RDNA 3 Rumor Discussion

[SuperGeil]

• Infinity Cache will play a bigger role
• RDNA 3 is looking like chiplets
• Infitiy Fabric could be key to RDNA 3

My thoughts:

1. 3 6nm or 5nm compute dies
2. 1 7nm I/O die
3. 384mb/256mb/128 Infinity Cache pools
4. 384bit or 256 bit memory bus
5. Chiplets and memory will be using AMD's infinity fabric technology
6. I/O die could also have the video processing hardware, ROP's and Ray accelerators
7. Compute dies will be mostly shaders

This chiplet setup is primed and ready for a FPGA take over in due time from an outsiders perspective looking in. Not just for RDNA but Zen as well. The Xilinks purchase is looking pretty genius.

Theoretically you could see the 7800, 7700 and 7600 all come out at the same time since they will be using the same compute chip. Smaller less complex chips that are can be scaled by mere count rather than complexity or turning off sections of the chip will have dramatic impacts on yields, time to market, power, and price.

[Gandalfthewhite]

this will be AMD going back to HBM as well. calling it now. Prices for HBM have dropped SIGNIFICANTLY in the last 2 years.

[bill dennison]

AMD is expecting the next-gen RDNA 3 GPUs scheduled for 2021 to deliver 50% improved performance-per-watt over the new RX 6000 series

https://www.notebookcheck.net/AMD-is....503412.0.html

AMD promises RDNA 3 GPUs will be as big a leap forward as Big Navi

https://technosports.co.in/2020/08/1...like-zen-cpus/

[pax]

One thing NAAF didnt mention is that one of the more consistent rumors about navi3 is that the geometry engine is overhauled and much stronger than navi2. They want to do away with what has been apparently been a bit of a weak spot on amd gpus.

So not just an obviously bigger IF cache and more shaders and RT units but some significant architectural changes. I hope they match nvidia nvenc.

[Nunz]

I hope they don't gimp the memory bandwidth so I can actually consider buying one. Come on baby, give us the goods.

If this progresses as rapidly as Ryzen did, then either RDNA3 or 4 will be the GPUs to buy. The only thing different, is that I don't see NV laying over and being lazy. The rumor of them moving back to TSMC means I expect a massive leap over Ampere..

[caveman-jim]

Originally Posted by SuperGeil

This chiplet setup is primed and ready for a FPGA take over in due time from an outsiders perspective looking in. Not just for RDNA but Zen as well. The Xilinks purchase is looking pretty genius.

What does this mean?

Typically you can't replace perf/w and outright performance sensitive ASICs with FPGAs.

There are no FPGAs running 5GHz CPU cores, or 3GHZ GPU cores.

This is why I don't comprehend the 'FPGA take over' comment.

What does FPGA take over mean, and why does a chiplet strategy prime it?

IMO the Xilinx acquisition is only for more datacenter and networking market share opportunities.

[acroig]

Happy new year Jim! Hope you and fam are doing well.

[pax]

Originally Posted by caveman-jim

What does this mean?

Typically you can't replace perf/w and outright performance sensitive ASICs with FPGAs.

There are no FPGAs running 5GHz CPU cores, or 3GHZ GPU cores.

This is why I don't comprehend the 'FPGA take over' comment.

What does FPGA take over mean, and why does a chiplet strategy prime it?

IMO the Xilinx acquisition is only for more datacenter and networking market share opportunities.

Yea the Xilinx thing was for IP Id read and the upcoming 3d chip stacking which xilinx has some knowhow for.

[NIGELG]

Originally Posted by acroig

Happy new year Jim! Hope you and fam are doing well.

Nice to know he still reads the forum and can still slam the hammer down on nonsense when it appears.

[LordHawkwind]

Originally Posted by caveman-jim

What does this mean?

Typically you can't replace perf/w and outright performance sensitive ASICs with FPGAs.

There are no FPGAs running 5GHz CPU cores, or 3GHZ GPU cores.

This is why I don't comprehend the 'FPGA take over' comment.

What does FPGA take over mean, and why does a chiplet strategy prime it?

IMO the Xilinx acquisition is only for more datacenter and networking market share opportunities.

This is what financial analyst's think of the deal.

- AMD is paying 74 times XLNX earnings, based on by far XLNX's highest earnings in the last five years.
- AMD are paying a 25% premium on the Xilinx current stock value.
- The FPGA market is not a big growth area anyway. Their CAGR is 5.9% over the last five years.
- XLNX's revenue looks lackluster.
- Neither company generates enough free cash flow to take on competitors Intel or Nvidia long term.
- AMD's rocket ship stock price has a CAGR of 95% over the last five years.

