navi 31 Radeon RX 7900 XT PCB Layout Leaks Out: Reveals AMD's Chiplet Design, 24GB VRAM Configuration and 450W Power Specs

[BiG StroOnZ]

Summary

Igor’s Lab has illustrated the PCB for the upcoming Radeon RX 7900 XT. He says his drawing is based on real information he’s received. In order to protect his sources, it’s an artistic rendering of the PCB in crude form, but it still provides some interesting information. Igor says his “artwork” is based on real circuit boards and blueprints sent to him, but he can’t share them for obvious reasons. It looks like AMD's Radeon RX 7900 XT will feature seven chipets, 24GB of memory, and a maximum power draw of 450 watts. 

 

 

 

Quotes

Quote

The Radeon RX 7900 XT will use the codenamed Navi 31 GPU based on the RDNA 3 architecture. With RDNA 3, AMD has stated that they can deliver power efficiency (performance per watt) increases of over 50% when compared to RDNA 2. AMD's RDNA 3 GPUs will feature re-architected compute units, next-gen Infinity Cache, an optimized graphics pipeline, and enhanced ray tracing capabilities

 

We can see that AMD's RDNA 3 GPU core uses seven chiplets, six smaller chiplets, which are expected to be memory controller chiplets (MCDs), and a larger chiplet (GCD) that is expected to be AMD's compute/graphics die. All interconnected using AMD's Infinity Fabric. The GCD will be more complex than AMD's current-generation Navi 21 GPU, and deliver substantially more performance.

 

Based on the VRAM chips and chiplet design shown, AMD's RDNA 3 flagship is expected to feature a 24GB frame buffer that uses twelve 2GB (16Gb) GDDR6 modules over a 384-bit memory bus. This memory bus is divided over six chiplets, each of which controls two GDDR6 DRAM chips over a 64-bit memory bus. 

 

We can also see that AMD's alleged RX 7900 XT design uses three 8-pin (6+2-pin) PCIe power connectors, giving this GPU a maximum power draw of 450W using these three connectors (and 75W from the PCIe 4.0 x16 slot, if AMD uses Gen4). This does not rule out that AMD will use 16-pin 12VHPWR connectors, though, but thus far we have not seen any indication of that.

 

The card is equipped with 21 voltage converters, which should be divided between GPU, memory, SOC and VDDCI.

 

It is revealed that the card has four display connectors, three DisplayPorts and one HDMI. There is no information on the standard supported, but it is very likely to be DisplayPort 2.0 and HDMI 2.1 at this point. It looks like AMD is ditching USB Type-C display connectors with this release. 

 

According to the rumors, AMD RX 7900XT should feature full Navi 31 GPU with 12288 Stream Processors.

 

My thoughts

From the looks of it, AMD is not afraid of high power consumption for its next-Gen product, much like NVIDIA is; seems they are both going in the same direction in that regard. Although, senior director of marketing for AMD gaming, Sasa Marinkovic tweeted “bigger doesn’t necessarily mean better” after RTX 4090 pics leaked. AMD might be going more conservative in its design then, because even though maximum board power of the card is 450 watts, the actual TBP should be far below that. Regardless, it looks like AMD will be quite competitive again for next-Gen judging by the specs of this card. I'm interested in seeing how the chiplet design pays off, latency wise. Should be interesting to see frametime variance tests (frame pacing) coming back into play; like when they were prominent some years back. Supposedly, November is the rumored launch date for RDNA3 GPUs. Which will be right around NVIDIA's launch for Ada. So it will be interesting to see if that pans out, as if they both launch around the same time, AMD won't be late to the party.

 

Sources

https://www.overclock3d.net/news/gpu_displays/radeon_rx_7900_xt_pcb_design_leaks_reveals_amd_s_chiplet_design_vram_configuration_and_power_specs/1

https://videocardz.com/newz/amd-radeon-rx-7900xt-pcb-layout-leaks-out-with-navi-31-gpu-and-three-8-pin-power-connectors

https://www.tweaktown.com/news/88455/amd-radeon-rx-7900-xt-pcb-layout-navi-31-gpu-7-chiplets-up-to-450w/index.html

https://www.kitguru.net/components/graphic-cards/joao-silva/alleged-amd-radeon-rx-7900-xt-pcb-diagram-leaks-out/

https://www.extremetech.com/gaming/339584-radeon-rx-7900-xt-pcb-illustrated-based-on-insider-info

https://www.techpowerup.com/298895/amd-rdna3-radeon-rx-7000-flagship-gpu-pcb-sketched

https://www.tomshardware.com/news/amd-radeon-rx-7900-xt-board-design-memory-power

[HomeBoi]

Lool the name Hotpink Bonefish really appeals to me. Guess the upgrade is a 7600 for me if it don't suck

[porina]

My first through when seeing the 6 smaller chiplets around the big one was that they were the ram, missing the chips on the outside. So AMD decided to break out the memory controllers? I wonder if that's because it gives them more options on memory type e.g. HBM?

