA VCORE voltage between 2.5V and 3.1V in 0.1V steps
-implements bidirectional communication with the host OS via the PCIe-PCI bridge IC so that various things can be set and read
-and some other things that are not useful for this project

Wow, that is awesome!
One question though - can the fan speed be set through software somehow? I mean, it greatly varies what kind of airflow any given fan will push at a certain rpm.
Btw, I assume your card only consumes power from the PWR_SRC rail? I think I'll add 5V and 3.3V rails to my connector as well, just to facilitate making other cards easier.

The software can be modified to allow for manual fan speed, or just the parameters changed for the respective fan.
At the moment it is temperature controlled, so if the card gets hotter the fan will spin faster. I think it should work fine with any fan?
Yes, the card uses only PWR_SRC. While it also needs 5V and 3.3V rails, I decided to generate them on the board, as some laptop manufacturers don't even provide these two rails on the MXM connector.

sdz wrote on 2024-09-24, 09:51:

The software can be modified to allow for manual fan speed, or just the parameters changed for the respective fan.
At the moment it is temperature controlled, so if the card gets hotter the fan will spin faster. I think it should work fine with any fan?

Thanks! Yeah, that sounds like a good solution for sure!

sdz wrote on 2024-09-24, 09:51:

Yes, the card uses only PWR_SRC. While it also needs 5V and 3.3V rails, I decided to generate them on the board, as some laptop manufacturers don't even provide these two rails on the MXM connector.

👍

Slightly off-topic, but I thought it would be interesting nonetheless:
I finally heard back from my contact at DMP - they've gotten the silicon for the upcoming EX3 CPU back from TSMC, tested it and it looks good.
They're currently working on getting the substrate for the IC just right, supposedly it'll be in the ballpark of 4 months until the chip is done.
I also got a photo of the bare silicon die, with flip-chip balls attached and protective polymide coating removed.

Eivind wrote on 2024-09-25, 08:00:

Slightly off-topic, but I thought it would be interesting nonetheless: I finally heard back from my contact at DMP - they've got […]
Show full quote

Slightly off-topic, but I thought it would be interesting nonetheless:
I finally heard back from my contact at DMP - they've gotten the silicon for the upcoming EX3 CPU back from TSMC, tested it and it looks good.
They're currently working on getting the substrate for the IC just right, supposedly it'll be in the ballpark of 4 months until the chip is done.
I also got a photo of the bare silicon die, with flip-chip balls attached and protective polymide coating removed.

The attachment EX3-noBump.png is no longer available

Is the EX3 a faster drop in replacement for Vortex86EX used in Llama's ? 😀

janih wrote on 2024-09-25, 08:35:

Is the EX3 a faster drop in replacement for Vortex86EX used in Llama's ? 😀

Faster: yes by a ton, most likely. 1.2 GHz, 28nm-process, DDR4 RAM - compared to the EX's 300-ish MHz, 90nm, DDR3.
Drop-in replacement: no, different pin layout.

Have they told you anything more about what it actually is, @Elvind? Is it still fundamentally a similar, single-core MP6-derived chip like the the Vortex86EX (SOM) with a speed bump, fatter pipeline and/or process upgrade; or some kind of dual-core chip like the DX3 (where we will see little benefit to the second core); or some unnecessarily weird master/slave core setup like an EX2? I guess the question is, architecturally, is what DM&P is developing with the EX3 something that will be of value for a project like this? (I assume if you're continuing to ask them about it, the answer is yes). Is it closer to P6+ levels of per-clock performance than the EX? I feel like the dream is something like a Pentium M, but public info about it is staggeringly scarce.

Will come back to you on the clock control ideas - I don't think it'll really influence hardware design at this stage anyway.

jerms wrote on 2024-09-25, 08:46:

Have they told you anything more about what it actually is, @Elvind? Is it still fundamentally a similar, single-core MP6-derived chip like the the Vortex86EX (SOM) with a speed bump, fatter pipeline and/or process upgrade; or some kind of dual-core chip like the DX3 (where we will see little benefit to the second core); or some unnecessarily weird master/slave core setup like an EX2? I guess the question is, architecturally, is what DM&P is developing with the EX3 something that will be of value for a project like this? (I assume if you're continuing to ask them about it, the answer is yes). Is it closer to P6+ levels of per-clock performance than the EX? I feel like the dream is something like a Pentium M, but public info about it is staggeringly scarce.

Will come back to you on the clock control ideas - I don't think it'll really influence hardware design at this stage anyway.

jerms wrote on 2024-09-25, 08:46:

Have they told you anything more about what it actually is, @Elvind? Is it still fundamentally a similar, single-core MP6-derived chip like the the Vortex86EX (SOM) with a speed bump, fatter pipeline and/or process upgrade; or some kind of dual-core chip like the DX3 (where we will see little benefit to the second core); or some unnecessarily weird master/slave core setup like an EX2? I guess the question is, architecturally, is what DM&P is developing with the EX3 something that will be of value for a project like this? (I assume if you're continuing to ask them about it, the answer is yes). Is it closer to P6+ levels of per-clock performance than the EX? I feel like the dream is something like a Pentium M, but public info about it is staggeringly scarce.

From what I've gathered, it's a dual-core CPU in a heterogeneous setup. Meaning, each core runs completely separately from the other. They both carve out a portion of the memory, and can share data through an internal FIFO serial connection or a memory hole. Quite interesting actually, but as a simple starting point, the second core can be completely shut off.
Not quite sure about the microarchitecture, I assume it's still based on P6. My guess is that it's still got a worse IPC than Intel P6 (all the previous Vortexes have). I've asked my contact this very question, will post the reply here.

The main value of this chip for me/us will be: It's going to be way faster than the EX + will have more IO - full ISA bus, for a start (more DMA channels, yes thank you very much!)

jerms wrote on 2024-09-25, 08:46:

Will come back to you on the clock control ideas - I don't think it'll really influence hardware design at this stage anyway.

👍

Edit: thanks, Vridek - I think that spec sheet matches what I wrote above.

I made another version of the V4 M4800, derrived from V4 M4800 LVDS A01, just removed things that wouldn't make sense for this project (and reduced the cost a bit), LVDS connector, RTD scaler IC + supporting components, and the HDMI MUX (even if the scaler and LVDS connector weren't soldered, the mux IC was still needed, to mux the HDMI output to the MXM connector).

sdz wrote on 2024-09-25, 09:34:

I made another version of the V4 M4800, derrived from V4 M4800 LVDS A01, just removed things that wouldn't make sense for this project (and reduced the cost a bit), LVDS connector, RTD scaler IC + supporting components, and the HDMI MUX (even if the scaler and LVDS connector weren't soldered, the mux IC was still needed, to mux the HDMI output to the MXM connector).

Oh man, that looks bloody fantastic! 🤤
Can't wait to get my hands on it - I'll probably be stuck playing Unreal or something for days! 😆