AMD Ryzen 7000U "Phoenix" Processor iGPU Matches RTX 3060 Laptop GPU Performance: Rumor

[btarunr]

AMD is planning a massive integrated graphics performance uplift for its next-generation Ryzen 7000U mobile processors. Codenamed "Phoenix," this SoC will feature a CPU based on the "Zen 4" microarchitecture with a higher CPU core count than the Intel alternative of the time; and an iGPU based on the RDNA3 graphics architecture. AMD is planning to endow this with the right combination of a CU count and engine clocks, to result in performance that roughly matches the NVIDIA GeForce RTX 3060 Laptop GPU, a popular performance-segment discrete GPU for notebooks, according to greymon55. Other highlights of "Phoenix" include a DDR5 + LPDDR5 memory interface, and PCI-Express Gen 5. The SoC is expected to be built on the TSMC N5 (5 nm) process, and debut in 2023.



View at TechPowerUp Main Site | Source

 

[ExplodingCaps]

Just out of my curiosity, why amd just doesn't integrate gddr memory for their apu since it will be in laptop anyway?

 

[Kapone33]

That is a bold statement the 3060 Laptop is a really Good GPU.

 

[ratirt]

Oh my. 3060 in a CPU as an iGPU? That is extraordinary if true

 

[btarunr]

Just out of my curiosity, why amd just doesn't integrate gddr memory for their apu since it will be in laptop anyway?

Because it uses the same APU die even on desktop, where it will have to use standard DDR5.

 

[Vayra86]

Will it finally happen? A truly proper gaming IGP?

Instamotherfckin buy if true

And also, Intel's Alchemist is DOA definitively now.

 

[zlobby]

Sweet!

 

[Rhein7]

According to rumors it will be around 60w of rtx 3060 mobile, the the lowest wattage of that gpu but still really impressive for an APU.

 

[Richards]

Once igpu reach 3080 level performance.. desktop's and consoles will be irrelevant.

 

[zlobby]

According to rumors it will be around 60w of rtx 3060 mobile, the the lowest wattage of that gpu but still really impressive for an APU.

People don't realize that this is the greater feat here i.e., the insanely low power draw for the entire package!

 

[SL2]

Phoenix is up to 45 W, not just 15 W. 15 W is something vcardz made up in an earlier post for no reason at all. This time they say 35 - 45 W, but forgot to update their own roadmap at the end of the article.

The Phoenix APUs will operate in the 35 to 45W range, which probably means Ryzen 7000HS and 7000H series.

That is a bold statement the 3060 Laptop is a really Good GPU.

Yeah, it doesn't make sense.
CPU + IGP running at 45 W without GDDR matches 60 W discrete GPU with GDDR.

Even if AMD have somehow magically cut the power usage in half (let's say it's 30 W for the IGP) compared to the 3060, how's that still possible with the inferior RAM?

 

[Valantar]

Just out of my curiosity, why amd just doesn't integrate gddr memory for their apu since it will be in laptop anyway?

Because GDDR on its own isn't that fast - you need a lot of channels for it to be useful. LPDDR5 is comparable in per-channel speeds (though it does have 2x the channel width, FWIW), and is far better suited for use as general system memory (GDDR has horrible latency). GDDR also has many times the power draw of LPDDR. They could still of course add a few GDDR controllers to the die, but that would be expensive and difficult to implement - the die size would increase, you'd need more pins on the package, and motherboard design would become very difficult due to all the traces requred. Given that (LP)DDR and GDDR are rather different memory technologies it's also likely quite the challenge to make a hybrid controller capable of all three.

My hope for this would be for them to add a quad channel (4x2x32-bit) DDR5/LPDDR5 interface on these chips. It would cost a bit of die space, sure, and it would increase idle power draws (though that should be somewhat fixable with selective power gating etc.), but they wouldn't need to populate those channels on all packages (only those for high performance gaming applications), and it would serve double duty as adding a larger system memory pool. This would of course still increase motherboard costs and trace routing difficulty, but it still would be worth exploring IMO. Some sort of LLC for the iGPU would obviously also be very beneficial.

Still, I would absolutely love for this to be true. Bring on the 35W-45W gaming thin-and-lights! This in a ROG Flow X13 chassis would be a dream.

 

[Vya Domus]

Even if AMD have somehow magically cut the power usage in half (let's say it's 30 W for the IGP) compared to the 3060, how's that still possible with the inferior RAM?

Considering that RDNA2 is already significantly more efficient per shader than Ampere it doesn't sound that hard to believe that when you count in the improved node + whatever improvements RDNA3 has it could match a mobile 3060. And I bet this thing will have at least 32MB of dedicated cache, so the slower memory wont matter as much. I reckon they would need as little as 1536-1280 shaders to achieve this which really isn't a whole lot these days, especially on 5nm.

 

[InVasMani]

Could be using 3D Stacked Cache might have increased the amount over 5800XD. GDDR is high bandwidth, but not great on latency CPU cache is much better on latency and better bandwidth at the same time. CPU cache doesn't have navigate the PCIE bus which is a bottleneck. RNDA3 should bring optimization over RDNA2. It'll have DDR5 which is still a lot of bandwidth. There is plenty of potential ways it could beat a RTX 3060 laptop GPU at least in some scenario's.

 

[Valantar]

Considering that RDNA2 is already significantly more efficient per shader than Ampere it doesn't sound that hard to believe that when you count in the improved node + whatever improvements RDNA3 has it could match a mobile 3060. And I bet this thing will have at least 32MB of dedicated cache, so the slower memory wont matter as much. I reckon they would need as little as 1536-1280 shaders to achieve this which really isn't a whole lot these days, especially on 5nm.