When Lisa Su announced it all she talked about was opportunities in datacentre, NICS and 5G. Nothing whatsoever to do with CPU's and GPU's. She did mention $300m synergies in the first year through cost duplication (getting rid of staff) which is nothing new.

Happy New Year Caveman nice to see you posting again.

[acroig]

Originally Posted by NIGELG

Nice to know he still reads the forum and can still slam the hammer down on nonsense when it appears.

, I miss him slamming the hammer.

[caveman-jim]

Originally Posted by pax

Yea the Xilinx thing was for IP Id read and the upcoming 3d chip stacking which xilinx has some knowhow for.

Originally Posted by LordHawkwind

This is what financial analyst's think of the deal.

- AMD is paying 74 times XLNX earnings, based on by far XLNX's highest earnings in the last five years.
- AMD are paying a 25% premium on the Xilinx current stock value.
- The FPGA market is not a big growth area anyway. Their CAGR is 5.9% over the last five years.
- XLNX's revenue looks lackluster.
- Neither company generates enough free cash flow to take on competitors Intel or Nvidia long term.
- AMD's rocket ship stock price has a CAGR of 95% over the last five years.

When Lisa Su announced it all she talked about was opportunities in datacentre, NICS and 5G. Nothing whatsoever to do with CPU's and GPU's. She did mention $300m synergies in the first year through cost duplication (getting rid of staff) which is nothing new.

Happy New Year Caveman nice to see you posting again.

I think Xilinx portfolio is strong by itself for the datacenter and enterprise networking (5G/RAN, Datacenter networking, DPU, etc.). They have strong market share and good IP. Xilinx products have a lot of synergies for market attach with AMD datacenter and enterprise platforms, plus future products bringing the IP roadmaps together.

FPGAs are in strong demand right now for lots of different applications using AI, as the models change 2x -3x faster than ASICs can be designed and produced. As the model evolution calms down domain-specific hardware (or workload-focused, at least) will be more prevalent, and will need to be more integrated than chips in sockets either side of a PCIe or similar bus; integration will have to happen both in-package, and in-SoC.

Originally Posted by NIGELG

Nice to know he still reads the forum and can still slam the hammer down on nonsense when it appears.

Originally Posted by acroig

, I miss him slamming the hammer.

Well I wasn't trying to club anyone, just asking cos I didn't get the connection between chiplets and FPGAs

[SuperGeil]

Originally Posted by LordHawkwind

This is what financial analyst's think of the deal.

- AMD is paying 74 times XLNX earnings, based on by far XLNX's highest earnings in the last five years.
- AMD are paying a 25% premium on the Xilinx current stock value.
- The FPGA market is not a big growth area anyway. Their CAGR is 5.9% over the last five years.
- XLNX's revenue looks lackluster.
- Neither company generates enough free cash flow to take on competitors Intel or Nvidia long term.
- AMD's rocket ship stock price has a CAGR of 95% over the last five years.

When Lisa Su announced it all she talked about was opportunities in datacentre, NICS and 5G. Nothing whatsoever to do with CPU's and GPU's. She did mention $300m synergies in the first year through cost duplication (getting rid of staff) which is nothing new.

Happy New Year Caveman nice to see you posting again.

My favorite person with the world's most amount of anxiety for video cards just one month ago was saying he never heard of Xilinks, now he is a market expert on the deal.

AMD did not buy Xilinks 35 billion for the short term in Data Center or 5G networking. FGPA on it's own will never pay off the deal if that was their goal. Nor did they buy them for 3D stacking chips. There is a reason why the two x86 giants bought the biggest FPGA companies.

Moores Law is dying and x86 needs a friend. The idea of offloading to an FPGA is not new, and there are tons of papers on this. FPGA's do a lot of things better than the GPU and CPU from a performance standpoint as well as a power standpoint. This is not to say they are going to replace them, but this about reinforcing them. Actually... I could probably see CDNA being an mostly an FPGA in 5 years or less. You are going to see FPGA logic integrated one way or another. Whether it's on fabric or in the logic itself.

Originally Posted by caveman-jim

Well I wasn't trying to club anyone, just asking cos I didn't get the connection between chiplets and FPGAs

I did not take offense, not sure what they are talking about. Chiplets would make it easier to transition task and engines to a custom FPGA without committing your entire architecture to it. At least in my mind. An FPGA for instance can do the ray tracing, handle complex AI, the I/O, as well as the video encoding and decoding and do it more efficiently. If RDNA 3 is chiplet with a decoupled controller die, then things would be set and primed for integration in future RDNA architecture.

reply in the Xlinks thread if you would like. : )
I'll share a cool link.

