Discussion ARM Cortex/Neoverse IP + SoCs (no custom cores) Discussion

[soresu]

This looks like an interesting board A720: https://www.cnx-software.com/2024/1...core-armv9-soc-with-a-30-tops-ai-accelerator/

Geez, it's even got an x16 slot, albeit only carrying x8 lanes.

With the m.2 slot already on the board that's potentially 3x m.2 nvme at full PCIe v4 bandwidth, perfect for an upgradable media PC.

 

[DZero]

This looks like an interesting board A720: https://www.cnx-software.com/2024/1...core-armv9-soc-with-a-30-tops-ai-accelerator/

HOLY CRAP!

• 8GB RAM for $199
• 16GB RAM for $239
• 32GB RAM for $299
• 64GB RAM for $449

16GB is already a sweet spot. With the correct drivers, we are witnessing a high tier Raspberry Pi for the gamers.

Now, imagine it being the base of a USD 399 gaming handheld...

 

[DrMrLordX]

Very interesting! It even has "medium" A720s. It'd be a great dev board for code profiling on ARMv9 systems.

 

[Shivansps]

Here a pdf with a lot more details.

https://dl.radxa.com/orion/o6/radxa_orion_o6_announcement_presentation_2024_12_18.pdf

The first consumer ARM board that is not based on a TV Box SoC... it was about time. Also over time prices for boards with this soc should go down a bit, because the pricing here without the discount it is a little bit high.

Something to keep in mind is that power and TDP are two things that are being keep hidden.

Btw, i already pre-ordered one.

 

[DZero]

WoA comming soon is a nice touch for that board. Now I am interested on how Microsoft will open their drivers. And who will benefit from that.

 

[Shivansps]

Microsoft dosent need to do anything, it all depends on the EDK2 firmware on the board to be able to boot it up. And Windows generic drivers should be enoght to handle storage IF the pcie controller on the soc is up to spec. The rest of the drivers arent Microsoft problem.
There are know problems with RK3588 boards and M.2 disks to the point you need a specific custom driver for them to work on WoA, but thats only because Rockchip PCIE is not compilant.

This is the first time a company pledges to develop the EDK2 and support WoA... i trust Radxa a lot more than other Chinese companies that launches one SBC, release 1 OS version, never to be updated again.

Now ARM really need to work on Windows Mali drivers, they cant expect 3rd party companies to do it.

 

[MS_AT]

https://newsroom.arm.com/blog/arm-nvidia-project-digits-high-performance-ai interesting I thought that would be using V2 cores for Grace in Digits, but it seems they are using Grace as an umbrella name seeing here it is using 10 x925 and 10x A725

 

[igor_kavinski]

https://newsroom.arm.com/blog/arm-nvidia-project-digits-high-performance-ai

Don't you think it's basically useless starting at $3000 and them citing FP4 performance claims (not even bfloat)?

 

[MS_AT]

Don't you think it's basically useless starting at $3000 and them citing FP4 performance claims (not even bfloat)?

I guess it depends on the use case, but for inference the Mem BW will become the perf limit sooner than TOPS, in my experience, and it's cheaper than M4 PRO macbook with 128GB of RAM, though it is not a laptop. They of course did not care enough to put the numbers on MemBW, but I guess something on Strix HALO level or maybe better, but with full CUDA support so without having to deal with AMD software stack.

 

[soresu]

https://newsroom.arm.com/blog/arm-nvidia-project-digits-high-performance-ai interesting I thought that would be using V2 cores for Grace in Digits, but it seems they are using Grace as an umbrella name seeing here it is using 10 x925 and 10x A725

If it is targeting consumer use it wouldn't have Neoverse IP which has extra stuff for servers/datacenters that 99.9999% of the target market won't use.

Also I suspect that the licensing on Neoverse is pricier than Cortex to boot.

 

[soresu]

There are know problems with RK3588 boards and M.2 disks to the point you need a specific custom driver for them to work on WoA, but thats only because Rockchip PCIE is not compilant.

Hopefully that will be fixed + PC-BSA compliance for RK3688 generation of boards.

 

[MaksDampf]

They of course did not care enough to put the numbers on MemBW, but I guess something on Strix HALO level or maybe better, but with full CUDA support so without having to deal with AMD software stack.

I can see two rows of 4 memory packages each in their rendering. So it is likely a 256bit interface just like Strix Halo and M4 pro. So if it is clocked at 8533 like M4 series, its 276gb/s.

Thats the same level as a 4060.

They might clock the ram a little lower though as it is a power optimized part. The chassis still looks like a pure mockup that is far from final as there is no cooling solution and the metal cover has unrealistic sharp bends that you would't have with real sheet metal. The metal mesh and front to back Airflow that DGX has does not make any sense for a product this small. I'd prefer if they rather make the top plate a mesh and put a big, quiet fan and heatsink underneath.

 

[Shivansps]

Hopefully that will be fixed + PC-BSA compliance for RK3688 generation of boards.

ARM released in November the "Arm PC Base System Architecture 1.0" it is intended to solves all these issues so ARM SoCs can be made in order to make ARM pcs as it includes the design guidelines.

Documentation – Arm Developer

developer.arm.com

This was very very late into the game, they should have made this years ago. Now is up to the companies to follow it. Thats a start so ARM SoCs can work just like x86 cpus do.

