5 ms·
Micron Kicks Off Production of HBM3E Memory
- brucethemoose2 3y agoHow about higher capacity GDDR6X? 48GB consumer cards (or 96GB pro cards) would sell like hotcakes if AMD/Intel dare to break the artificial VRAM segmentation status quo.
- cyanydeez 3y agountil they drain the main vein, no MBA is going to greenlight artificial price inflation.
- teaearlgraycold 3y agoBuild it and people will start porting their CUDA stuff to run on other architectures.
- Brananarchy 3y agoYou over estimate the 'semi-pro' market for graphics cards. Gamers are barely willing to pay for 20GB. There's no market for consumer cards with an order of magnitude more RAM until games are built to use that memory.
- cinntaile 3y ago48GB cards would sell like hotcakes. The problem is that they would sell way less cards aimed at professionals, where they have much higher margins.
- AnthonyMouse 3y agoIntel doesn't sell a lot of graphics cards whatsoever though. Be the first to offer 64GB of VRAM for under $1000 and that could change pretty fast.
- xadhominemx 3y agoNot without CUDA unfortunately.
- KeplerBoy 3y agoDon't underestimate the amount of shit people would be willing to deal with to make stuff work. A capable GPU with 24+ GB would sell if it significantly undercuts Nvidia. Just look at geohot building his tinyboxes with AMD cards.
- xadhominemx 3y agoI would personally love that project but there are already so many versioning issues in the space it would be a nightmare if ROCm randomly broke things all the time.
- KeplerBoy 3y agoI agree, ROCm seems to be a mess from the outside, but I'm glad people are putting in the effort.
- AnthonyMouse 3y agoAnd we're talking about Intel here. AMD is going to price competitively against Nvidia but they'd still rather you buy a $20,000 MI300 than a hypothetical 128GB Radeon for $2000. Intel could very easily just put a buttload of VRAM on their existing GPUs to stick it to their competitors and make out like bandits. All they'd have to do is charge a Big markup instead of an Enterprise markup. And Intel has a better history of not making broken libraries.
- cyanydeez 3y agoa lot of the true AI value is context window size limited, not compute limited.
- ianbutler 3y agoThis is for ML, not gamers. There is an entirely different market here.
- usrusr 3y agoBut "basement ML" is a thing, the market of people who are interested in PC gaming but not to the point of being lifestyle gamers who throw every cent they can spare at that altar. The GPU they bought long before the pandemic is still running every game they throw at it, but they never completely stop eyeing the new stuff. Dipping their toes in ML, even if it's just getting through 80% of some stable diffusion setup tutorial, can be a very welcome excuse to upgrade their gaming. A card sold for gaming but with generously overprovisioned VRAM (ideally in the range of the lowest bin of the biggest or second-biggest chip I think) could match that market segment very well - and it would not only compete with other price points, it would actually increase the market by some buyers (those who would not upgrade without the "ML excuse").
- HeWhoLurksLate 3y agoI swear I'm getting Deja Vu right now, I coulda sworn I've seen this thread before. There's gonna be a guy commenting that "you don't need it" somebody else saying "but I want it!" And a few trying to figure out the economics of it and whether or not it makes any sense. Personally I'd love to have as much VRAM as possible (and as high a bandwidth as is possible too) to mess around with simulations in- but that's definitely a pro workload. I'd love to see like a flagship card have a stupid amounts of VRAM spec option - like an RTX 4090 with 32-48gb of VRAM just to see what happens with it on the market.
- cyanydeez 3y agolast gen A6000 has 48gb. I just put two of them in a server. about 5k each.
- omneity 3y agoA friend just got an M3 Max with 128GB (V)RAM and he's extremely happy with it (AI workloads wise). That could be an option if you can run your simulations on macOS.
- AnthonyMouse 3y agoThat machine with 128GB is $5000, with 48GB is still well over $3000, and has as much memory bandwidth as a $400 GPU. At the current spot price, 128GB of GDDR6 is <$400 and 48GB is <$150, implying that they could be paired with any existing <$1000 GPU to produce something significantly faster for dramatically less money. If anyone could be bothered to make one.
- omneity 3y agoI agree with you, that’s how it should be, but that’s not how it currently is. Looking at what’s available right now. You need 3 A100 40GB to get this amount of VRAM which will cost you way north of 20000$. Doing it with A6000s is still about 15k$. There’s not that many high VRAM options out there you know..
- hydroreadsstuff 3y agohttps://semiconductor.samsung.com/news-events/tech-blog/a-bridge-between-worlds-how-samsungs-gddr6w-is-creating-immersive-vr-with-powerful-graphics-memory/ https://semiconductor.samsung.com/news-events/tech-blog/a-br...
- hkgirjenk 3y agoDon't you think Nvidia analyzed, sliced and diced the market to figure out how to maximize profits? AMD is doing the same thing, the only high memory cards they put out (MI300) are for data centers.
- andy_ppp 3y agoIt’s possible to come up with many strategies and different companies will. Why are you so sure that Nvidia’s strategy is right for AMD or Intel who need to offer differentiation to get over the CUDA moat?
- xadhominemx 3y agoNVIDIA has actually not really sliced and diced the market. They only sell Cadillacs, which is fine for now because they're the only game in town.
- hmottestad 3y agoAnd Samsung just launched 36 GB packages of HBM3E: https://www.anandtech.com/show/21278/samsung-launches-12hi-36gb-hbm3e-memory-stacks-with-10-gts-speed https://www.anandtech.com/show/21278/samsung-launches-12hi-3... Combined with this[1] interesting paper from summer 2023 on HBM combined with Xeon processors which would now allow for 144 GB on a single CPU. In theory at least. 1: https://lenovopress.lenovo.com/lp1738-implementing-intel-high-bandwidth-memory https://lenovopress.lenovo.com/lp1738-implementing-intel-hig...
- hiddencost 3y ago> launched Not true. They announced completion of development, with intention to begin mass production in H1. So a couple months out, at least. > Micron's memory roadmap for AI is further solidified with the upcoming release of a 36 GB 12-Hi HBM3E product in March 2024. So likely competitive timing with Samsung for 36/12.
- hmottestad 3y agoWent by the title, but yeah, announced is more fitting.
- omneity 3y agoWould this explain why I started seeing so many ads from HBM manufacturers with no product an individual like me can purchase? (I'm not even close to hardware, professionally) Not naming the company but seems like HBM manufacturers might be going all-in to benefit from Nvidia's stock surge.
- fcanesin 3y agoBy the article (March production) new GPU eng. samples should be flying in H1 which means new high memory GPUs coming end of the year. 2025 will be extremely exciting for ML.
- spintin 3y agoThis is an empty article: - CAS latency? - Wattage? Also: https://piped.video/watch?v=2G4_RZo41Zw https://piped.video/watch?v=2G4_RZo41Zw (is this memory same size as 5nm?)
- jdblair 3y agoThe linked press release is a better source, anandtech didn't add much. https://investors.micron.com/news-releases/news-release-details/micron-commences-volume-production-industry-leading-hbm3e https://investors.micron.com/news-releases/news-release-deta...
- mechagodzilla 3y agoWhat is the read latency like on HBM3E?