[Falck]

[pax]

https://wccftech.com/amd-files-mcm-b...o-radeon-gpus/

AMD has filed a patent for something that everyone knew would eventually happen: an MCM GPU Chiplet design. Spotted by LaFriteDavid over at Twitter and published on Freepatents.com, the document shows how AMD plans to build a GPU chiplet graphics card that is eerily reminiscent of its MCM based CPU designs. With NVIDIA working on its own MCM design with Hopper architecture, it's about time that we left monolithic GPU designs in the past and enable truly exponential performance growth.

The patent points out that one of the reasons why MCM GPUs have not been attempted in the past is due to the high latency between chiplets, programming models and it being harder to implement parallelism. AMD's patent attempts to solve all these problems by using an on-package interconnect it calls the high bandwidth passive crosslink. This would enable each GPU chiplet to communicate with the CPU directly as well as other chiplets via the passive crosslink. Each GPU would also feature its own cache. This design appears to suggest that each GPU chiplet will be a GPU in its own right and fully addressable by the operating system.

Not sure if this will be navi3 or more likely later but its interesting nonetheless.

[bill dennison]

Originally Posted by pax

https://wccftech.com/amd-files-mcm-b...o-radeon-gpus/



Not sure if this will be navi3 or more likely later but its interesting nonetheless.

more than likely and in 10 to 12 months

I'm getting more and more reasons to wait till next time every day

[Nascar24]

Originally Posted by bill dennison

more than likely and in 10 to 12 months

I'm getting more and more reasons to wait till next time every day

Stop it Billy, looks like this isnt happening for years, looks like next new chip for both NV and AMD are still old design.

“MCM design for its GPUs after RDNA3”

[bill dennison]

Originally Posted by Nascar24

Stop it Billy, looks like this isnt happening for years, looks like next new chip for both NV and AMD are still old design.

“MCM design for its GPUs after RDNA3”

keep laughing fuzzball

AMD patent points to new MCM GPU design for RDNA3

https://www.techradar.com/news/amd-p...let-technology

if RDNA 3 is MCM or not RDNA 3 will be out this year and RDNA 4 will be MCM and out next year

the sand is running out of that high wattage hourglass you got fast

[Nascar24]

Originally Posted by bill dennison

keep laughing fuzzball



https://www.techradar.com/news/amd-p...let-technology

if RDNA 3 is MCM or not RDNA 3 will be out this year and RDNA 4 will be MCM and out next year

the sand is running out of that high wattage hourglass you got fast

After the jacket changed things up I believe he knows more then us and AMD is years out as well with MCM.

Dont know what to tell ya Billy, I'm happy with my 3080!

[bill dennison]

Originally Posted by Nascar24

After the jacket changed things up I believe he knows more then us and AMD is years out as well with MCM.

Dont know what to tell ya Billy, I'm happy with my 3080!

AMD is not really MCM it's more like Ryzen

AMD patents a chiplet GPU design quite unlike Nvidia and Intel's

In the new patent dated December 31, AMD outlines a chiplet design fashioned to mimic a monolithic design as closely as possible. Their hypothetical model uses two chiplets connected by a high-speed inactive interposer called a crosslink.

A crosslink connection sits between the L2 cache and L3 cache on the memory hierarchy. Everything beneath it, such as the cores and L1 cache and L2 cache, are aware of their separation from the other chiplet. Everything above, including the L3 cache and GDDR memory, are shared between the chiplets.

This design is beneficial because it is conventional. AMD claims that compute units can access low-level cache on other chiplets almost as fast as they can access local low-level cache. Should that prove true, software won't need updating.

https://www.techspot.com/news/88138-...dia-intel.html

the jacket is at a disadvantage at this point as he can't violate all those Ryzen chiplet & crosslink patents AMD has

[Trunks0]

There are other ways to do it, really doubt patents will be a barrier for either side. That said... I see this more for cloud, server and data crunching than I do gaming from either side.

[SuperGeil]

I would be a little surprised if Nvidia has a chiplet design in the pipeline for commercial consumer use to be used in a GUI environment. Hopper is a data center card (I assume) and Lovelace is the commercial card. The problems posed by GPU chiplets are non-issues when it comes to the data center market, which is why Hopper being a chiplet does not need to worry about latency or communication with the CPU. Interesting to note, nonetheless, Hopper is delayed.

