News Nvidia counters AMD DeepSeek benchmarks, claims RTX 4090 is nearly 50% faster than 7900 XTX

[YSCCC]

Sometimes I feel funny how we are racing to cut the final drop of fossil fuel for saving the world, yet ppl are easily spending/consuming a few times more fossil fuel/heat generated in these sort of home run AI in consumer GPUs which overlapps with one another for... gimmick more than actual useful things... these don't even produce entertainment...

 

[watzupken]

The main differentiator is the price. While both are last gen GPUs, the RTX 4090 may cost more than 2x the price of the RX 7900 XTX. So yeah, around 50% improvement but cost 2.5x more is not good value. However, if you have limited PCI-E slots and you need to maximize performance, then that leaves one with the RTX 4090 as the better option.

 

[watzupken]

Sometimes I feel funny how we are racing to cut the final drop of fossil fuel for saving the world, yet ppl are easily spending/consuming a few times more fossil fuel/heat generated in these sort of home run AI in consumer GPUs which overlapps with one another for... gimmick more than actual useful things... these don't even produce entertainment...

This is indeed the irony. On one hand, people are talking about the environment, global warming, etc. Other the other hand, tech companies and techies are racing to use even more power for the purpose of profits. Where is the governance and control here? Not that AI is all bad, but it is evident that it is very power intensive, and every big tech companies are deploying more and more hardware, instead of focusing on optimizing the algo/ logic. This is where I find that Deepseek really caught big US tech companies' pants down.

 

[fybyfyby]

Isn't RTX 4090 more than 2x the price of RX 7900 XTX so 47% faster officially confirms that it's worse? For RTX 5090 it's the same even with 2.2x faster.

So if you buy 2 7900 you will do better. At least for inference. And even will have more ram for inference. But for it its maybe better to buy 2 used 3090 instead.

 

[YSCCC]

This is indeed the irony. On one hand, people are talking about the environment, global warming, etc. Other the other hand, tech companies and techies are racing to use even more power for the purpose of profits. Where is the governance and control here? Not that AI is all bad, but it is evident that it is very power intensive, and every big tech companies are deploying more and more hardware, instead of focusing on optimizing the algo/ logic. This is where I find that Deepseek really caught big US tech companies' pants down.

Yea, and the irony to me is that personal AI models are even allowed to be promoted as such, for these stuffs to be working you need those massive hundreds of million worth of server and well monitored by pros to train any AI to be of remote usefulness, you and I running a few 5090 is only burning the nature for nothing

 

[DougMcC]

Huge win for AMD here. Prior to this, everyone would have just assumed that the best AI performance was NV. Now people will probably want to benchmark their use case before buying, and that is an absolutely huge change in favor of AMD. Even if they lose every time in absolute performance, people may consider the cost/perf outcomes at that point, and that may well favor AMD.

 

[Peksha]

This is possible.

I know for a fact LM studio supports ROCm.

LM Studio - Yes. But for nv - No. See small text in the right-bottom corner in the diagram

 

[Pierce2623]

This is possible.

I know for a fact LM studio supports ROCm.

Yep. LM studio is probably the most common way for everyday civilians to experiment in AI. AMD performance has literally improved by leaps and bounds in LM studio to the point that my 7800xt actually outperforms my 4070 super in models under 8B parameters or so and OBVIOUSLY outperforms the 4070 super in anything that requires between 12GB and 16GB.

 

[Pierce2623]

LM Studio - Yes. But for nv - No. See small text in the right-bottom corner

Sometimes I feel funny how we are racing to cut the final drop of fossil fuel for saving the world, yet ppl are easily spending/consuming a few times more fossil fuel/heat generated in these sort of home run AI in consumer GPUs which overlapps with one another for... gimmick more than actual useful things... these don't even produce entertainment...

Wait….do you think playing with something like Llama or ChatGPT just has a GPU blasting along for hours at the TDP limit? Playing in LM studio uses way less electricity than gaming. Get over yourself.

 

[YSCCC]

Wait….do you think playing with something like Llama or ChatGPT just has a GPU blasting along for hours at the TDP limit? Playing in LM studio uses way less electricity than gaming. Get over yourself.

Wattage per hour is lower, but for local running, you still need to train it quite a lot for it to be useful, the time is much longer than gaming (not counting those play Diablo to death cases). It’s just wasting money and power for no real reason for most ppl.

 

[edzieba]

Huge win for AMD here. Prior to this, everyone would have just assumed that the best AI performance was NV. Now people will probably want to benchmark their use case before buying, and that is an absolutely huge change in favor of AMD. Even if they lose every time in absolute performance, people may consider the cost/perf outcomes at that point, and that may well favor AMD.

People buying hundreds of millions of dollars worth of GPUs already benchmark before purchasing them. They haven't been choosing Nvidia cards just through guesswork or tealeaf reading.

Remember, training workloads are not inference workloads, even if both fall under the general umbrella of "AI". The Optimum card for training is not necessarily the optimum card for inference.

 

[Pierce2623]

Wattage per hour is lower, but for local running, you still need to train it quite a lot for it to be useful, the time is much longer than gaming (not counting those play Diablo to death cases). It’s just wasting money and power for no real reason for most ppl.