 

I guess it is also a (small) saving for lower configuration memory offerings down the road too.

[ZetZet]

7 minutes ago, porina said:

My first through when seeing the 6 smaller chiplets around the big one was that they were the ram, missing the chips on the outside. So AMD decided to break out the memory controllers? I wonder if that's because it gives them more options on memory type e.g. HBM?

 

I guess it is also a (small) saving for lower configuration memory offerings down the road too.

It gives them more margin, the GPU is on newer more expensive 5nm, memory controllers on more mature 6nm, which is cheaper. 

[porina]

2 minutes ago, ZetZet said:

It gives them more margin, the GPU is on newer more expensive 5nm, memory controllers on more mature 6nm, which is cheaper. 

That's a tradeoff between silicon cost and packaging/manufacturability complexity therefore cost. In a way not new since they've been doing it since Zen 2, just new to GPU.

[ZetZet]

Just now, porina said:

That's a tradeoff between silicon cost and packaging/manufacturability complexity therefore cost. In a way not new since they've been doing it since Zen 2, just new to GPU.

Considering both Intel and Nvidia have been trying to copy this approach I think it is working. AMD has figured out how to make infinity fabric not bottleneck their stuff so they can now use it like this. But I think they still can't do super high speed/bandwidth connections, that's why there is only one graphics die, if they figure out that too they will be using multiple smaller chips for graphics side too. 

[Tan3l6]

Where are all the PSUs thet have 12-pin power? Though more demanded by Nvidia in near future, still  12-pin seems to become a standard.

[ZetZet]

Just now, Tan3l6 said:

Where are all the PSUs thet have 12-pin power?

You can already buy some, but others are only coming soon. 

 

Silverstone HELA 850R

Seasonic Vertex series

MSI MEG AI1000P and AI1300P

Lian Li SP850

 

[TetraSky]

Looking forward to the reviews, that's for sure.

Just too bad I have to wait until 2023 for something for affordable to come out.

[porina]

3 minutes ago, ZetZet said:

Considering both Intel and Nvidia have been trying to copy this approach I think it is working.

Multiple chips on a substrate is not new, even if it is made famous by AMD CPUs since Zen 2. It's a more common thing on mobile CPUs, but my first multi-chip consumer desktop CPU was Intel Broadwell. At the end of the day it is a tradeoff of various things. AMD thinks it works best for them in some areas, that's fine, but they're not using it everywhere e.g. APUs. It's up for each company and product offering to decide what is best.

 

3 minutes ago, ZetZet said:

AMD has figured out how to make infinity fabric not bottleneck their stuff so they can now use it like this.

It is really bad on Ryzen if you do anything more serious than Cinebench. Why do you think they're throwing on ever bigger caches to help get around it? I'm still looking forward to see what they're doing with Zen 4. Can single CCD Zen 4 CPUs make full use of DDR5 bandwidth? Or is that why they're moving to iGPU as standard?

 

I see it as much less of a problem on a GPU since it is inherently more parallel. Data is more predictable than on multi-core CPUs.

 

3 minutes ago, ZetZet said:

But I think they still can't do super high speed/bandwidth connections, that's why there is only one graphics die, if they figure out that too they will be using multiple smaller chips for graphics side too. 

It takes the silicon edge or better interconnects to do that. Apple are using whatever TSMC's version of it is already, I can't remember the name right now. Again, it will come at a cost, and it has to be worth doing it for a given product design. Intel have EMIB but I think for now it is only really used on supercomputing products and upcoming server CPUs.

[StephanTW]

honestly hope its gonna be a good card, also noticed in the last couple of days that the Video Core Next in which a new column was added that was AV1 encoding, so i keep my hopes up it will have it and then i will very likely get one, once price goes down a bit

 

[dizmo]

I won't buy it, as I just see this as being a potential driver nightmare, but I'm very interested to see how it performs. I hope we actually see massive increases in performance. 

 

There's a 4k 144hz monitor in my future and I need something to drive it. 

[diegoaccord]

Where's the 6X? 

[WarbossChoppa]

1 hour ago, diegoaccord said:

Where's the 6X? 

Look at the die and you'll see the 6 spots around the "I/O" chip

[Shzzit]

Funny how no one is mentioning the 450 watts when the same people be crying about Nvidia using 450 lol.  So was all this famboyism or what?

