Discussion RDNA4 + CDNA3 Architectures Thread

[DisEnchantment]

With the GFX940 patches in full swing since first week of March, it is looking like MI300 is not far in the distant future!
Usually AMD takes around 3Qs to get the support in LLVM and amdgpu. Lately, since RDNA2 the window they push to add support for new devices is much reduced to prevent leaks.
But looking at the flurry of code in LLVM, it is a lot of commits. Maybe because US Govt is starting to prepare the SW environment for El Capitan (Maybe to avoid slow bring up situation like Frontier for example)

See here for the GFX940 specific commits

History for llvm/lib/Target/AMDGPU - llvm/llvm-project

The LLVM Project is a collection of modular and reusable compiler and toolchain technologies. - History for llvm/lib/Target/AMDGPU - llvm/llvm-project

github.com

Or Phoronix

More AMD "GFX940" Enablement Work Landing In LLVM - Phoronix

www.phoronix.com

There is a lot more if you know whom to follow in LLVM review chains (before getting merged to github), but I am not going to link AMD employees.

I am starting to think MI300 will launch around the same time like Hopper probably only a couple of months later!
Although I believe Hopper had problems not having a host CPU capable of doing PCIe 5 in the very near future therefore it might have gotten pushed back a bit until SPR and Genoa arrives later in 2022.
If PVC slips again I believe MI300 could launch before it

This is nuts, MI100/200/300 cadence is impressive.



Previous thread on CDNA2 and RDNA3 here

Question - Speculation: RDNA3 + CDNA2 Architectures Thread

Man I have been dying to make this one for a while now. First rumours for RDNA3 are here so new thread time! Just going to start off with this one for now: kopite7kimi on Twitter: "@VideoCardz Ah, I mean a simple mcm design with 10240 cores is not enough. Because the lift from RDNA2 to RDNA3...

forums.anandtech.com

 

[xpea]

cleaned up this mess

esquared

 

[Hitman928]

Nah.

nope. being used by Nvidia for more than 4 years on cadence and Synopsys

I’ll reach out to my support contacts this week and ask. Last I talked to them, none of these tools were in production.

 

[adroc_thurston]

cleaned up the mess


esquared

 

[xpea]

cleaned up the mess



esquared

 

[Hitman928]

@xpea do you know which tools support GPU acceleration so I know which ones to ask about?

Edit: I looked up in the support docs, at least for Cadence, their transient spice simulator has GPU accelerated support now but it's fairly recent (a little over a year). I'm surprised as I was told right around that time that it wasn't ready yet, but I also use other tools that don't support GPUs yet so maybe that's why I was told that. I expect it will be at least a few years before most companies start to incorporate it as even just software versions tend to lag due to in house validation time and stability being more important and those don't require hardware changes. You were right though in that there are tools that have started to support GPUs. I'd be interested in seeing many core vs. GPU comparisons as the ones provided in the doc were only small core counts versus GPU.

 

[adroc_thurston]

cleaned up the mess


esquared

 

[Abwx]

For ignorants like you, Nvidia used their H100 cluster to design Blackwell family with their in house Cadence and Synopsys pluggins. They even publish a white paper on it. You are so last decade...

IC design is not rendering, eventualy GPUs can be used for VHDL since that s simulations at the functions blocks level but for anything SPICE i can guarantee
you that Nvidia has a full building of CPUs to do the work, as it amount to solve
non linear equations with inherent huge data dependencies, i would think that
SPICE simulating a whole CPU or GPU can be done only at very low frequencies, a few kHz, and still, it takes ages.

To give you an idea a modern CPU require a few seconds to SPICE simulate a basic circuit using a few dozen transistors, and that s transistors that have 10x less parameters than say a 22nm finfet one, i let you imagine one billion transistors
with the die routings parameters included on top.

 

[Hitman928]

IC design is not rendering, eventualy GPUs can be used for VHDL since that s simulations at the functions blocks level but for anything SPICE i can guarantee
you that Nvidia has a full building of CPUs to do the work, as it amount to solve
non linear equations with inherent huge data dependencies, i would think that
SPICE simulating a whole CPU or GPU can be done only at very low frequencies, a few kHz, and still, it takes ages.

To give you an idea a modern CPU require a few seconds to SPICE simulate a basic circuit using a few dozen transistors, and that s transistors that have 10x less parameters than say a 22nm finfet one, i let you imagine one billion transistors
with the die routings parameters included on top.

Chips with billions of transistors aren't simulated through SPICE simulators.

 

[Abwx]

Chips with billions of transistors aren't simulated through SPICE simulators.

They SPICE simulate the basic blocks to extract the input/output characteristics wich then can be used as black boxes in VHDL, of course whole CPUs are also characterised this way, SPICE is too low level to be of use for rapid iterations.

 

[beginner99]

Stop comparing it to the previous generation and look at the 5070 instead.

the 5070 has less SMs than a 4070 Super and might not even match it. if 9070 non-xt is 7900xt in raster, it will be clearly faster than a 5070 with more vram.

I stand by my previous statement:

9070 XT for $599
9070 for $499

With 9070 having less CU and lower clocks.