NOW ARM needs to go out and make a windows mali driver, they cant really expect 3rd party companies to do it.

 

[soresu]

This was very very late into the game, they should have made this years ago

Ye, especially weird because the server equivalent SBSA is a decade old now.

 

[Panino Manino]

A bit off topic.
"People" are going crazy with the rumored Switch 2 specs, and I wanted to ask, how good was the A78 core, really? How strong is a 1GHz A78 compared with AMD CPUs, specially Zen 2?

I always had this understanding that ARM designs were weak compared with desktop x86 CPUs, that things only started to improve with the X1 generation. Was I wrong? The A series was stronger than I believed?

 

[LightningZ71]

The A78 was OK, but nothing to write home about. It IS significantly better than the existing Switch SOC CPU cores, so there will be an improvement that you can feel.

 

[Nothingness]

There are some A78 SPEC scores (among other CPUs, both Arm and x86) here: https://blog.hjc.im/spec-cpu-2017

 

[Shivansps]

Some more news on the Radxa Orion O6..

It was been showed off booting regular linux arm64 iso linux installers with a RX 6400 GPU with the pre-release firmware.

https://twitter.com/x/status/1879023500196991392

Some more info on a sbc-bench run...

https://0x0.st/8opg.bin

Looks like all 8 A720 cores have 512KB L2 so the only difference from high to mid is frecuency

 

[Shivansps]

A bit off topic.
"People" are going crazy with the rumored Switch 2 specs, and I wanted to ask, how good was the A78 core, really? How strong is a 1GHz A78 compared with AMD CPUs, specially Zen 2?

I always had this understanding that ARM designs were weak compared with desktop x86 CPUs, that things only started to improve with the X1 generation. Was I wrong? The A series was stronger than I believed?

Its not bad, but its not great either, its a 5 year old core at this point...

 

[NostaSeronx]

"People" are going crazy with the rumored Switch 2 specs, and I wanted to ask, how good was the A78 core, really? How strong is a 1GHz A78 compared with AMD CPUs, specially Zen 2?

Its not bad, but its not great either, its a 5 year old core at this point...

Cortex-A78 is a bit more efficient than Cortex-A76. Cortex-A76 beats Zen in TDP-locked regions.

Cortex-A57/NS1 < 4x Jaguar @ 1 GHz
Cortex-A78/NS2 > 8x Zen @ 1 GHz

For what it is, it probably is great. If we get to release and it is sub-3 watt it would be especially great.

Trend of the switches.
Switch 1 v0 = 15w
Switch 1 v1 = 7w
Switch 1 v2 = 5w
Switch 2 v0 =

 

[Meteor Late]

From what I see, the main limitation in terms of battery life is the battery capacity more than anything else, from rumors and speculation, it seems Switch 2 will have a 20 Wh battery, that's absolutely terrible and merely smartphone-like.
i understand Switch wants to have much lower weight than other handhelds, but there is a lot of margin to have at least 30-35 Wh or so in a 400-500g weight range.

 

[Panino Manino]

From what I see, the main limitation in terms of battery life is the battery capacity more than anything else, from rumors and speculation, it seems Switch 2 will have a 20 Wh battery, that's absolutely terrible and merely smartphone-like.
i understand Switch wants to have much lower weight than other handhelds, but there is a lot of margin to have at least 30-35 Wh or so in a 400-500g weight range.

Nintendo is really yakuza of the gaming world.
Like they did with the antecessor I believe that in a few years will come a firmware update that will almost double the clock that will allows them to keep selling the same old hardware at the same price.
Maybe there'll be a new model with the battery capacity it always needed, for a higher price.

 

[Shivansps]

Cortex-A78 is a bit more efficient than Cortex-A76. Cortex-A76 beats Zen in TDP-locked regions.

Cortex-A57/NS1 < 4x Jaguar @ 1 GHz
Cortex-A78/NS2 > 8x Zen @ 1 GHz

For what it is, it probably is great. If we get to release and it is sub-3 watt it would be especially great.

Trend of the switches.
Switch 1 v0 = 15w
Switch 1 v1 = 7w
Switch 1 v2 = 5w
Switch 2 v0 =

Im not completely convinced of a 5 year old A78 core for a 8 year minimum product that many companies will try to launch their games to.

 

[LightningZ71]

I think it isn't appreciated how most modern games aren't heavily CPU compute limited these days. With four performant threads, this device should be well balanced with respect to it's specs and planned games.

 

[MS_AT]

I think it isn't appreciated how most modern games aren't heavily CPU compute limited these days. With four performant threads, this device should be well balanced with respect to it's specs and planned games.

Well, it will hopefully be able to run original Switch 1st party games without framedrops

Switch 2 is rumoured to have 8 cores, each clocked at 1 GHz in mobile mode. I don't consider that performant. Of course 1T performance will exceed 1T performance of A57 clocked at 1 GHz (I am not sure what are the rules that allow Switch 1 to boost to 1.8GHz), but I don't see how it will give them a room to do something more with certain games. They could have done 4 X2 cores at 2GHz each, wouldn't that be a better balance? Or even boost those A78 cores to smarthphone level clocks. I mean 2 GHz.

Of course rumours might be wrong. Maybe there is some sort of dynamic boost that we don't know about, that is able to cover the low threaded situation, hard to say.