![[H]ard|Forum](/styles/hardforum/xenforo/logo_dark.png){kind=logo}
As well as AMD can do, as long as Jen Hsun Huang is fit and in charge, I don't see a way in for AMD.
Alternatively, can Nvidia's ARM offerings give them an even bigger lead into the server and PC market.
It is a fight for survival for AMD
If Intel can actually deliver 18A, I see someone big jumping over to use their fabs. If not two or more. Anyways would AMD really be a competitor with Intel if they use their fabs or more like a partner? Keeping X86 alive for another decade or two. As for being a decade behind, ahmm no.
Intels biggest issue is their E Cores,
Here's the thing....Not all HPC is equal. For a LOT of math, we need double precision or the errors grow too fast and the computation is worthless. I don't get the feel that Nvidia prioritizes FP64 given Blackwells specs..
FPGA SOCs have a lot of applications in edge compute.
From my limited understanding you are right, like Jensen said, cpu is about spending a limited amount of very high quality chips at a precise problem, GPU is about a massive amount of low quality non-precise chips for things that often do not have an actual exact answer to be found
Not sure about single depending what we mean, that over 1000mm of silicon sold for like $10-20k that look like this, I am not sure if it can fit in a single mono chip:
Strict FP64, is one of the few remaining locations where AMD shines.
CUDA hitting FP64 performance? There's NO chance double pumping fp32 is gonna match full 64 bit register math. There's NO chance poor condition matrix with fp32. Matlab is easily the most commonly used application in this space, and I don't see this at all on my home rig. I don't use GPUs on my work PC for matlab....but those are all quadro gpus.
$ for $ then Nvidia doesn't come close, but the MI300x doesn't do FP64 as fast as the new Blackwell parts, but the Blackwell parts cost more than 2x the MI300.
I think it hands down crushes the Blackwell in the sense that only the Super chip (gpu and grace cpu) hit 90 TFLOPS in total. The AMD MI300X is a single chip at nearly 82 TFLOPS.
Yeah Matlab uses the official AMD drivers as those are the only ones certified, AMD hasn't done a good job of updating them and it's painful.
Fairly sure the AMD part is not heterogeneous compute (feels weird to say homogeneous). Writing efficient cpu code and writing efficient gpu code are very different.From my limited understanding you are right, like Jensen said, cpu is about spending a limited amount of very high quality chips at a precise problem, GPU is about a massive amount of low quality impressice chips for things that often do not have an actual exact answer to be found
Not sure about single depending what we mean, that over 1000mm of silicon sold for like $10-20k that look like this, I am not sure if it can fit in a single mono chip:
View attachment 689475https://www.amd.com/content/dam/amd/en/images/products/data-centers/2882196-amd-instinct-mi325x.png
This seem like 4 different cpu title, with 8 cpu titles/hbm stack or something like that, "fused" in a way that look similar to nvidia with the GB200.
Unless Nvidia gets a little cocky starts over promising and under delivering and they leave room for others to step in and take their place.
AMD's gaming GPU sales sucks because AMD doesn't want to lower prices, and their Ray-Tracing performance sucks. As much as people hate the RTX 4060, it is a better GPU for the price. As for server sales, we know it's all for AI. The question is, how long does the AI bubble have before it pops? I really doubt it'll last beyond 10 years. The AI bubble would be lucky to last past 2026, which at that point AMD doesn't need to catch up anymore because the demand for AI hardware would have declined. Also everyone and their grandma is making AI hardware, so either way the market for AMD is going to be lesser but this also applies to Nvidia.
)Is it anything when talking Nvidia vs AMD, it is not like AMD as not been using TSMC and using very similar node, with some of the best collaboration in that regard, 3d cache, mi300x packaging, as Nvidia any advantage on that front ?
Surpass in what?
There up and down, late 90s for moments I think ?
The first consumer x86 processor to reach 1 gigahertz (GHz) was the Athlon, on June 23, 1999.
