More powerful iGPU?

[Sarra]

Is there a chance that we will start to see more powerful iGPU setups in the future? I would actually like to buy a 7900X CPU with a 7900XT or XTX integrated on the same substrate. I'd rather have two dies next to each other on the same plane, makes cooling easier. Moving the memory and VRM to the board would save a lot of space, and having the GPU and CPU in close proximity would reduce complexity in board wiring and latency.

 

I know this is a pipe dream, but if it was an actual option, I would gladly do it.

[Needfuldoer]

I'm sure someday we'll see performance equivalent to a 7900X and 7900 XTX on a single APU, but that's years away. The power and cooling requirements to run today's dies together would be astronomical.

[tkitch]

10 minutes ago, Sarra said:

Is there a chance that we will start to see more powerful iGPU setups in the future? I would actually like to buy a 7900X CPU with a 7900XT or XTX integrated on the same substrate. I'd rather have two dies next to each other on the same plane, makes cooling easier. Moving the memory and VRM to the board would save a lot of space, and having the GPU and CPU in close proximity would reduce complexity in board wiring and latency.

 

I know this is a pipe dream, but if it was an actual option, I would gladly do it.

no, we won't.

 

At least not in the timeframe that makes it usable.

 

Also, you're talking about a 2500$ CPU at that point due to size and complexity, and an additional 100$ of motherboard cost for VRMs.

 

a 200+ Watt CPU combined with a 300+ Watt GPU on a single chip would require insanely high VRM Counts on the mobo, and a lot of added cost.

 

also:  Cooling that 500+ Watt chip will *not* be easier.  

[Sarra]

2 minutes ago, Needfuldoer said:

I'm sure someday we'll see performance equivalent to a 7900X and 7900 XTX on a single APU, but that's years away. The power and cooling requirements to run today's dies together would be astronomical.

Yeah, but we have a really ugly solution for this. An AIB that puts the GPU die 90° off from the CPU, longer traces, and shared PCIE lanes, it's not ideal.

 

If I could run a larger size mainboard and have a more intelligent layout, I would gladly do it.

[Sarra]

Just now, tkitch said:

Also, you're talking about a 2500$ CPU at that point due to size and complexity, and an additional 100$ of motherboard cost for VRMs.

What if both socketed separately? Or the GPU was actually part of the board, as opposed to the CPU?

[Spotty]

The 7900X die is 71mm² (x2) and the IO die is 122mm². The 7900XTX (Navi 31 XTX) die is 529mm². The actual GPU package for the Navi 31 XTX is 40mm x 40mm which is the same as the AM5 socket size. You would need a Threadripper/Epyc sized socket just to fit both the CPU and GPU on the same substrate.

 

If the GPU is sharing system memory (DDR5) you're going to see much worse memory performance. You could add GDDR6 memory to the package or use HBM3 but that's going to make it even bigger. Forget something in the size of a CPU. With GDDR6 memory around the dies you're basically getting in to ITX motherboard size by this point.

 

While we wouldn't see this in the consumer space, and especially not on an AM5 system, the concept is possible. Nvidia is already basically doing this with their Grace Hopper CPU + GPU cards for AI.

 

 

[tkitch]

16 minutes ago, Sarra said:

What if both socketed separately? Or the GPU was actually part of the board, as opposed to the CPU?

wouldn't work for similar reasons.

 

Also, you'd need to feed 4 or 5 8 pin PCIE Cables onto the mobo.

[Sarra]

10 minutes ago, Spotty said:

The 7900X die is 71mm² (x2) and the IO die is 122mm². The 7900XTX (Navi 31 XTX) die is 529mm².

Didn't AMD go with chiplets for navi31? Why not just mingle your chiplets (don't take that out of context, please)?

 

I mean, yeah, the die size is an issue, but... I'd rather run a threadripper size package, if it meant I could focus on more sensible cooling solutions.

 

Also... What about mingling cache? You could really start doing crazy stuff if you had direct cache communication, and you had a 3D vcache.

[Agall]

20 minutes ago, Sarra said:

Didn't AMD go with chiplets for navi31? Why not just mingle your chiplets (don't take that out of context, please)?

 

I mean, yeah, the die size is an issue, but... I'd rather run a threadripper size package, if it meant I could focus on more sensible cooling solutions.

 

Also... What about mingling cache? You could really start doing crazy stuff if you had direct cache communication, and you had a 3D vcache.

I'm just waiting on the Ryzen 7040 Phoenix APUs to launch on AM5 socket.

 

3D v-cache is probably too thermally sensitive to have a +100W component nearby or on the same IHS.

 

I'm hoping AMD just makes a custom I/O die designed to handle the north bridge and RDNA3's MCD so then they can throw a single zen4 CCD and 1-2 GCDs on the same substrate, ideally in an AM5 socket.

 

I think they'd have to start making TR4 mITX/mATX boards to make TR4 useful for an APU, otherwise why not just have a dGPU if you're already going with an ATX board?

[Lurking]

1 hour ago, Sarra said:

Is there a chance that we will start to see more powerful iGPU setups in the future? I would actually like to buy a 7900X CPU with a 7900XT or XTX integrated on the same substrate. I'd rather have two dies next to each other on the same plane, makes cooling easier. Moving the memory and VRM to the board would save a lot of space, and having the GPU and CPU in close proximity would reduce complexity in board wiring and latency.

 

I know this is a pipe dream, but if it was an actual option, I would gladly do it.

Isn't that what Apple is doing with their m2 et al? I suspect their GPU is much weaker than top end dGPU these days.

 

Good idea, though. But you need many SKU to cover the needs. One person just needs simple iGPU, the other needs the top GPU, another the middle. So now that CPU needs 3 SKUs. There would have to be quite some cutting of CPU SKUs first. 

 

You also can't just upgrade the GPU. 

 

You'd need Intel to buy Nvidia or something to have top of the line Intel and AMD combos. Not sure Intel itself can do top GPU, or if Nvidia could do good CPUs. We still want at least 2 manufacturers 

 

As others said, size, cooling and power delivery are issues. To make it work, GPUs would need a significant boost in efficiency and smaller structure.

 

Lots of cache may help with the need for faster memory. I don't think an MB with DDR5 and DDR6 makes sense in practical terms. You would use the same memory for both. So either extremely costly DDR6, or slower DDR5 for both CPU and GPU.

 

Being on the same die probably had some great speed advantages. So a compromise with RAM or similar could be mitigated to some degree