NVIDIA introduces Hopper H200 GPU with 141GB of HBM3e memory

[filpo]

Summary

NVIDIA Hopper H200 GPUs Supercharged With World’s Fastest HBM3e Memory, Grace Hopper Superchips Power Jupiter Supercomputer

 

Quotes

NVIDIA H200 GPU: Supercharged With HBM3e Memory, Available In Q2 2024

Quote

The NVIDIA H200 GPUs are equipped with Micron's HBM3e solution with memory capacities of up to 141 GB and up to 4.8 TB/s of bandwidth which is 2.4x more bandwidth and double the capacity versus the NVIDIA A100. This new memory solution allows NVIDIA to nearly double the AI inference performance versus its H100 GPUs in applications such as Llama 2 (70 Billion parameter LLM).

 

In terms of solutions, the NVIDIA H200 GPUs will be available in a wide range of HGX H200 servers with 4 and 8-way GPU configurations. An 8-way configuration of H200 GPUs in an HGX system will provide up to 32 PetaFLOPs of FP8 compute performance and 1.1 TB of memory capacities.

NVIDIA Grace Hopper Superchips Power 1-Exaflop Jupiter Supercomputer

Quote

In addition to the H200 GPU announcement, NVIDIA has also announced a major supercomputer win powered by its Grace Hopper Superchips (GH200). The Supercomputer is known as Jupiter and will be located at the Forschungszentrum Jülich facility in Germany as a part of the EuroHPC Joint Undertaking and contracted to Eviden and ParTec. The supercomputer will be used for Material Science, Climate Research, Drug Discovery, and More. This is also the second supercomputer that NVIDIA announced in November with the previous one being the Isambard-AI, offering up to 21 Exaflops of AI performance.

 

Quote

In terms of configuration, the Jupiter Supercomputer is based on Eviden’s BullSequana XH3000 which makes use of a fully liquid-cooled architecture. It boasts a total of 24,000 NVIDIA GH200 Grace Hopper Superchips which are interconnected using the company's Quantum-2 Infiniband. Considering that each Grace CPU packs 288 Neoverse cores, we are looking at almost 7 Million ARM cores on the CPU side alone for Jupiter (6,912,000 to be exact).

 

Performance metrics include 90 Exaflops of AI training & 1 Exaflop of high-performance compute. The supercomputer is expected to be installed in 2024. Overall, these are some major updates by NVIDIA as it continues to lead the charge of the AI world with its powerful hardware and software technologies.

My thoughts

141GB of VRAM is crazy. I can only imagine how much it'll cost tho

 

Sources

NVIDIA introduces Hopper H200 GPU with 141GB of HBM3e memory - VideoCardz.com

NVIDIA Hopper H200 GPUs Supercharged With World's Fastest HBM3e Memory, Grace Hopper Superchips Power Jupiter Supercomputer (wccftech.com)

[bcredeur97]

141/8 ~= 17 Gamers, 1 GPU? 

Lol

[filpo]

4 minutes ago, bcredeur97 said:

141/8 ~= 17 Gamers, 1 GPU? 

Lol

Let's hope they don't play The Last Of Us at the same time 

[igormp]

Is it basically an H100 with even faster memory?

[Fasterthannothing]

And this is why Nvidia doesn't give a crap about the consumer market anymore. As soon as they are having to try and sell consumers grade GPUs they will bow out entirely. The only reason it's still viable for them is they have enough of a performance lead not to need to pay for development nor change their pricing structure. If one generation flops they will pull it entirely. Nvidia has secured the enterprise sector so that's all they really need.

[starsmine]

46 minutes ago, Fasterthannothing said:

And this is why Nvidia doesn't give a crap about the consumer market anymore. As soon as they are having to try and sell consumers grade GPUs they will bow out entirely. The only reason it's still viable for them is they have enough of a performance lead not to need to pay for development nor change their pricing structure. If one generation flops they will pull it entirely. Nvidia has secured the enterprise sector so that's all they really need.

yea. no thats crap. 
Its seperate markets and both make money. They are not cannibalizing themselves and the R&D done in one area inform the R&D in the other. Tensor came from VOLTA and now is their main selling feature for DLSS.

They do give a crap, they just are not reliant on it and are not worried about fully sinking from a bad generation. 

Like they now make chips FOR tsmc to speed up design
https://nvidianews.nvidia.com/news/nvidia-asml-tsmc-and-synopsys-set-foundation-for-next-generation-chip-manufacturing
Its assists them (and everyone else), and we all know TSMC is willing to spend millions on chips that will simplify their work flow. 

