Follow Tom's Hardware on Google News , or add us as a preferred source , to get our latest news, analysis, & reviews in your feeds.
Anton Shilov Social Links Navigation Contributing Writer Anton Shilov is a contributing writer at Tom’s Hardware. Over the past couple of decades, he has covered everything from CPUs and GPUs to supercomputers and from modern process technologies and latest fab tools to high-tech industry trends.
bit_user The article said: Micron claims that in memory-constrained Spark Support Vector Machine (SVM) analytics workloads Wow, who even uses SVMs, any more? Neural networks started outperforming them like entire decades ago, and you don't even need a very big or costly ones to do it. BTW, I always find it interesting to look at the ratio of system memory capacity to bandwidth. Essentially, it tells you how many seconds it takes to actually read or write all of that memory. I have a theory that, when it exceeds a certain point, the performance benefit of adding more RAM vs. swapping to fast SSDs loses to the cost-savings and potentially higher capacity of the SSDs. The transition point should vary by application, but I expect that even in-memory databases struggle to justify more RAM, beyond a certain point. At the high end of the curve, you end up in mainframe territory, because you need lots of uptime just to fill the RAM, let alone actually get the benefits of reading it vs. NAND. Reply
Fomdoo Just so people can go ask ChatGPT why memory costs so much. Thanks, AI bubble. Reply
bit_user Fomdoo said: Just so people can go ask ChatGPT why memory costs so much. Thanks, AI bubble. I honestly don't know how much demand there will be for these. Due to the use of TSVs, they seem to require the use of advanced packaging, which is probably still somewhat of a bottleneck, making them rather costly. If you run the numbers, a 256-core/512-thread CPU with 16 of them would have a whopping 16GB of memory per vCPU. That seems like rather a lot, to me. As the article mentions, 512GB DDR5 DIMMs aren't new and there hasn't seemed to be a lot of demand for the older ones. Micron claims these are 50% faster and ??? more efficient, but I'm not sure that's why Samsung's didn't see a lot of uptake. Clearly, Micron's market research people thought there was a worthwhile market for these, yet I don't expect them to be the memory of choice for most new servers. For one thing, I'm a little surprised that they seem to be conventional RDIMMs and not MRDIMMs. Reply
usertests bit_user said: As the article mentions, 512GB DDR5 DIMMs aren't new I wandered in here to talk about that, without looking at the article first, but you've said it all better than I could. These should be cheaper than MRDIMMs while delivering decent bandwidth, so maybe there's something to it. Gonna drop this in here since I thought it was overlooked: Intel Nova Lake 12Xe reportedly supports upgradable DDR5-8000 CSO-DIMM memory Reply
bit_user usertests said: These should be cheaper than MRDIMMs while delivering decent bandwidth, so maybe there's something to it. Not sure about that. I think the main difference between RDIMMs and MRDIMMs is in the buffer chip (i.e. the register responsible for the R in RDIMM). Compared to the DRAM itself, I think the cost of those buffer chips isn't significant. Perhaps the reason why Micron opted to cite regular RDIMMs is because the energy-efficiency case they were trying to make is better supported by them than MRDIMMs. usertests said: Gonna drop this in here since I thought it was overlooked: Intel Nova Lake 12Xe reportedly supports upgradable DDR5-8000 CSO-DIMM memory But, since then, it came out that 12Xe got dropped from the Nova Lake lineup. https://www.tomshardware.com/pc-components/cpus/intel-reportedly-cans-12xe-option-for-nova-lake-s-desktop-gaming-apu-design-said-to-resurface-with-razor-lake Anyway, I guess it's nice if SODIMMs have a bit longer to live, before being fully displaced by LPCAMMs and SOCAMMs. I still expect SODIMMs to eventually disappear, and they'll probably go before full-size DIMMs. Reply
usertests bit_user said: But, since then, it came out that 12Xe got dropped from the Nova Lake lineup. https://www.tomshardware.com/pc-com…-apu-design-said-to-resurface-with-razor-lake The VC article is referring to the mobile version, which should still be coming. I'd expect it to be basically just a drop-in slight improvement on Panther Lake w/ 12Xe / B390 graphics, i.e. the Intel® Core™ Ultra X9 Processor 388H. Reply
3ogdy I hope all these servers will end up burning to the ground entirely. For the water pollution, for the crazy prices they've caused in the hardware market, for the electricity demand. For everything that's wrong with them. Reply
usertests 3ogdy said: I hope all these servers will end up burning to the ground entirely. For the water pollution, for the crazy prices they've caused in the hardware market, for the electricity demand. For everything that's wrong with them. 512 GB RAM modules existed long before you cared about AI, and can be used in any kind of server. The water pollution is a myth btw. Unfortunately, your hope will have no impact. Good luck with the arson career. Reply
magbarn So this is where all the Crucial memory went that was supposed to go to the warranty department. Instead you get a $100 check for your 32gb of failed ddr5 you bought on black Friday 2024 that you mailed in at your expense. Reply
Key considerations
- Investor positioning can change fast
- Volatility remains possible near catalysts
- Macro rates and liquidity can dominate flows
Reference reading
- https://www.tomshardware.com/pc-components/dram/SPONSORED_LINK_URL
- https://www.tomshardware.com/pc-components/dram/micron-announces-512gb-ddr5-9200-memory-modules-with-16w-power-draw-up-to-12tb-per-server-claims-60-percent-less-energy-intensive-than-four-128gb-modules#main
- https://www.tomshardware.com/membership
- New York State recommends demanding AI data centers pay $1 million in community investment per megawatt — framework advises towns to plan for maintenance costs,
- Acer Swift Air 14 review: Wildcat Lake and lots of ports at $699
- Poland lost $230 million in cryptocurrency trying to buy Venezuelan oil in 2023 — USB drives with crypto handed directly to scammers
- Chinese state media counters Anthropic's call to put brakes on AI development — paper says move is ‘a response to Chinese competition’
- Solo dev enables running CUDA on AMD hardware in Windows, getting multiple CUDA libraries running on a gaming Radeon RX 9060 XT GPU in Windows — CUDA-exclusive
Informational only. No financial advice. Do your own research.