![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
They didn't skip Volta for consumer, it was never meant to be for consumers, just like hopper isn't. Really now? THAT'S your example? You really think they won't improve raster for two full gens? That's absurd.
Well, Enterprise gets Blackwell in 2024 with the big GH100, the rest of us need to wait until 2025, before we see it so no RTX 5000 BK, 6000 BK, or 5080/5090 for a full year after the big boys get their toys, so it will be old hat for them by then.
I don't know if you got the message from the CEO. Nvidia is an AI company now. How does raster benefit them?
at most 25% make it sound way more than it was from most that have I raid, the Turing 1660 super was RT-Turing core free and despite having about half the silicon used for cache the core count by mm2 does not match 25% of the 2060 being used for both RT and Tensor, it was probably more around 4 to 7% for RT cores, same for the Tensors.
Well, some are fewer, but on the AD102 it breaks down to some 24% of the silicon being dedicated to the Ray Tracing, AI, blah blah blah, going down to the AD106 where it is closer to 16%, the CUDA logic takes up the bulk of both as that is identical across the whole range which that takes up a set footprint within each of the SM's and then some 20-25% of the physical area doing the raster work. But on any given part of the AD series less than 50% of the chip itself is processing the actual graphics be it raster, ray-traced, or generated, the majority of the silicon is taken up by IO, memory, or logic.
The other side of this though is games as a whole are doing less and less with raster natively, and advancements brought forward in DX12U and Vulkan have improved raster performance by some 20% when those functions are used. Mesh Shaders and their associated culling methods reduce a lot of unnecessary processing that never gets seen or even displayed, and advancements in texture codecs have brought forward much better scaling which reduces the need for a lot of background processing and in some cases that is possibly to be offloaded to dedicated silicon and no longer a part of the raster process.
That is the potential hope for gamers IMO. Yes AMDs first chiplet attempt here had some issues. That is probably the solution going forward. The AI stuff still needs a small number of standard raster style compute cores, a solution where gaming cards could simply stack a couple of them is really the best hope of not having hardware stagnate, or jump in price so high it makes no sense.
You are probably not wrong as apparently TSMC 3nm limit die size quite a bit, we can expect grace hopper type a la M1ultra setup of pairing multiple GPU for the AI stuff and the 5090 to be smaller than the 4090 if it is monolithic.
Not quite, TSMC has a maximum reticle limit of 858mm2, but they have the ability to tile that out 2x3 for a 6x increase resulting in a max size of 5148mm2.
I have very little idea of what they are and their implication so I doubt it, I think this is what I had in mind:
OK yeah for N2 that is different, but the N2 process is being done on the EXE 5200 series and not the NXE 3600 and it's a whole other animal, pair that with a completely new transistor design bringing in GAA and new layer masking technologies and something that took the full 858mm2 on N3 with a few redesigns will fit well within that 429 with better clocks and lower power to show for it.
Jensen: NVIDIA is now an AI company.
That seem a bit hyperbolic. Raster only, a 4080 is quite faster than a 3080 and that would be true (just less and less) down to the 4060, the 4060 had a msrp significantly lower than the 3060 back in the days too:
This is why I miss real competition, because if Nvidia was intentionally selling sub par GPU’s you would think AMD would use it as an opportunity to walk all over them and gobble up market share. It’s not like gaming GPU’s are sold at a loss there is money to be made, and it’s obvious we’re willing to spend.
No doubt real competition would either force Nvidia to make proper gaming cards... or officially exit the market. AMD has had decent mid range offerings... but to really compete they need to outright win a flagship fight at some point. Nothing less will change minds. They can release all the 5800 / 7900 type hardware they want... and you can argue they are the best cards in their price class. If they can't dethrone them at the top end though Nvidia can continue selling cut down cards that rely on generation locked DLSS tags.
You're the last person I figured I would have to give this response to, as it has come up from time to time for at least the last two decades. Usually in regards to things like, not wanting to have to rebuy vRAM over again, etc.
I ask because I got dumped/volunteered with a stupidly frustrating problem not at all of my making and the only solution I can find makes everybody unhappy because they really have no idea what they signed up for but they are too far in to back out but too underfunded to proceed, and it's their hardware "requirements" that are killing them, so now I am trying to talk them down and figure out what they are actually trying to accomplish and if any of the hardware that they purchased incorrectly can be returned/resold.
Haha. Well that's a totally different, far a field thing there.
Not even my job, Village Council got funding to do some Tech Startup initiatives to get better internet infrastructure in the area to attract business and blah blah blah, and they need to make some reports back to the province on how they are progressing and the answer is they haven't made a lot of progress at all.