That is the most optimistic pricing. these won't be any cheaper than that and much more likely the price is higher.

 

[adroc_thurston]

and much more likely the price is higher.

No, they're pretty cheap parts to build overall.

 

[Shmee]

Hello all, just a friendly reminder to keep profanity out of the tech section, and keep conversations civil. If you suspect someone has broken our rules, please use the report button, do not attack someone back. Also, remember that in vendor threads, cheering for said vendor is allowed, but trolling for a competitor is not. If you want to cheer for a vendor, do so in their thread. Thanks.

 

[del42sa]

I'm looking forward to today's full presentation of the mighty RDNA4 architecture and for the price announcement. Radeons gonna be back with all it´s glory

I bought three packs of popcorn, so I hope it doesn't go to waste 🍿🍿🍿 😎

 

[Keller_TT]

the 5070 has less SMs than a 4070 Super and might not even match it. if 9070 non-xt is 7900xt in raster, it will be clearly faster than a 5070 with more vram.

The 4070S has less SMs and RT cores than a 3080 going by that metric and it blows it away with the same GDDR6X interface, let alone GDDR7.

It doesn't say anything without knowing the gains they've made with the architecture. It just needs to be a 4070Ti in raster and near TiS in RT to be the darling card of the mid-range along with the 4070S. The 4070S is going to be available for a while at $500 probably.

 

[adroc_thurston]

The 4070S has less SMs and RT cores than a 3080 going by that metric and it blows it away with the same GDDR6X interface, let alone GDDR7.

It has a helluva lot more frequency and a chungus L2. It's basically RDNA2 to Ampere's RDNA1.
Blackwell does not have all that. Just a new SM and it doesn't look all that impressive?

 

[Keller_TT]

It has a helluva lot more frequency and a chungus L2. It's basically RDNA2 to Ampere's RDNA1.
Blackwell does not have all that. Just a new SM and it doesn't look all that impressive?

It does come with a 9% increase in base clocks, a big increase (1/3) in memory bandwidth with GDDR7 speed, plus whatever their new SM and RT cores are capable of. RT has made much bigger gains gen on gen for NV, and even if that lessens a bit, it's still going to be near the TiS, probably memory limited by 12 vs 16 GB at 1440P native.
I'm not pinning my hopes on NV pulling a 4060Ti with the 5070.

 

[adroc_thurston]

increase in base clocks

Meaningless number.

a big increase (1/3) in memory bandwidth with GDDR7 speed,

Indeed.

plus whatever their new SM and RT cores are capable of.

Kinda the problem really, doesn't seem like a major™ capability expansion.

 

[Keller_TT]

Kinda the problem really, doesn't seem like a major™ capability expansion.

My reference for raster is a meagre 7-8% increase for the 5070 from 4070S, because going by historic NV tiers and price-performance, it should slot in for the displaced 4080 12GB, wink wink.
So, if the IPC gain is neutralized with 48 vs 56 SM, the game clock and mem bandwidth are going to be the horse power, and where does all that fall is the Q.

 

[adroc_thurston]

the game clock and mem bandwidth are going to be the horse power, and where does all that fall is the Q.

Yea.
But again, doesn't look all that impressive.

 

[gaav87]

The 4070S has less SMs and RT cores than a 3080 going by that metric and it blows it away with the same GDDR6X interface, let alone GDDR7.

It doesn't say anything without knowing the gains they've made with the architecture. It just needs to be a 4070Ti in raster and near TiS in RT to be the darling card of the mid-range along with the 4070S. The 4070S is going to be available for a while at $500 probably.

3090=4070ti (non-s)=6800xt all with 9800x3d in cyberpunk 1440p native no gimicks

 

[gaav87]

 

[gaav87]

Trouble in paradise ?

 

[beginner99]

So, if the IPC gain is neutralized with 48 vs 56 SM, the game clock and mem bandwidth are going to be the horse power, and where does all that fall is the Q.

true but we can guess from a CES presentation full of AI and fake frames and no mention on basic raster improvements. If raster improvements were good they would have shown them.

and I think the one title that they showed that doesn't have all the gimmicks it was deduced that the 5090 is maybe 20-30% faster than a 4090, much less than expected and it's the only card with a price increase. From that it was also estimated that a 5070 likely not consistently beat a 4070 super in raster.

the 5070 = 4090 is complete non-sense

 

[reaperrr3]

the 5070 has less SMs than a 4070 Super and might not even match it. if 9070 non-xt is 7900xt in raster, it will be clearly faster than a 5070 with more vram.

I stand by my previous statement:

9070 XT for $599
9070 for $499

With 9070 having less CU and lower clocks.

That is the most optimistic pricing. these won't be any cheaper than that and much more likely the price is higher.

I wouldn't rule out 549 for the XT and 479 for the non-XT.

AMD is aware of their brand and feature disadvantage, and with slightly cheaper memory, lower TDP (can go with somewhat cheaper cooling and board) and no extra cost from InFO packaging, the 9070 should be at least a little cheaper to make than the 7800XT.

 

[SolidQ]

Trouble in paradise ?

You mean rumor about new BIOS to improve perfomance and raise price?