It seems likely to certain, AMD is going to catch up to Nvidia. I'm not saying AMD will be on exactly the same level as Nvidia but they will get into the same ballpark, more or less. AMD has an impressive PowerPoint. That should be a nice boost for AMD that will come in about 1Q25. The AMD camp will be happy until 3Q25 when Nvidia is expected to roll their next gen. These days, both AMD and Nvidia are throwing money at AI. The race is on. It's Nvidia's to lose. I expect AMD will do well but I do not predict they will either catch Nvidia or pass them. I do predict that Microsoft will require about 40 TOPS to run Copilot locally. AMD will meet this spec by the end of the year. Intel will not. Intel is a distant third place, in the AI race.
It is clear you do your research before making statements, especially ones as bold as your first sentence. I have not come across anything to support this belief thus far. Do you have any links or sources that I could look at so I can see where you are coming from? Also, your post is a bit confusing in that you say AMD will catch up to Nvidia, but then you later state that "AMD will do well but I do not predict they will either catch Nvidia or pass them." I would love to put money into AMD, but I am not convinced just yet.
Nothing I wrote is fact. The future cannot be predicted as fact. In my defense, I ask that you cast your gaze toward the phrase, "It seems likely to certain..." Having said that, AMD's second generation AI accelerator is a huge step forward. What's more, it's a chiplet design so they can put it into any CPU complex or put a bunch together to form an AI specialty device. Obviously, it's not this simple and there are limitations but the idea stands. I believe the design would allow for a 4 way chiplet AI accelerator. That would get them to about 150 TOPS, based on the numbers I've seen. Perhaps more interesting would be a 2 AI chiplet, 1 CPU chiplet, 1 GPU chiplet complex. That would get them to 8 cores (16 threads) [Zen 5 will be 8C/16T per chiplet]. When later iterations of Zen 5 move to N3P, they are expected to move to 16C/32T per chiplet. Add 75 TOPS to that and a very competent GPU. You might literally have a device that redefines the world in AI technicolor. AMD will be a short term leader, IMO. I would not bet against them on CPUs. I'm not saying they will beat Intel but they are the leader and the incumbent usually remains on top. I would not bet against Nvidia on AI. Same reason. I'm just saying they will be in the ballpark for a while, as Nvidia waits on TSMC to cook the N2 sausage. When you see Intel and Nvidia commit to process nodes which are not stable, that means they need the small boost a process node step will bring. AMD is bringing out Zen 5 on N4. N4 is currently in production so is no risk and is cheaper than any of the N3 steps. By sacrificing a few percent in performance and power efficiency, they have a guaranteed win. If there are teething problems with N3P, they will still have N4P product to flood the channel. With semiconductors, the company in the lead has a massive advantage. AMD was a close second to Intel for years but Intel had a 6~18 month process node advantage and they dangled AMD in the wind for two decades. Now the tables have turned. As for sources, I won't go down that rabbit hole. There are people I believe have information and I assume everyone else is making shit up. What you are reading is opinion based on my personal information filter. Do with it what you wish or, better still, share any information you have that might be in agreement or conflict.
Yes, my post was worded terribly. For that, I apologize. I believe AMD will get into the same ballpark as Nvidia, with AI. They might even take the lead but, if so, it won't be by much and it won't be for long. Far more likely, AMD will get close for about 6 months before Nvidia pulls away. Keep in mind, if it's a close battle, AMD can move to N2 to match Nvidia on process and gain a little bit. If the accelerator battle is close, strategy becomes the deciding factor. One company can decide to sacrifice a bit of performance for cheaper parts and go after the lower end while the other company goes all out. Or, both companies can go all-out. It will be an interesting future!
