Discussion RDNA 5 / UDNA (CDNA Next) speculation

[Magras00]

And another consideration could be Apple. The two Medusa Halo versions (12C / 24C) together with AT3 and AT4 could challenge M5/M6 Pro & Max.

Very interesting idea. Assuming there's a small (1x video encode/decode), standard (2x <) and ultra (4x <<) sized MID AMD will have unprecedented flexibility and could counter Apple quite effectively without wasting silicon on the rest of the SKUs.

Regarding ML/AI, AT3 and AT4 with LPDDR6 would operate in a similar bandwidth range per tensor/matrices TFLOPS like an RTX 5090. Crazy, if you think about that. So AT3 and AT4 could be quite decent "accelerator cards" for ML/AI. Together with the huge VRAM capacity you could even outmatch RTX Pro 6000 with 96 GByte or GB202 successor with max. 192 GByte for some tasks (if 32 Gbit GDDR7 modules are available), despite featuring much lower raw TFLOPS numbers. And because the chips are small and use cheap memory: Much better bang for the buck. And having the possibility to pair that in APU style with a CPU is just the cherry on top.

I see a strategy there...

Edit:
One funny thought I had now was US export restrictions to China. Because AT3 and AT4 are much below the bandwith and TFLOPS limits of those restrictions, they could potentially sell well in China.

Hasn't AMD already shown that to some extent (IIRC 12-32GB NV cards) with Strix Halo? But sure the nextgen could extend that lead even further with 4x memory + much stronger HW: An upgraded AT3 pro market card with 9070 level FP8 AI tops + FP4 but proper version akin to NVFP4, softmax FF HW and lowest ML hanging fruits from CDNA + 256-512GB would probably destroy even the RTX Pro 6000, perhaps even nextgen titan. Having tons of TFLOPS is pointless if you run out of VRAM.

RTX PRO 6000 already uses 24Gb x 16 clamshell. 32Gb GDDR7 would only increase maximum capacity to 128GB for a Vera Rubin based RTX Pro card. There's 48Gb and 64Gb GDDR7 in the spec, but that's so far out that LPDDR by then would have moved the goalpost yet again. NVIDIA really needs to counter this effectively but perhaps the upcoming NV1 and NV1X are a part of this.

Sure higher GB/$ at iso-GM. Those 24Gb GDDR ICs aren't cheap and 32Gb likely won't be any cheaper nextgen.

Would be interesting if someone could find somewhat accurate figures, but AFAIK rn LPDDRx is half or less $/GB than GDDRx. If it's as low as a third of the cost (vs standard 16Gb GDDR) then that's very disruptive.

Summarizing this seems like at least a three pronged strategy by AMD:

1. Counter future Apple Mx with MID flexibility for content creation + increased memory capacity for LLM
2. Counter NVIDIA and disrupt LLM market with very high memory capacity pro cards
3. Evade US export restrictions to China on HW for LLMs

 

[basix]

192 GByte for next RTX Pro card were wrong, correct. 128 GByte are right (if 32 Gbit GDDR7 chips even exist).

If I may add 4.) to your list:
-> 4. Great APUs for Rendering/EDA/ML/AI workstations for Windows (not as direct Apple competitor but for existing Windows applications and use cases which Apple devices do not support) as well as Mobile Gaming

Regarding your point 2.):
Nvidia might jump on the LPDDR6 bandwagon as well. But AMD has a more stringent use case for that due to APUs.

And true, Nvidias NV1X is also in the wilds with similar features and target markets like Medusa Halo. But Windows with x86/x64 is far more mature than Windows on ARM.

Edit:
AT3 should be doable with 150W (if power gets limited). This would then allow for single slot server GPUs (passive cooled). Imagine deploying 16x of them with 512 GByte each in a single server blade. 16 times x16 PCIe 5.0 would also match well with 2x EPYC CPUs. Quite compact, powerful and decent cost for such a system. Each card could additionally feature a local SSD for KV-Cache (inferencing). It just requires x4 PCIe 5.0 lanes in addition to the existing x16 ones on the MID.

Napkin math predicts FP4 throughput and memory bandwidth of all 16 cards combined to be on a similar level like a single Nvidia B200 or AMD MI350X. Power draw will be higher (~2x) and you cannot build NVLink alike GPU-to-GPU networks. But having up to 8 TByte of total VRAM with much lower cost might compensate that for some use cases (inferencing only).

 

[adroc_thurston]

1. Counter NVIDIA and disrupt LLM market with very high memory capacity pro cards
2. Evade US export restrictions to China on HW for LLMs

dead market

 

[poke01]

Am I missing something or is AMD using LPDDR6 for RDNA5 dGPUs?