Doesnt mean because of that chip they dont give a crap about their other markets.

[Arika]

1 hour ago, Fasterthannothing said:

And this is why Nvidia doesn't give a crap about the consumer market anymore

If I was nvidia, I wouldn't either. Gamers are entitled brats who will never be happy unless they can get a xx90 gpu for $200 and even then still find something to whine about.

[Kisai]

2 hours ago, starsmine said:

 
Its seperate markets and both make money. They are not cannibalizing themselves and the R&D done in one area inform the R&D in the other. Tensor came from VOLTA and now is their main selling feature for DLSS.

They do give a crap, they just are not reliant on it and are not worried about fully sinking from a bad generation. 
 

 

Trickle-down in how it's supposed to work

 

2 hours ago, Arika said:

If I was nvidia, I wouldn't either. Gamers are entitled brats who will never be happy unless they can get a xx90 gpu for $200 and even then still find something to whine about.

 

The thing is, it's been pretty much impossible to "cost reduce" the GPU's since the RTX series, since what are you going to do? take a laser and cut off the enterprise FP64 section? No. Games almost never use FP64, and there's more of a push to use FP16 (half precision) in areas where precision doesn't matter, only an approximation does.

 

Like there are default settings in pytorch and such to do half-precision by default.

https://docs.nvidia.com/deeplearning/performance/mixed-precision-training/index.html

Quote

2. Mixed Precision Training

Mixed precision training offers significant computational speedup by performing operations in half-precision format, while storing minimal information in single-precision to retain as much information as possible in critical parts of the network. Since the introduction of Tensor Cores in the Volta and Turing architectures, significant training speedups are experienced by switching to mixed precision -- up to 3x overall speedup on the most arithmetically intense model architectures. Using mixed precision training requires two steps:

1. Porting the model to use the FP16 data type where appropriate.
2. Adding loss scaling to preserve small gradient values.

 

The ability to train deep learning networks with lower precision was introduced in the Pascal architecture and first supported in CUDA 8 in the NVIDIA Deep Learning SDK.

The thing is, certain workloads should NEVER be half precision, where accuracy is preferred. 

 

Like one of the dead giveaways to deepfakes is how low resolution the result is, which is a consequence of trying to fit the deepfake data into a commodity GPU and not something with 5 times as much memory. Like, you just know that's a deepfake because it sounds like it was recorded on VHS.

 

[leadeater]

5 hours ago, Kisai said:

The thing is, it's been pretty much impossible to "cost reduce" the GPU's since the RTX series, since what are you going to do? take a laser and cut off the enterprise FP64 section? No. Games almost never use FP64, and there's more of a push to use FP16 (half precision) in areas where precision doesn't matter, only an approximation does.

No need, there are no FP64 execution units in any Nvidia GPU die other than Gx100 and that has been true for so many generations. Kepler was the last to feature FP64 in hardware across more than a single die and inside non datacenter GPUs even if the performance was locked out. Maxwell had none at all, totally zero FP64 in all die variants so if you needed FP64 in the datacenter and from GPU then you would stick to Kepler. Do note though GPGPU back then was by no means what it is now,

 

Pascal was the first and set the current standard on GPU dies for compute and dies for graphics, Gx100 dies since then do not feature any display output hardware in the die so any workload that relies on that (screen connected or not) don't work, they are strictly for compute.

 

So FP64 has no bearing at all on gaming GPU prices, other than Nvidia wanting to make the one die(s) that do, the Gx100 dies.

[Kisai]

1 hour ago, leadeater said:

No need, there are no FP64 execution units in any Nvidia GPU die other than Gx100 and that has been true for so many generations. Kepler was the last to feature FP64 in hardware across more than a single die and inside non datacenter GPUs even if the performance was locked out. Maxwell had none at all, totally zero FP64 in all die variants so if you needed FP64 in the datacenter and from GPU then you would stick to Kepler. Do note though GPGPU back then was by no means what it is now,

 

Pascal was the first and set the current standard on GPU dies for compute and dies for graphics, Gx100 dies since then do not feature any display output hardware in the die so any workload that relies on that (screen connected or not) don't work, they are strictly for compute.

 

So FP64 has no bearing at all on gaming GPU prices, other than Nvidia wanting to make the one die(s) that do, the Gx100 dies.