I don’t have to train models to run Stable Diffusion of ChatGPT at home. What are you even talking about? We went from people playing with AI at home to now we’re talking about training models which is a complex and intensive process for any model. That has nothing to do with your original comment saying people playing with Ai at at home are wasting electricity.

 

[Pierce2623]

People buying hundreds of millions of dollars worth of GPUs already benchmark before purchasing them. They haven't been choosing Nvidia cards just through guesswork or tealeaf reading.

Remember, training workloads are not inference workloads, even if both fall under the general umbrella of "AI". The Optimum card for training is not necessarily the optimum card for inference.

Yep and surprisingly enough, most of the idiots buying tons of Nvidia GPUs up just because they were told to don’t even need to do intensive training on a fleet of H100s.

 

[InnatelyAcidic89]

AMD didn’t run their tests well and nVidia got the opportunity to refute them. Major smack down.

Of course an Nvidia cuck would give them the benefit of the doubt, even after they lied straight to your face and told you a 5070 had comparable perf to a 4090

 

[YSCCC]

I don’t have to train models to run Stable Diffusion of ChatGPT at home. What are you even talking about? We went from people playing with AI at home to now we’re talking about training models which is a complex and intensive process for any model. That has nothing to do with your original comment saying people playing with Ai at at home are wasting electricity.

IIRC if you want to use something like Chat GPT or LM studio you could run local service for small models using 2060 or so and works perfectly fine, or even use cloud, if you want the TOTL GPUs one are more often trying to train it to do some specific tasks where the pre-trained models are not capable of.

What I meant was the irony that ppl just rush into the AI bubble with those TOTL GPU and keeping the price so unrealistically high and burning electricity for no real benefit or so compared to using the cloud ver. with much better trained models in super computers, how many playing AI really need to run it locally vs playing chat gpt on a phone?

 

[InnatelyAcidic89]

Because if you punch the biggest guy in the face you better be sure you knock him out.

That metaphor makes literally no sense in this instance. If i knock him out, he's able to get back up at some point, no? I'd better kill him. If you weren't an nvidia shill, the logical response would be, we need to see third party, non-biased results for both parties. Especially given Nvidia, and to a lesser extent AMD's track records with half truths and outright lies.

 

[InnatelyAcidic89]

That’s simply not true. Nvidia clearly ran AMD in inferior configurations using Vulkan instead of ROCm but running their cards in CUDA instead of Vulkan. You literally commented without even reading any of the real sources.

You're clearly arguing with a mouth breather. You do you though

 

[InnatelyAcidic89]

The irony of your rebuttal there is Deepseek made a huge fuss about it costing far less to train the model. So cost was immediately made a significant metric. You can buy 2 XTXs for less than the cost of a 4090.

I agree, but the irony of that ironic statement is deepseek completely just blatantly lied about their funding, and spent 1.5+ billion on their model.

 

[beyondlogic]

Claiming that your overpriced 4090 is almost as good as a 7900xtx at something is quite the idiotic statement lol.

 

[beyondlogic]

Sometimes I feel funny how we are racing to cut the final drop of fossil fuel for saving the world, yet ppl are easily spending/consuming a few times more fossil fuel/heat generated in these sort of home run AI in consumer GPUs which overlapps with one another for... gimmick more than actual useful things... these don't even produce entertainment...

Don't worry about it because I'm sure they will clap themselves on the back with we did this using solar panels that takes as much resources to make in probly fossil fuels lol.

 

[Gururu]

That metaphor makes literally no sense in this instance. If i knock him out, he's able to get back up at some point, no? I'd better kill him. If you weren't an nvidia shill, the logical response would be, we need to see third party, non-biased results for both parties. Especially given Nvidia, and to a lesser extent AMD's track records with half truths and outright lies.

It's all about perception, not reality. As I said, it doesn't matter what the truth is. nVidia is the authority on AI hardware at this time and their response will be what is considered true. Where is AMD's follow up? That will be interesting.

 

[Gururu]

Of course an Nvidia cuck would give them the benefit of the doubt, even after they lied straight to your face and told you a 5070 had comparable perf to a 4090

My last nVidia card was a hand me down GTX 570. I don't like them, but I am impressed with what they have been able to do, even though my gut screams "Bubble"!

 

[YSCCC]

Don't worry about it because I'm sure they will clap themselves on the back with we did this using solar panels that takes as much resources to make in probly fossil fuels lol.

From my background the data says absolutely... especially considering how much energy you are using to produce, transport and setup them vs the wearing and subsequent disposal of them during typhoon/hail etc. But in the grand scheme of things, scalping a ton of consumer/pro GPU to run those LLM is just better than mining crypto but not a whole lot more better

 

[LaDiabless]

AMD didn’t run their tests well and nVidia got the opportunity to refute them. Major smack down.

You mean Nvidia didn't run their tests well and used the weight of their hooting fanboy army to call it an own.

Nvidia benchamaks their cards with CUDA and Vulkan with the AMD cards and declares victory

Imagine if the AMD benchmarks had used Vulkan for the Nvidia cards and ROCm for the AMD cards?

people would be crying blue murder

 

[LaDiabless]

This is possible.

I know for a fact LM studio supports ROCm.

I think like half the performance features are missing like flash attention when you run under vulkan instead of ROCm