News Framework moves into desktops, 2-in-1 laptops at Second Gen event

[usertests]

48GB and 96GB should be possible, but the highest density 32-bit LPDDR5X packages are 128Gb (Samsung and Micron at least, SK Hynix has horrible public part data but I'd be surprised if they were higher) which makes the maximum capacity 128GB.

Ok, I think I was missing that single packages for smartphones are 64-bit, and those are usually going to have the largest capacity on the market before anything else.

The main downside is if you want a dGPU. There, it seems the most they could do is PCIe 4.0 x8. That's still enough for mid-range GPUs.

I think it will be rare that anyone would want to pair a dGPU with Strix Halo. You might want to do it when retiring the board from AI purposes years after getting it, or to enable some feature not provided by the 8060S iGPU (like DLSS or pre-50 series PhysX acceleration). Otherwise, you have a roughly 7600 XT GPU with "unlimited" VRAM to play around with.

 

[theverge]

$1999, no discreet GPU, no storage, no OS license.

$1999?

 

[usertests]

$1999, no discreet GPU, no storage, no OS license.

$1999?

Start adding parts together and see what your build costs. You need to add a 16-core Ryzen 9 9950X, 128 GB of DDR5 RAM, maybe an RX 7600 XT to approximate the iGPU in gaming, and other parts.

You will probably end up below $1999, but that's not really the point since most people will be getting this to outperform an RTX 4090/5090 in VRAM-limited AI workloads. Those GPUs are currently going for over $2k for just the GPU alone.

 

[bit_user]

Start adding parts together and see what your build costs. You need to add a 16-core Ryzen 9 9950X, 128 GB of DDR5 RAM, maybe an RX 7600 XT to approximate the iGPU in gaming, and other parts.

You will probably end up below $1999, but that's not really the point since most people will be getting this to outperform an RTX 4090/5090 in VRAM-limited AI workloads. Those GPUs are currently going for over $2k for just the GPU alone.

Yeah, if you add case, motherboard, cooler, CPU, dGPU, and memory, it probably comes to about $1600. As you said, that still missing the NPU, has only like half the memory bandwidth, and the GPU can't access main memory like this one can. It also can't be quite as small as this is.

It's a specialty item and those command a premium.

 

[heffeque]

Pre-ordered the 395 with 128 GB version.

It has A LOT of punch for such a small PC.

It's going to be extremely quiet with the single 12cm Noctua fan, extremely low power consumption (saw a reviewer playing a not so old 3D game pushing 60 fps using just 7W in total, so it's extremely efficient), and it uses standard parts, so if the PSU breaks, I can easily replace it. The fan? That's easy to replace too. Etc.

I need it small enough to fit behind my TV (which is the case), silent enough for HTPC (which is also the case), and fast enough for power use (CPU-wise it's a beast; GPU-wise it's very decent and efficient; and 128 GB of RAM should keep it well fed for some time).

It fits my requirements to the detail!

 

[usertests]

As you said, that still missing the NPU, has only like half the memory bandwidth, and the GPU can't access main memory like this one can. It also can't be quite as small as this is.

I'll assign a value of $0 to the NPU because it's there for some edge cases where not much perf is needed but it can do the operations with better efficiency (typically in a laptop). Everyone is going to be using the higher performing 32-40 CUs for LLMs. Maybe both can be leveraged simultaneously for a small performance boost (I think Intel was trying to push this angle with Meteor Lake) but I don't know about it.

Other than that, yeh

 

[theverge]

Start adding parts together and see what your build costs. You need to add a 16-core Ryzen 9 9950X, 128 GB of DDR5 RAM, maybe an RX 7600 XT to approximate the iGPU in gaming, and other parts.

You will probably end up below $1999, but that's not really the point since most people will be getting this to outperform an RTX 4090/5090 in VRAM-limited AI workloads. Those GPUs are currently going for over $2k for just the GPU alone.

NGL, saw Ryzen AI Max+ 395 and my brain just skipped right to the price of 2k. Didn't see 128gb of memory, i'm sure that adds quite a bit. 700 or so on the asus laptop.

 

[bit_user]

I'll assign a value of $0 to the NPU because it's there for some edge cases where not much perf is needed but it can do the operations with better efficiency (typically in a laptop). Everyone is going to be using the higher performing 32-40 CUs for LLMs. Maybe both can be leveraged simultaneously for a small performance boost (I think Intel was trying to push this angle with Meteor Lake) but I don't know about it.

They said just the NPU can do 50 TOPS. I didn't see a spec for the RDNA 3.5 iGPU, but it's probably not much more than that if you consider that the RX 7900 XTX desktop GPU is only good for 123 TOPS.

 

[SonoraTechnical]

I'd be interested in just the Motherboard w/ the Ryzen AI Max 395+ APU and 128GB of LPDDR5X 8000 Memory. The board supports two DP ports (I'm interested in that) and then the spare HDMI, two M2.NVMe slots (I imagine they are both PCI4)?

 

[bit_user]

two M2.NVMe slots (I imagine they are both PCI4)?

Yes, the SoC only has PCIe 4.0.