Yeah, my point was that prior before nvidia got into the GPGPU game, their parts were basically divided between Geforce and Quadro, and the differences between the two didn't exist at the die level, just RAM, DP connectors and Coolers. When Nvidia started going big with the GPGPU stuff, that's when we stopped seeing SLI on consumer hardware, and if you wanted it, you needed the parts that would have previously been called Quadro's. Now we're even past that.

 

It's hard to believe that Nvidia basically graduated from doing consumer graphics cards as their core business, and now only make them to what I can only believe is to minmax the dies they get from a wafer.  Can't suddenly turn unsold dies into cheaper cards by cutting them in half. But it might also be their loss if Intel starts to eat the entire low-end (eg the parts equal to nvidia xx50/xx60 parts) segment. AMD is still making cards in that segment, but only Nvidia is actually making cards at the high end. But unless there is a sudden push for 8K or 4Kp144 as standard, all that leaves is VR, and that's pretty much been a niche market as well. Finding use cases for the highest end parts is deceptive because Nvidia is actually nerfing the FP64 use cases. Sure, in some cases you can get improved visual quality from a game, but we plateau'd in 2015 in how much visual fidelity is needed to sell a game. People are perfectly fine with games looking like FFXIV and Fortnite in online competitive environments. No RT or DLSS necessary. Even stuff like Alen Wake, Last Of Us, Cyberpunk 2077 and such aren't really using high end parts in a way that justifies purchasing a 4090. I don't know of any game that raytracing being turned on is justified. But that's largely because RT features on a entry-level card craters the performance.

 

 

[leadeater]

3 hours ago, Kisai said:

but only Nvidia is actually making cards at the high end.

Since when is the 7900 XTX not high end? Sure it's on average 17% slower (depends on games used) but it's also not like there are zero games where the 7900 XTX is not faster. The idea that there can only be one "high end" card is a little silly or that if one company has a bit better card in performance that only they are in the high end too.

 

7900 XTX is high end, it's just not the absolute best and that is fine. Really is a odd way to look at things since we don't do it as much in other areas of our lives and products outside of computer tech.

[Arika]

58 minutes ago, leadeater said:

Since when is the 7900 XTX not high end? Sure it's on average 17% slower (depends on games used) but it's also not like there are zero games where the 7900 XTX is not faster. The idea that there can only be one "high end" card is a little silly or that if one company has a bit better card in performance that only they are in the high end too.

 

7900 XTX is high end, it's just not the absolute best and that is fine. Really is an odd way to look at things since we don't do it as much in other areas of our lives and products outside of computer tech.

The one thing this forum has taught me, is that if you don’t make the best of the best of the best, then you may as well not even try and just not release anything.

[Techstorm970]

1 hour ago, Arika said:

The one thing this forum has taught me, is that if you don’t make the best of the best of the best, then you may as well not even try and just not release anything.

Indeed, there comes a point where happiness is a choice.  Just because we didn't beat Ohio State doesn't mean we suck.

 

Wait...this is a tech forum...

[starsmine]

3 hours ago, leadeater said:

Since when is the 7900 XTX not high end? Sure it's on average 17% slower (depends on games used) but it's also not like there are zero games where the 7900 XTX is not faster. The idea that there can only be one "high end" card is a little silly or that if one company has a bit better card in performance that only they are in the high end too.

 

7900 XTX is high end, it's just not the absolute best and that is fine. Really is a odd way to look at things since we don't do it as much in other areas of our lives and products outside of computer tech.

I wonder if to be charitable about it, they meant that for the RDNA 4 generation, the high end big chip is canceled. So it was a forward-looking statement, then a current state of things statement.

(Big RDNA 5 is still on track)

[williamcll]

141 is very odd of memory setup

[Kisai]

6 hours ago, leadeater said:

Since when is the 7900 XTX not high end? Sure it's on average 17% slower (depends on games used) but it's also not like there are zero games where the 7900 XTX is not faster. The idea that there can only be one "high end" card is a little silly or that if one company has a bit better card in performance that only they are in the high end too.

 

7900 XTX is high end, it's just not the absolute best and that is fine. Really is a odd way to look at things since we don't do it as much in other areas of our lives and products outside of computer tech.