 

[adroc_thurston]

Am I missing something

no.

using LPDDR6 for RDNA5 dGPUs?

For baby ones indeed.
They're trading shoreline bandwidth density for capacity and commonality with premium APU configs.

 

[marees]

Am I missing something or is AMD using LPDDR6 for RDNA5 dGPUs?

The replacements for strix point z2e (medusa premium z3e) & strix halo (medusa halo) are rumored to share chiplets with entry level & budget RDNA 5 GPUs

Basically the 10060xt & 10050xt will now have lpddr5x or lpddr6 vram instead of gddr6/gddr7

 

[adroc_thurston]

The replacements for strix point z2e (medusa premium z3e

They're premium laptop parts.
Handhelds are a tiny irrelevant market not a single soul cares about.

 

[ToTTenTranz]

Handhelds are a tiny irrelevant market not a single soul cares about.

2.3 million "irrelevant single souls" estimated for 2025, a 32% increase from 2024.

https://www.newsnationnow.com/business/tech/pc-handheld-sales-2025

This is before Microsoft launches its Win11 "gaming version" UI, which is only coming to one model in Q3 but will become widespread in 2026.

 

[branch_suggestion]

2.3 million "irrelevant single souls" estimated for 2025, a 32% increase from 2024.

https://www.newsnationnow.com/business/tech/pc-handheld-sales-2025

This is before Microsoft launches its Win11 "gaming version" UI, which is only coming to one model in Q3 but will become widespread in 2026.

PC market is ~250Mu, handhelds also have below average ASPs/GMs.
It is <1% of the market, something AMD only bothers with because they can get away with the same silicon and they have the market captive.

 

[marees]

2.3 million "irrelevant single souls" estimated for 2025, a 32% increase from 2024.

The soul that matters doesn't care about handhelds !? 🤪

 

[marees]

PC market is ~250Mu, handhelds also have below average ASPs/GMs.
It is <1% of the market, something AMD only bothers with because they can get away with the same silicon and they have the market captive.

Going forward xbox will not have a series S version

Instead there will be a first party box console plus 3rd party 'xbox' handhelds + steam deck (not counting PS handheld & Switch as they are custom hardware)

So effectively windows handhelds will play the role of series S

That is a non-insignificant volume

 

[soresu]

+ steam deck

So effectively windows handhelds

Steam Deck is not and was never intended to be a Windows handheld, even if it is possible to operate Windows on it.

They didn't launch it until Wine/Proton/DXVK and Linux AMD gfx drivers were in a state of relative maturity precisely because they were aiming to displace the WIndows bottleneck.

 

[marees]

Steam Deck is not and was never intended to be a Windows handheld, even if it is possible to operate Windows on it.

They didn't launch it until Wine/Proton/DXVK and Linux AMD gfx drivers were in a state of relative maturity precisely because they were aiming to displace the WIndows bottleneck.

Agree. Steam Deck is anti-windows

That gives great clarity to xbox, & Microsoft
Their competitor is not Sony anymore. It is Valve

It is a do or die moment for Xbox & they know it
They need to beat valve with the windows handhelds (which replace series S)

 

[soresu]

Their competitor is not Sony anymore. It is Valve

I mean, it's still Sony too, they just know (or should know) they don't have a chance based on current and past generations.

X360/PS3 gen was the closest they came to coming on top, and IMHO that was mainly down to just how much of a ballache it was to develop for the Cell architecture and the initial cost of the console.

 

[Tigerick]

Guys, I have explained the usage of AT3 and AT4 in LPDDR6 thread, we could forget both dies to be used as discreet GPU due to inclusion of IOD and NPU. Please continue with AT0 and AT2. And I have listed possible RDNA5 lineup with AT2 below with the help of MLID:

 

[marees]

Guys, I have explained the usage of AT3 and AT4 in LPDDR6 thread, we could forget both dies to be used as discreet GPU due to inclusion of IOD and NPU. Please continue with AT0 and AT2. And I have listed possible RDNA5 lineup with AT2 below with the help of MLID:

View attachment 129159

The last bit 9050xt 8gb won't sell. AMD is not in a position to sell 8gb cards. They could just reduce the price on 9060xt 16gb to below $280 in 2027

7400xt will continue because it is board powered & 6nm which is a discounted process node

 

[Tigerick]

The last bit 9050xt 8gb won't sell. AMD is not in a position to sell 8gb cards. They could just reduce the price on 9060xt 16gb to below $280 in 2027

7400xt will continue because it is board powered & 6nm which is a discounted process node

Yeah, AMD could just bundle 16GB for the card above 7400 w/8GB by 2027. OR BOTH.