So if I am reading you right, you believe that AMD will gain on Nvidia but only for the short term and then Nvidia will widen the gap again? Or do you believe that the gap will narrow and become more of a tit for tat battle? Again, I respect your views immensely, but I am not seeing AMD closing the gap anytime soon to put Nvidia's AI moat in danger. I can, however, see AMD being a cost-effective alternative to Nvidia that will allow them to grow into this frantically expanding AI market like they are beginning to do. It would emulate the existing consumer GPU market and provide AMD with a much needed high margin product line to bolster their R&D budget. Having said all this, Nvidia has an enormous amount of cash, the majority of the AI chip market, a thriving hardware/software ecosystem, and a brilliant CEO that is not making any mistakes that Intel made. Nvidia can and will throw its weight around to get priority treatment from TSMC just like Apple does. I am essentially doing the same thing you are when you say that your stances are based on an aggregation of sources that I deem historically accurate and informed. The only difference is my sources are leading me to believe there will be a maintaining of current Nvidia dominance for the near/mid term. Your statement of AMD closing the gap is noteworthy to me. When you say that you won't go down the rabbit hole of citing sources, it is these sources that I want to use to gauge my beliefs against yours to see if there are any errors in my line of reasoning.
This. I expect the gap to narrow. AMD has a huge step coming online by end of year. Nvidia will then widen the gap with their next step but I don't think it will be as wide as it is now. I believe the nuance here is the idea of "good enough". Rumors are floating around that Microsoft will move Copilot local on machines with 40 TOPS. For this use case, assuming the rumor is remotely accurate and/or true, that will make Intel's NPU useless, AMD will have the middle ground, and Nvidia will have the top end with dedecated AI coprocessing. Of course, Copilot is niche at this time. Also, we don't know that AMD will achieve 40 TOPS. And then there is the idea that Intel could pull a rabbit out of a hat and achieve this. I have no doubt Intel could gaslight their performance and get Microsoft to certify their NPU, as long as they get reasonably close. I have no idea about other applications. We are looking at initial adoption. If AI follows the typical path, there will be a burst of interest followed by diminished demand followed by AI being a must-have for everybody creating a massive upgrade cycle. AMD's chiplet design appears brilliantly poised for the next generation compute ecosystem. Intel's monolithic designs are not looking all that viable, to be honest. (more drivel in the next post)
I hope this post doesn't offend anyone. It is not intended to insult with it's simplicity. Why are chiplets better than monolithic dies? Chiplets add complexity, cost, and failure modes. It seems like monolithic would be superior? Any silicon wafer has a few inclusions. These are electron microscopic blemishes which render the part at that random location unusable. If we know that a wafer averages 5 inclusions (I've heard the average is closer to 7 but don't have any data on this): - we would have a yield of 0% if we tried to lithograph a single part that consumes the entire wafer. - we would average a yield of 95% for wafers with 100 dies - we would average 99.5% for wavers with 1000 dies While it's not this simple, it is almost this simple. For this and other reasons, smaller dies provide better yield. Process shrink helps. Adding stuff like NPU to the part hurts. Intel will be producing their largest die yet on a brand new, more expensive, process. If there are any yield difficulties at all, Intel will be hurt more than AMD would be, on that same process. So, they will be paying more for what are likely to be lower yields. AMD will be paying less for known excellent yields. Intel is stretching to compete so they are all but forced to sign up for the new process. If everything goes well, Intel will be fine. For some reason, Intel continues to have a lot of street cred. Particularly in servers, they continue to dominate due exclusively to franchise value but that franchise value is diminishing.
A quick clarification. Intel uses chiplets and will continue to do so. What I have written above is misleading. As I understand it, Intel does not subdivide their CPUs. They cannot take a pair of 8 core dies and glue them together for a 16 core CPU complex. Further, the leaks I've seen (and mostly trust) show the NPU will share a die with CPU on Arrow Lake. Where the chiplets come in (Intel calls them "Foveros") is primarily for the GPU. In fact, current generation Intel chips are built on the Intel 4 with integrated graphics built with TSMC's N3E process node.
The corporate salesmanship and the fanbois do not seem connected to the production realities of Nvidia. To a significant extent, Nvidia has been pumped up like Tesla.
$NVDA closed the week at $886.45. It sounds like the next gen GPU will be released at the end of June. That will surely boy the stock. I doubt it will make a significant difference for the company and bottom line but people love halo products, particularly if they can use them to play Halo.
The 5090 will be a total beast. It will sell out on day one, though it'll be way too rich for my blood for sure.