People don't use AMD cards to do things that only are offered by NVidia cards (eg CUDA, NVENC, etc), AMD might offer a high end card, but as far as consumer cards go, they can only be used for games unless there is a push against the inertia. As an example OBS didn't offer hardware encoding on AMD cards largely because AMD kept changing the underlying hardware and API's. This problem originates from FFMPEG itself. Even the Intel encoder is better supported. Davinci Resolve doesn't support the AMD encoder. In Deep Learning Nvidia is pretty much the only thing there is. AMD ROCm is pretty much unsupported by everything. Only where a software program explicitly uses OpenCL/Vulkan does AMD get a chance to offer something comparable.

 

3 hours ago, starsmine said:

I wonder if to be charitable about it, they meant that for the RDNA 4 generation, the high end big chip is canceled. So it was a forward-looking statement, then a current state of things statement.

(Big RDNA 5 is still on track)

A high end chip that can only be used for gaming is not as valuable as one that can be used for everything. The primary reason I quit buying AMD cards was because they tend to always have lower build quality, so I've been unwilling to spend money on a high end "gaming only" card from a vendor that has a reputation for low build quality. My use cases span both those that require CUDA, and those that only support Nvidia encoders (eg Davinci Resolve Studio.) I wish AMD would offer a compelling product that fits all my use cases, but Davinci and OBS are use case #2 behind games, and CUDA AI stuff #3. 

 

If I was building a new computer, the only use for an AMD part would be for functionality that is exposed through Vulkan. So pretty much games, and possibly OBS if that API remains stable. 

[leadeater]

3 hours ago, Kisai said:

As an example OBS didn't offer hardware encoding on AMD cards largely because AMD kept changing the underlying hardware and API's.

 

They have for quite a while and from what I remember that was more OBS fault than AMD's, since OBS was the one changing stuff and support for Nvidia was a little more important 

 

Don't blame AMD for something that was directly OBS, they changed their core backend not AMD. Brand name changes to APIs and stuff doesn't mean anything was actually changed and the hardware encoder and how it works hasn't changed. Luckily the encoder itself has gotten better but how is no different to NVEC upgrades, usage remains consistent.

 

And a RTX 4090 is exactly no better than a RTX 4060 for NVENC so how that factors in to it being the only high end card when literally every Nvidia card is equally capable across a generation and model range bar the lowest of low end where they may have cut it out or relegated to previous generation NVEC

 

3 hours ago, Kisai said:

A high end chip that can only be used for gaming is not as valuable as one that can be used for everything.

That doesn't make it not high end and that's also the vast majority of what consumer GPUs are used for. In total numbers very few care about anything else than playing games on a RTX 4090/4080/4070 Ti. Nvidia learnt that VERY strongly with the Titan cards, no matter how much they insisted they weren't for gaming the only thing that actually mattered was gaming heh.

 

Not that I don't agree about the pain points and feature/support issues with AMD GPUs, it's just that it doesn't really make them not have a high end GPU option broadly suitable to most people. Use case specific needs do matter, but that's up to the person buying.

[igormp]

6 hours ago, leadeater said:

They have for quite a while

A bit over 1 year ago.

6 hours ago, leadeater said:

from what I remember that was more OBS fault than AMD's

No, it was actually AMD's fault with their poor SDK and no official support. It still has no proper linux support, and AMD also had poor ffmpeg support back then.

[leadeater]

49 minutes ago, igormp said:

No, it was actually AMD's fault with their poor SDK and no official support. It still has no proper linux support, and AMD also had poor ffmpeg support back then.

Wasn't the AMD hardware encoder supported and working prior to AMF (2016)? Even after then there was a plugin to get it working so it's not like it was that hard. I distinctly remember OBS changing their core and AMD not being supported after that for ages, of course chances are I'm remembering a different software situation lol 

[Kisai]

On 11/14/2023 at 12:43 PM, leadeater said:

 

They have for quite a while and from what I remember that was more OBS fault than AMD's, since OBS was the one changing stuff and support for Nvidia was a little more important 

 

Don't blame AMD for something that was directly OBS, they changed their core backend not AMD. Brand name changes to APIs and stuff doesn't mean anything was actually changed and the hardware encoder and how it works hasn't changed. Luckily the encoder itself has gotten better but how is no different to NVEC upgrades, usage remains consistent.

Because it IS AMD's fault? On AMD's user support pages, OBS's forum, OBS's github and FFMPEG you see references to "new AMD driver, broken support" and a lot of these are the RDNA 1.0 Radeon 5xxx parts.

 

It also doesn't change the fact that it's unsupported in Davinci Resolve Studio, and the same issues with OBS also plague other video tools like Adobe Premiere Pro and Vegas Pro.