AI bubble won't pop. Luddite thinking. Here is just one reason why. Ever hear of Folding@Home? 24 years of intense computer usage replaced by AI.
Don't be surprised if NVidia goes on an acquisition spree. And Trump, unlike Biden, is much less likely to challenge an acquisition for being anti-competitive. So NVidia can overpay just to deny all competitors a chance to acquire a given company.
and more ... Sadly.
Well he had the good sense to stop Broadcom from buying Qualcomm previously so there’s that at least. His reason was a little strange but he still did it.
If we only talk about Ryzen: 5800X3D is the point at which they surpassed Intel. So, about 5 years. All on the same socket, too.
You don't think it'll happen because of a niche case? I'm not saying AI doesn't have a future, I'm just saying it doesn't have one that can sustain Nvidia and anyone else thinking of jumping into the AI train. Look at Microsoft and how they're trying to shoe horn AI into Windows with CoPilot and Recall. They need a way to make AI work for everyone and that's not going to happen this way.
Gaming is a different story, but AMD could do it if they made some simple changes. One is that they need to get better Ray-Tracing performance. If I were trying to compete with Nvidia i would try to find a way to get Ray-Tracing without any performance loss, because Nvidia hasn't cracked that code. The second thing AMD needs is to lower prices. Something investors don't wanna hear but they need market share and this is how you do it. The third thing AMD should do is leverage their graphics with their CPU. Ryzen is popular but Radeon graphics is not, so why not try to put good graphics with their CPU? The 9800X3D has built in graphics but it's terrible. It's not even as good as laptop graphics found in AMD's Ryzen AI 300 series. An easy way to expand Radeon graphics is to just include their latest graphics technology in their CPU's. This seems like the direction Intel is going, and it's not stupid.
Maybe for zen 6 (or later)
AMD gaming segment had 2% margin last quarter, I am not sure how much room they have in that regard.
Simple change.... like doing real time ray tracing faster than raster (what would that even mean.... how many rays, how many bounce, etc...there is no limit on how fast or slow raytracing get, it is a bit of a choice on the quality as well) also achieve to make better gpu (ie more performance by $ they take to make),
Are you suggesting they put 16 CU or more level of GPU in all their cpu and not just in their G line (8700G type of offer) ? Or you mean using only 2-4 cu still but the latest one ? zen5 come with "RDNA 3+" or rdna2 according who you read...
Also how old is Jensen What about his health. At some point he will be replaced. Is there a succession plan in place?
He seems extremely healthy from when I saw him. Watch him speaking....he easily has 20 more years, probably 30.
Just like auto makers in US and EU have to figure out how to compete against China, so too will AMD have to come up with new and creative ways to make cheaper and better hardware. Does nobody wonder why modern graphic cards are so big and heavy? What ever happened to using HBM memory like Vega and R9 Fury cards? It's gotten to the point where you can crack the PCB due to the weight.
Simple as in invest into it before Nvidia does. Right now Nvidia doesn't care about gaming, because they're now an AI company. Nvidia's solution to poor RTX performance was DLSS, and this is how Nvidia plans to solve this problem. At some point someone is going to get Ray-Tracing to perform as well on and off, it might as well be AMD.
Whatever they put in their Ryzen AI laptops they should also put in their desktop chips. This game with their G products needs to end. Intel's Lunar Lake is already better in GPU performance, and it's just a matter of time before Intel starts putting good graphics in all their CPU's.
Market share is market share. You know what's worse than a GPU nobody buys? A GPU that nobody supports. AMD has more advantages than Nvidia right now, that is until Nvidia figures out they're not an AI company. Right now AMD is their own worst enemy when it comes to how they use their GPU technology and they need to use their latest tech in everything.
I would have too that outside of cutting cost (R&D once it work accross sku line and better yield manufacturing wise), the only advantage* of chiplets is easier heat management, has you have a series on distinct heat spot easier to manager instead of a single one.