 

Amyine know if this will be faster then 4090?  Haven’t seen any leaks,  exited to buy which ever is fastest in 4K 

[diegoaccord]

5 hours ago, WarbossChoppa said:

Look at the die and you'll see the 6 spots around the "I/O" chip

I meant GDDR6X memory.

 

4 hours ago, Shzzit said:

Funny how no one is mentioning the 450 watts when the same people be crying about Nvidia using 450 lol.  So was all this famboyism or what?

 

Amyine know if this will be faster then 4090?  Haven’t seen any leaks,  exited to buy which ever is fastest in 4K 

I think its because Nvidia moved the bar yet further, so now AMD next gen being 450, seems great compared to this gen Nvidia running the same, especially when AMD's next gen *should* be much more powerful than what Nvidia did at 450.

 

That's just a thought as to why the leeway. 

[ZetZet]

4 hours ago, Shzzit said:

Funny how no one is mentioning the 450 watts when the same people be crying about Nvidia using 450 lol.  So was all this famboyism or what?

 

Amyine know if this will be faster then 4090?  Haven’t seen any leaks,  exited to buy which ever is fastest in 4K 

450 MAX. Which means it will be 350-400 real consumption. Just like current cards. 

[leadeater]

7 hours ago, Shzzit said:

Funny how no one is mentioning the 450 watts when the same people be crying about Nvidia using 450 lol.  So was all this famboyism or what?

Nvidia was getting heat (HAH! Get it? ) because of the 600W-800W rumors, which were obviously BS but whatever, I tried to say no way not ever. Nvidia does sell GPUs with those sorts of TGP/TDP but they are Tesla compute cards not consumer air cooled cards for standard form factor cases and standard PSU's.

 

450W is fine, 450W Nvidia consumer cards have already existed from AIB's so just more of the same.

[Shzzit]

3 hours ago, ZetZet said:

450 MAX. Which means it will be 350-400 real consumption. Just like current cards. 

I’m hoping these new gpus are gonna be really good, I personally have never cared about watt usage.  But it is annoying having to upgrade my PSU also.  
 

With EVGA gone I might just go AMD this time, very sad, EVGA we’re my favorite. 
 

Also pretty crazy we might see stock 3000+ gpus and 6000mhz cpus, what a time.

 

You guys thinking a 1200 watt psu or 1600 for the new stuff coming out?  The top end GPU from either company with the 24 core intel or up coming ryzen 16 core.  

[Doobeedoo]

Interesting, I did expect a larger die. Also since it's MCM I wonder why not with latest HBM though, isn't it supposed to be cheaper quite a bit by now, lower consumption and yea faster. Solid memory buffer.

I take it with custom cards we may see even more memory, new power connector. We'll see.

[ZetZet]

12 minutes ago, Doobeedoo said:

Interesting, I did expect a larger die. Also since it's MCM I wonder why not with latest HBM though, isn't it supposed to be cheaper quite a bit by now, lower consumption and yea faster. Solid memory buffer.

I take it with custom cards we may see even more memory, new power connector. We'll see.

Cheaper than GDDR6 or GDDR6X?

[Doobeedoo]

33 minutes ago, ZetZet said:

Cheaper than GDDR6 or GDDR6X?

No, of course not, but in very start it was quite more expensive yet still used back then on older GPU and with each gen it's supposedly getting cheaper. So I'd expect maybe viable at least for flagship. 

[ZetZet]

1 hour ago, Doobeedoo said:

No, of course not, but in very start it was quite more expensive yet still used back then on older GPU and with each gen it's supposedly getting cheaper. So I'd expect maybe viable at least for flagship. 

But in this particular generation they either expected to sell to miners with insane margins or to gamers with extremely low ones, so in no scenario they would want to actually spend extra for performance. 

[Doobeedoo]

2 hours ago, ZetZet said:

But in this particular generation they either expected to sell to miners with insane margins or to gamers with extremely low ones, so in no scenario they would want to actually spend extra for performance. 

I mean, they're the only ones who know exactly how it would fit in their design and all. Just saying, they did HBM GPU before with first gen when it was new and very expensive, to save on power. Now, few games later, also improved, price was said to be lower. Especially with new high end GPU they could benefit from it.

As we're not engineers, would be great to hear some official take on HBM option. Maybe around release we may get some extra info we'll see.

[Senzelian]

The infinity cache is smaller than last gen. 

128MB 6900XT

96MB 7900XT

 

Does anyone have an idea why that could be?

I guess the wider bus probably helps.