 

 

 

On 11/14/2023 at 12:43 PM, leadeater said:

And a RTX 4090 is exactly no better than a RTX 4060 for NVENC so how that factors in to it being the only high end card when literally every Nvidia card is equally capable across a generation and model range bar the lowest of low end where they may have cut it out or relegated to previous generation NVEC

 

That doesn't make it not high end and that's also the vast majority of what consumer GPUs are used for. In total numbers very few care about anything else than playing games on a RTX 4090/4080/4070 Ti. Nvidia learnt that VERY strongly with the Titan cards, no matter how much they insisted they weren't for gaming the only thing that actually mattered was gaming heh.

You do realize that there is still exactly 0 support for CUDA on AMD right? There might have been a brief bubble in prices due to Ethereum mining when someone could write the crypto against the AMD card directly, but after that, those wealth-seeking greed monkies switched to AI-generated content, and none of that stuff runs on AMD parts, only Nvidia.

 

On 11/14/2023 at 12:43 PM, leadeater said:

Not that I don't agree about the pain points and feature/support issues with AMD GPUs, it's just that it doesn't really make them not have a high end GPU option broadly suitable to most people. Use case specific needs do matter, but that's up to the person buying.

If you only need a "Gaming card", then there's nothing wrong with picking the AMD card. But if you're doing anything else that uses the GPU, the AMD card ranges from "broken support" to "no support" in the tools people use. Davinci Resolve supports the Intel GPU.

You are pretty much SOL if you have an AMD CPU + AMD GPU if you want to use the hardware encoder on Davinci Resolve Studio. You would be better off with a MacMini M1 if that is your use case.

 

My point is that a "high end" marketed card without support for a feature preclude it from being "the right choice", and thus calling it "high end" is maybe being dishonest. What AMD should be doing, is what Nvidia did with FFMPEG and actually just do the work for them so drivers stop being the thing that breaks it.

 

AMD has a mixture of API's VAAPI, VDPAU in addition to AMF

 

AMD can say they support video encoding and AI training until they are blue in the face, but half-assed support and users consistently complaining about support being broken for video encoders, and still no means of actually using anything written against PyTorch CUDA on ROCm

 

https://rocm.docs.amd.com/en/latest/release/windows_support.html

 

Tensorflow support is even worse. I'll just be honest here and say that Tensorflow under windows is basically nerfed or useless, no matter what GPU you use, so I can't comment if the AMD support is any better than the Nvidia support, but previous attempts to use tensorflow always result in some part of the process running on the CPU under Windows.

 

And who knows, maybe it will eventually work where CUDA stuff can be retargeted to ROCm, but I've yet to see something actually work.

https://github.com/ROCm-Developer-Tools/HIP

 

https://github.com/Lightning-AI/lightning

 

But if I were just wanting to develop something on hardware I already have, the hardware I already have is the Nvidia RTX 3090. As well as  3070Ti and a 1080, RTX 2070 mobile and a 1050Ti mobile. I already have hardware I can spare to try stuff on. The problem always comes back to "Well what If something dies?" The laptops are pure hell to run inference on, their pitiful fans screaming 'halp! halp!" as something spins up for 10 seconds. If I decided to flat out buy a prebuilt with an AMD Ryzen 9 7950X3D and a AMD RTX 7900 XTX. I can't simply plop whatever Python install I have on my current desktop or laptop over to it. It won't work. No magic wand will suddenly make something written by Nvidia (eg Tacotron2) to only use CUDA suddenly work. That would require actually going in and updating it against a newer Pytorch version that supports AMD ROCm, and also basically doing a lot of work that would require more understanding of the hardware than I really care to.

 

In some ways it feels like the AI training side is just re-learning all the solutions that were already learned before about optimization for CPU's (remember MMX?)  No matter how good-intentioned your solution is, if you're not the first to do it, your solution will never win over once there is an installed base of software that assumes that feature is standard and you are just the weirdo who wants to speak French while refusing to speak English to people speaking English who have never spoken French. Sometimes you'll luck out and find out yes, that software library speaks your language, but you're largely just running into "Speak English dammit!" libraries.

 

[leadeater]

1 hour ago, Kisai said:

If you only need a "Gaming card", then there's nothing wrong with picking the AMD card. But if you're doing anything else that uses the GPU, the AMD card ranges from "broken support" to "no support" in the tools people use. Davinci Resolve supports the Intel GPU.

