LPDDR GPUs are incoming

[Marees]

AMD rumored to use lpddr5x or lpddr6 based GPUs at the low end

Why is AMD using lpddr now for dGPU (timestamped video below that raises this question)

Possible reasons

1. They are now faster — newest lpddr stuff will be as fast the gddr5 in base PS5
2. They are now not as slow as GDDR competition — they will only be half as slow as GDDR — which can be compensated by doubling the bandwidth
3. Improved caching — larger L1, L2 cache plus extra Infinity cache
4. They will not be capacity constrained — the latest & greatest gddr stuff always came with crippling capacity constraints which hopefully will not be the case with lpddr
   1. More vram for AI — professional cards can go upto 512gb vram
   2. More vram for gaming — marketing advantage like more megapixels in a camera
   3. No more excuses for low vram at budget — traditionally budget GPUs had less bus width such as 128gb (for cost & power saving) resulting in less gddr RAM. this will not be the case anymore with lpddr
5. Can be shared with laptop APU
   1. Medusa premium — shares 24 CU AT4 (10050xt) — this is a drop in replacement for strix point & will be used for all windows PC gaming handhelds, starting 2027-2028
   2. medusa halo — shares 48 CU AT3 (10060xt) — premium tablets / laptops / NUCs & edge inference use case, maybe a few $2500 handhelds too

View: https://youtube.com/watch?v=HEdT4lCvWaY&t=1500s&pp=2AHcC5ACAQ%3D%3D

 

[ChadD]

Most importantly. They don't have to worry about a competitor buying up years of supply from under them. The 7000 series got big memory pools, during design AMD assumed a price and figured it was worth it. Then Nvidia bought up so much supply that chips they expected to be paying single digits for each were costing them $20-40 each. It killed margin, and also greatly reduced the general supply as OEMS were not in a big rush to produce a ton of cards with low margins. lpddr is so plentiful that such a thing happening again is almost impossible. For the lower end and probably even mid range cards it will be plenty fast enough. 16-32gb mid range cards on lpddr why not, no real down side until you start talking about higher end cards that can saturate the extra bandwidth.

 

[ChadD]

I think the key to this working out... probably as much as the increase in the speed of lpddr. Is AMD improving their caching. I think the 9070s have shown with a good cache system you can squeeze a lot of extra performance out of slower ram. The gddr6 in the 9000s doesn't seem to hold them back compared to NVs gddr7 cards. (at least in the mid range... might have been an issue if they did try and do some 9090 version) I would assume the next gens cache will be even a little better.

 

[Marees]

I think the key to this working out... probably as much as the increase in the speed of lpddr. Is AMD improving their caching. I think the 9070s have shown with a good cache system you can squeeze a lot of extra performance out of slower ram. The gddr6 in the 9000s doesn't seem to hold them back compared to NVs gddr7 cards. (at least in the mid range... might have been an issue if they did try and do some 9090 version) I would assume the next gens cache will be even a little better.

Right. Will add that point also above — added now as the 3rd point

 

[DukenukemX]

Been saying AMD should innovate to save costs and they finally did it. Hopefully this along with chiplet GPU's will reduce the cost of graphic cards.

 

[Lakados]

Been saying AMD should innovate to save costs and they finally did it. Hopefully this along with chiplet GPU's will reduce the cost of graphic cards.

I hope this leads to a 120w AMD APU, their existing 65w units like the 8700g are damned impressive, a slightly larger one with some extra juice using what is hopefully user upgradable LPDDR would be awesome.

 

[ChadD]

I hope this leads to a 120w AMD APU, their existing 65w units like the 8700g are damned impressive, a slightly larger one with some extra juice using what is hopefully user upgradable LPDDR would be awesome.

The first company to use LPCAMM2 in a GPU wins. lol
Only half joking how awesome would that be. Ship 16gb LPCAMM2 LPDDR5x base model GPUs. Let people pop those off and replace them with 32 or 64gb mem modules.
Not realistic of course. I could see it now people buying cheap ass Lenovo oem 32gb parts that cause more issues then they are worth. We can dream.

 

[WhoBeDaPlaya]

Only half joking how awesome would that be. Ship 16gb LPCAMM2 LPDDR5x base model GPUs. Let people pop those off and replace them with 32 or 64gb mem modules.

Damn, we've come full circle...

 

[DukenukemX]

The first company to use LPCAMM2 in a GPU wins. lol
Only half joking how awesome would that be. Ship 16gb LPCAMM2 LPDDR5x base model GPUs. Let people pop those off and replace them with 32 or 64gb mem modules.
Not realistic of course. I could see it now people buying cheap ass Lenovo oem 32gb parts that cause more issues then they are worth. We can dream.

I would love to see LPCAMM2 on a graphics card. We would never see this in a commercial product, but maybe in server GPU's? Since LPDDR memory is so cheap then we could finally see GPU's with more VRAM on them. The main problem with LPDDR is that for it to reach that level of bandwidth would need a lot of connections, which means multilayered PCB's. Maybe that's why we're hearing 512-bit bus cards? Intel seems to be the one more interested in LPCAMM2 than AMD, which they hopefully stay interested.