 

[ToTTenTranz]

PC market is ~250Mu, handhelds also have below average ASPs/GMs.

This is like saying I shouldn't open a new pizza place in a crowded town without any other italian restaurants because there are plenty of vegan and meat restaurants in that zone.

It's a new, rising market that hardly overlaps with gaming PCs, let alone all the office/work PCs that take up most those 250M.
Companies setting the expectation of infinite growth towards investors will be going after new markets.

 

[gdansk]

This is like saying I shouldn't open a new pizza place in a crowded town without any other italian restaurants because there are plenty of vegan and meat restaurants in that zone.

No, analogies are stupid. Big GPU is going to be expensive even if you're throttling it so hard to doesn't look like it should be expensive. Expensive gaming device that doesn't have expensive gaming device performance is niche. Never not been the case.

Note Omdia describe the Steam Deck as being the PC handheld market leader yet completely eclipsed by Nintendo.
Note that they project in 2029 sales of still only 4.7 million. How many of these will be at 300mm+ GPU die size prices? Little Aerith sized chips make up the bulk of this volume today. Why would that change?

Note that by way of comparison Omdia describes cloud gaming chips as a $21 billion dollar market by 2029. One wonders. Maybe AMD did venture to gamble on AT0 themselves. But the gaming handheld chips so far have had some company behind them or are recycled laptop parts.

 

[adroc_thurston]

That is a non-insignificant volume

It does not exist.
Their is only one 'real' handheld and its name is Nintendo Switch.

It's a new, rising market

The sum total of all PC handheld sales yearly is a mere fraction of Switch 1/2.
Handhelds aren't new, they're old.

 

[Joe NYC]

Guys, I have explained the usage of AT3 and AT4 in LPDDR6 thread, we could forget both dies to be used as discreet GPU due to inclusion of IOD and NPU. Please continue with AT0 and AT2. And I have listed possible RDNA5 lineup with AT2 below with the help of MLID:

View attachment 129159

I disagree. There is a base tile that is basically a monolithic die and it has everything SOC/IOD: IO, LP cores, and Zen 6 cores. I think every Zen 6 client CPU will have it as a starting die.

And then, on one side, 12 core CPU die can be optionally attached
On the other side, either small (AT3) or a big (AT4) die can be attached.

As far as memory access, the small GPU die in Medusa Halo Mini will just get it from SOC.
The large, full Medusa Halo GPU die will get 2 channel worth from SOC and 2 channels of its own.

 

[adroc_thurston]

As far as memory access, the small GPU die in Medusa Halo Mini will just get it from SOC.
The large, full Medusa Halo GPU die will get 2 channel worth from SOC and 2 channels of its own.

oh what the hell. NO.

 

[Kepler_L2]

I disagree. There is a base tile that is basically a monolithic die and it has everything SOC/IOD: IO, LP cores, and Zen 6 cores. I think every Zen 6 client CPU will have it as a starting die.

And then, on one side, 12 core CPU die can be optionally attached
On the other side, either small (AT3) or a big (AT4) die can be attached.

As far as memory access, the small GPU die in Medusa Halo Mini will just get it from SOC.
The large, full Medusa Halo GPU die will get 2 channel worth from SOC and 2 channels of its own.

No, the memory interface is on the GPU GMD

 

[Joe NYC]

No, the memory interface is on the GPU GMD

Do you mean that the Medusa (both versions) will get its memory access for CPU from the GPU die?

It does not seem ideal for low power CPU operation, but then, GPU die still has to be working even in low power mode.

This would then allow a smaller, more limited SoC / IOD for Medusa Halo. No memory access, no GPU, no NPU. So there would have to be separate SoC / IOD dies between Halo and regular Medusa. Or just same one but using only fraction of the die functionality.

 

[Kepler_L2]

Do you mean that the Medusa (both versions) will get its memory access for CPU from the GPU die?

It does not seem ideal for low power CPU operation, but then, GPU die still has to be working even in low power mode.

This would then allow a smaller, more limited SoC / IOD for Medusa Halo. No memory access, no GPU, no NPU. So there would have to be separate SoC / IOD dies between Halo and regular Medusa. Or just same one but using only fraction of the die functionality.

There's Medusa Halo with SoC (CPU + I/O) and AT3 GMD (GPU + Memory), Medusa Premium with smaller SoC (CPU + I/O) and AT4 GMD (GPU + Memory) and Medusa Point with SoC (CPU + GPU + Memory + I/O) plus optional CCD.