It also supports AMD too, it's also relatively better performance wise than other applications like Adobe. You just don't see a tick box in the config screen but it is absolutely supported.

 

 

Quote

Check Task Manager to see if the GPU/Quick Sync is being used for decoding. Note that AMD GPUs do not report this data directly, so you have to infer whether it is being used based on the CPU load

https://www.pugetsystems.com/labs/articles/what-h-264-and-h-265-hardware-decoding-is-supported-in-davinci-resolve-studio-2122/

 

And Puget has been benchmarking AMD GPUs in Resolve for as long as I remember looking at their test suites

https://www.pugetsystems.com/labs/articles/what-h-264-and-h-265-hardware-decoding-is-supported-in-davinci-resolve-studio-2122/

 

Been true for Puget going back to AMD Vega so it's being working in that software for at a minimum that generation but I'll be willing to bet before then.

https://www.pugetsystems.com/labs/articles/davinci-resolve-14-nvidia-geforce-vs-amd-radeon-vega-1213/

 

And that and everything else you keep posting is way besides the point, AMD does make high end GPUs that many people can use and are well supported, and if you want to nit pick so much then look at what formats Nvidia GPUs don't support and ALL the ones Intel do (first link).

 

1 hour ago, Kisai said:

If I decided to flat out buy a prebuilt with an AMD Ryzen 9 7950X3D and a AMD RTX 7900 XTX. I can't simply plop whatever Python install I have on my current desktop or laptop over to it. It won't work. No magic wand will suddenly make something written by Nvidia (eg Tacotron2) to only use CUDA suddenly work.

Honestly what on earth are you saying, yes you can. Buy it with an NVidia GPU. Do you even double check what you say? Do the same with a Prebuilt with ONLY an Intel CPU and tell me how far you're going to get with a CUDA only use case.

 

Can we just TL;DR this, only your situation matters (to you) so the entire world revolves around your use case and as such you get to decide what is and is not "something". That'll make posts much shorter and not full of factual errors.

 

1 hour ago, Kisai said:

My point is that a "high end" marketed card without support for a feature preclude it from being "the right choice", and thus calling it "high end" is maybe being dishonest.

No, what IS dishonest is the exact above statement. The world and GPUs isn't only about AI/ML and video editing despite what you want to spout and a lot of what you say is wrong anyway. Before you go off saying things do or don't work first check the validity of the statement so honest discussion can happen.

 

And of all the most dishonest things ever is people complaining their CUDA things they wrote themselves doesn't work on anything not CUDA while refusing to do the work. Yes there have been long standing issues in that realm of the industry and on the hardware and tools side but that is getting better but the biggest problem of all is the people, the people not putting in the work because they are happy with CUDA and then in turn obliviously point to other "not CUDA" things being as good. So how did CUDA get so good? Time and effort, go do it.

[Kisai]

4 minutes ago, leadeater said:

Honestly what on earth are you saying, yes you can. Buy it with an NVidia GPU. Do you even double check what you say? Do the same with a Prebuilt with ONLY and Intel CPU and tell me how far you're going to get with a CUDA only use case.

 

Can we just TL;DR this, only your situation matters (to you) so the entire world resolves around your use case and as such you get to decide what is and is not "something". That'll make posts much short and not full of factual errors.

The TL;DR is AMD is always playing second fiddle and constantly reinventing the wheel, thus ensuring it keeps playing second fiddle. You see this right now with the entire DLSS/FSR stuff nobody asked for, and also the Mesh Shaders that only work on the most recent hardware.

 

If you pick AMD you are going to be waiting for middleware before that feature you want gets implemented, if ever.

[NumLock21]

On 11/13/2023 at 1:13 PM, filpo said:

Let's hope they don't play The Last Of Us at the same time 

They really need to play Cities Skylines 2!

[Erebuxy]

On 11/13/2023 at 11:09 PM, Kisai said:

The thing is, it's been pretty much impossible to "cost reduce" the GPU's since the RTX series, since what are you going to do? take a laser and cut off the enterprise FP64 section? No. Games almost never use FP64, and there's more of a push to use FP16 (half precision) in areas where precision doesn't matter, only an approximation does.

For sure, ML definitely accelerate the decline of FP64. ML models are basically an approximation of things. No one needs that much precision. Nvidia's 19bit TF32 is also pretty cool.

Nvidia cares so less about FP64 at this point, that AMD has HPC cards perform a lot better than H200 on FP64 (tho get obliterated on FP16.