Soldato
- Joined
- 17 Aug 2003
- Posts
- 20,160
- Location
- Woburn Sand Dunes
It's broken because it wont scale? Where did you read that?
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD Navi 23 ‘NVIDIA Killer’ GPU Rumored to Support Hardware Ray Tracing, Coming Next Year
- Thread starter Gregster
- Start date
- Status
256-bit GDDR6 memory means only 512 GB/s throughput. This is again mid-range performance.
If they think to offer it, though, for cheap and in low power consumption, it may be right there next to RTX 3070 or maximum RTX 3080 if very optimised software.
If they think to offer it, though, for cheap and in low power consumption, it may be right there next to RTX 3070 or maximum RTX 3080 if very optimised software.
some info here.
https://www.igorslab.de/en/chip-is-...-dispersal-at-the-force-rtx-3080-andrtx-3090/
Nvidia buys as far I know, working dies and pays for them not anything else.
so yields in a way wont be their concern only samsungs.
Good deal as they couldn't have tsmc 7nm and went with 8nm samsung and got a low payment deal.
Associate
- Joined
- 19 Jun 2017
- Posts
- 1,029
I think what's happening here is called "strategy". Amd has a marginal share of the PC GPU market, that causes some uncertainty in capacity allocation, so low risk approach like mass manufacturing mid range at competitive prices
If what you say becomes real, I will skip and not buy a Radeon this generation.
I am waiting for a real card with 512-bit memory interface, 1TB/s memory throughput and performance over RTX 3080, at the very least.
Kopite provided a lot of decent 30X0 info so it it would unwise to dismiss it out of hand.
.......with your 160CU 420W+ GPUs
Err?
Ya, ya, Big Navi aka Navi 21 with 80 CUs doubled?
The leaked card is 1/4 of that, perhaps Navi 23 with die size around Navi 10's and performance over Navi 10 equal to the IPC increase.
Associate
- Joined
- 19 Jun 2017
- Posts
- 1,029
That's second level of binning.. that kind of binning happens at the AIB level. But the first level of binning happens at Nvidia level... thats what folks are more interested in when they talk abt yields .. also the distinction between 3090 and 3080 in that article is somewhat confusing as the general consensus is that the 3080 is a binned version of 3090.
also everything needs to be measured wrt reference baseline.. none of us really know how those stats translate to reference clocks at given TDP
AMD's console business is part of their enterprise embedded and semi-custom division. It's nothing to do with Intel.
Associate
- Joined
- 19 Jun 2017
- Posts
- 1,029
.......with your 160CU 420W+ GPUs
I have a 1kw seasonic.. that could even handle 2 Rx Vega 64s without breaking sweat.. I don't remember when was the last time the psu fans actually powered up under load .. would be happy to buy 1.. I like big chips
This is a completely bizarre post. Those memes came out originally about the Fx cards from Nvidia and then again with the first generation Fermi cards.
I take it you haven't been reading the Ampere thread, plenty of mocking about the size of the cards and the power requirements going over there.
And it's bizarre for another reason, no enthusiast cares overly much about power consumption as long as the performance is there and the cooler can handle it relatively quietly. The problem with AMD for the past number of launches is the reference coolers have been terrible, the power consumption has been very high and the performance hasn't matched the power use. Hence the mocking.
If the performance of the Ampere cards is worse than AMD's RDNA 2 cards and use more power, then there will be a lot of more mocking of Ampere cards. But, if it turns out that AMD's cards use more power just to be slightly behind Nvidia's performance, then, yeah, you will see more memes about AMD.
Thats week old
Count the CU's in this, it might be real, it might be fake, it might be real and AMD are just not counting Dual CU Shaders as 2 CU's, like Nvidia do........with your 160CU 420W+ GPUs
Whatever it is this is what i was talking about.
4 Shader Engines
2 Shader arrays in each Engine
10 Shaders in each array
2 CU's in Each Shader
= 160 CU's
Soldato
- Joined
- 21 Jul 2005
- Posts
- 21,107
- Location
- Officially least sunny location -Ronskistats
It has this to the right:
I'll say it again.
The one thing that no one got right about Ampere was the CUDA count, because no one knew Nvidia was going to release a Dual CUDA core Shader, this is how a 3080 went from being leaked as a 4352 CUDA core card to being an 8704 CUDA core card (its X2) A CUDA core is what AMD call a CU.
Its exactly the same thing, This slide shows a doubling of CU's, ignore the blue tab on the side i suspect that is someone's interpretation, count the CU's in the Digram, its written in the Shader Arrays to make it easy for you. 20 CU's in each Array with 8 Arrays, 8X 20 = 160.
LOOK at the diagram not the blue tab on the side.
I agree about power and cooling but the elephant in the room is always heat. If you don't have AC the heat output of a high end GPU is no joke. I had 7970 XF and the room would hit 30°C with the window open. I'm sure I nearly had heatstroke a few times! For that reason I'll take lower power consumption if possible. I may invest in some AC one day as my projector also kicks out a tonne of heat too.
I'll say it again.
The one thing that no one got right about Ampere was the CUDA count, because no one knew Nvidia was going to release a Dual CUDA core Shader, this is how a 3080 went from being leaked as a 4352 CUDA core card to being an 8704 CUDA core card (its X2) A CUDA core is what AMD call a CU.
Its exactly the same thing, This slide shows a doubling of CU's, ignore the blue tab on the side i suspect that is someone's interpretation, count the CU's in the Digram, its written in the Shader Arrays to make it easy for you. 20 CU's in each Array with 8 Arrays, 8X 20 = 160.
LOOK at the diagram not the blue tab on the side.
I probably need to expand on this and own up to my own misunderstanding of what makes up a Shader.
This Block Diagram shows 5120 Shaders, not 10240 as i said before, Compute Cores are inside Shaders, not the other way round. As i said before.
So its 160 Compute Units across 5120 Shaders, not 5120 Shaders across 80 Compute Units, these Shaders are 'Dual Core' Shaders vs Single Core Shaders.
The Compute Capabilities of this GPU will quadruple over the 5700XT, it has 4X the Compute Cores. BUT, Just as this doesn't translate to Gaming Performance in Ampere vs Turing it will not do that for RDNA2 vs RDNA either. Each Shader will be better in gaming but by how much is anyones guess, if Ampere is anything to go by they are "A bit better".
Edit: I'm not an expert and this is how my understanding has evolved. I don't need to be an expert to count tho....
Last edited:
Associate
- Joined
- 19 Jun 2017
- Posts
- 1,029
From my limited research they call 1 single precision MAD ALU as a shader core..so 1 CU or CUDA core contains "n" such ALUs, they will let this be known before hand.. another shortcut is to just divide GFLOPS by GHz and then halve it.
The reason why these ALUs don't translate to real world performance is because the bottleneck is somewhere else
5700XT = 9.7 TF FP32
You're looking at around 40 TF FP32 for this Big Navi.
The RTX 3090 is 36 TF
Last edited:
- Status