A lot of talk about ''Advanced'' node. I think this is being given too much credence for nothing. I think it's just an open-ended question of it could be 6nm, or it could be 5nm.

The way the patent reads is that there will be a single master GPU that communicates to the CPU. This to me suggest that there will be no I/O die. And the logic will and should be the same between the master GPU and slave GPUs. Basically, some redundancy. Not actually a bad thing since there will only be one chip to worry about manufacturing. Now I'm not too sure how each GPU would talk to the Infinity Cache....

[LordHawkwind]

Originally Posted by bill dennison

keep laughing fuzzball

https://www.techradar.com/news/amd-p...let-technology

if RDNA 3 is MCM or not RDNA 3 will be out this year and RDNA 4 will be MCM and out next year

the sand is running out of that high wattage hourglass you got fast

Maybe this might enlighten you regarding AMD GPU's in 2021. https://www.techspot.com/news/88121-...2021-meet.html

That doesn't even include M$ and Zen 3. RDNA 3 in 2021 in sufficient numbers is a huge ask and one AMD probably won't be able to do.

Another rumour doing the rounds is the 6800/6900XT are just halo products to prove AMD can compete in the enthusiast segment. They never had any intention of producing large numbers so mass production will be centered on 6700/XT which is where the volume and money is. Sounds plausible as why would they produce load of enthusiast cards into a market where they have zero mark share. Just think about it.

[bill dennison]

Originally Posted by LordHawkwind

Maybe this might enlighten you regarding AMD GPU's in 2021. https://www.techspot.com/news/88121-...2021-meet.html

That doesn't even include M$ and Zen 3. RDNA 3 in 2021 in sufficient numbers is a huge ask and one AMD probably won't be able to do.

Another rumour doing the rounds is the 6800/6900XT are just halo products to prove AMD can compete in the enthusiast segment. They never had any intention of producing large numbers so mass production will be centered on 6700/XT which is where the volume and money is. Sounds plausible as why would they produce load of enthusiast cards into a market where they have zero mark share. Just think about it.

if they are going to make 16.8 million to 18 million PS5 consoles in 2021
plus near the same xbox's and then their CPU's


they sure won't have the capacity to do mass production on a even bigger seller like a 6700/XT
plus a 6700/XT is most likely a failed 6800/6900XT chip cut down


........


RDNA 2 maybe a stepping stone to RDNA 3 chiplets and they may never make a lot of any RDNA 2 for lack on TSMC production
and RDNA 2 is the biggest of their chips and most likely with the lowest yields so the logical one to cut


but by Q4 the game consoles sales will be over for the most part and or slowed a lot

and chiplets are easer to make and have better yields
so i still think they will hit or be very close to the RDNA 3 roadmap in 2021

[Nascar24]

Billy I am putting on the old tin foil hat with all your assumptions! Your going off the deep end. I also believe with 80 percent certainty that RDNA3 is not going to be the chiplet or whatever you want to call it design.

[bill dennison]

Originally Posted by Nascar24

Billy I am putting on the old tin foil hat with all your assumptions! Your going off the deep end. I also believe with 80 percent certainty that RDNA3 is not going to be the chiplet or whatever you want to call it design.

bitcoin 33,640.00 United States Dollar

not going to see a 3080 all year now

[Nunz]

Originally Posted by bill dennison

bitcoin 33,640.00 United States Dollar

not going to see a 3080 all year now

Apparently you ride NV harder than AMD even though you believe AMD is focusing more on consoles than RDNA2, which means they aren't even making many cards.

[bill dennison]

Originally Posted by Nunz

Apparently you ride NV harder than AMD even though you believe AMD is focusing more on consoles than RDNA2, which means they aren't even making many cards.

NV always deserves it

but RDNA2 is crap at mining compared to the 3080

the 5700 xt is better for mining than a 6800 xt or 6900 xt from what i have read so far


....

and i'm not happy with AMD but they made the consoles deal 2+ years ago and no one could predict last year


.........

and it was NV's fault they played games with TSMC and lost

[acroig]

If Samsung can get their chit together nV stands in a better position to churn out volume by end of Q1 21. TSMC has a better node but they seem to be at capacity.

[bill dennison]

Originally Posted by acroig

If Samsung can get their chit together nV stands in a better position to churn out volume by end of Q1 21. TSMC has a better node but they seem to be at capacity.

sorry ampere would have been killer on TSMC lower watts and higher OC
and no mater what it won't on Samsung ever .

and with crapcoin at 33k+ and rising and it doing well mining Samsung will never make enough now