Could be using 3D Stacked Cache might have increased the amount over 5800XD. GDDR is high bandwidth, but not great on latency CPU cache is much better on latency and better bandwidth at the same time. CPU cache doesn't have navigate the PCIE bus which is a bottleneck. RNDA3 should bring optimization over RDNA2. It'll have DDR5 which is still a lot of bandwidth. There is plenty of potential ways it could beat a RTX 3060 laptop GPU at least in some scenario's.

This sounds relatively reasonable, especially as AMD has all-but-confirmed that APUs won't move to a chiplet/MCM design until they can stop using through-package routing for IF connections (too much Z-height from the package for mobile use, mainly, but also too high idle power demands).

On the other hand, if this is an MCM APU that's built from the ground up for 3D stacking, we might see a CCD and GCD(?) stacked on top of an IO+cache+IF die below, which would both drastically lower the in-package trace complexity and lower IF power from the low tens of watts it can consume in current desktop systems to the milliwatt range (at least at idle). Or I guess the bottom layer might be several dice, with the cache made on an optimized node, if that makes sense economically and in terms of packaging complexity. And the 5800X3D has demonstrated that two stacked silicon dice can fit in the same Z-height as a single die, at least in some implementations.

Assuming RDNA3 makes for meaningful per-CU-per-clock performance improvements, we have that, a node improvement, packaging improvements that allow for a sizeable cache and low power interconnects - this could be feasible. I still don't think two channels of (LP)DDR5 would be sufficient ot match a 3060, but I guess we'll see how they handle that.

 

[InVasMani]

Wouldn't that depend more on resolution however in terms of matching 3060? Provided it's a lower resolution bandwidth needs are much lower.

 

[Denver]

Just out of my curiosity, why amd just doesn't integrate gddr memory for their apu since it will be in laptop anyway?

Expensive & Worst latency than DDRXxxx

 

[Valantar]

Wouldn't that depend more on resolution however in terms of matching 3060? Provided it's a lower resolution bandwidth needs are much lower.

To some degree, sure, but not enough to make up the difference between a 128-bit (LP)DDR5 bus and a 192-bit GDDR6 bus.

 

[InVasMani]

Will just have to wait and see what AMD ended up doing and how much truth there is to rumors.

 

[medi01]

Will it finally happen? A truly proper gaming IGP?

Instamotherfckin buy if true

And if priced right.

Once igpu reach 3080 level performance.. desktop's and consoles will be irrelevant.

That might be 6-8 years from now and uh oh no, you are wrong.

 

[Kapone33]

Phoenix is up to 45 W, not just 15 W. 15 W is something vcardz made up in an earlier post for no reason at all. This time they say 35 - 45 W, but forgot to update their own roadmap at the end of the article.




Yeah, it doesn't make sense.
CPU + IGP running at 45 W without GDDR matches 60 W discrete GPU with GDDR.

Even if AMD have somehow magically cut the power usage in half (let's say it's 30 W for the IGP) compared to the 3060, how's that still possible with the inferior RAM?

I am saying that I don't think it can happen. I am just shocked that an APU could have a GPU that is that fast. I have a 6500XT and cannot get over how small the GPU die is. The memory chip is about 12% larger than the actual GPU. That GPU could fit on a Ryzen chip. The wattage is where I am in line with you but efficiency is where AMD has focused it's R&D so it could happen. Just imagine something like a Chormebook with one of these but those little Handhelds (including the Steam Deck) is where AMD will make bank if it can be achieved.

Wouldn't that depend more on resolution however in terms of matching 3060? Provided it's a lower resolution bandwidth needs are much lower.

The 3060 Laptop is a proper 1080P card and ok 1440P card.

 

[ModEl4]

Another day, another AMD rumour to build hype...
It's just speculation without knowing anything concrete, graymon55 isn't sure if it is 1024SP or 1536SP by his reply the day after:

60W 3060m has 817MHz base clock, I'm sure you can find some tdp limited laptop designs that don't hit turbo frq (1.282MHz) and take the 1536SP scenario (why don't you) so Firestrike will be same or better, lol jk.
Without infinity cache addition the chances are for 1024SP imo, with infinity cache there is a chance for a 1536SP design.
Regarding desktop iteration, a 1024SP RDNA3 design without infinity cache how faster will be vs a 2.815MHz 6500XT? Probably around 5-10% depending on the frequency [taking account a 15% real IPC RDNA3 gain (20% advertised)]
If the testing is done at 720-1080p low we will see more difference due to Zen4 architecture, but if 6500XT is tested also on a zen4 (even a 6core Raphael) system with 1080p high settings then the results are going to be entirely different...

 

[zlobby]

To some degree, sure, but not enough to make up the difference between a 128-bit (LP)DDR5 bus and a 192-bit GDDR6 bus.

This new APU would greatly benefit from HBM but sadly, it's quite expensive.

 

[defaultluser]

it's only possible if hey add-on the same amount of Infinity Cache + shaders as a real 6500 RDNA3 replacement

 

[InVasMani]

720p/1080p seems modestly realistic for laptop if they want to insert these into cheaper more affordable options. In that scenario I can see them getting close to RTX 3060 especially in the case of 720p where some of the technology AMD can leverage will be leveraged to a higher extent. If AMD were to use 3D stacked cache not all performs has to derrive from the GPU itself for FPS uplift.