DegustatoR
Legend
Nvidia has been launching their lineups in 3 to a max of 6 months timeframes for literally years now. I fail to see how this is something new with Blackwell - and as such why it would be related to RDNA4 in any way.
They haven’t released everything in a 3 month period since Pascale I believe - and this is 2 months. The RDNA4 comment is admittedly speculation on my part - might not be a major consideration though surely they’ve considered it.
Last edited:
Wrong. 8 months between 4090 and 4060 launch.
DegustatoR
Legend
Jan-Mar is 3 months.
Turing was Sep-Dec - 4 months.
Ampere was Sep-Feb - 6 months.
Lovelace was Oct-Jun so technically 9 months but it was also the least interesting as a perf/price upgrade for anything below 4070Ti which launched in Jan which would also be 3 months.
I’m not sure I buy that bit of sophistry haha.
In any case even by your own estimate it’s a shorter time frame by a month. Look I never claimed this was a radical departure, just that it’s more accelerated than usual. Also, if you reread my initial post, you’ll see I said that one can also see this as a typical schedule just with the high end cards delayed. Either could be valid based on what we know right now.
(Also, if I’m in January, March is two months away, not three. If you go from early Jan to late March then it’s more like three, but these won’t release until at least the second half of January unless NVIDIA is sending reviewers their cards the day after the announcement.)
Last edited:
DegustatoR
Legend
4060 didn't bring anything new to the market, you could argue that it was even somewhat of a downgrade (due to VRAM size) in comparison to a similar priced 3060. So in this case its launch was rather irrelevant as well as when it had happened. Same is true for 4060Ti and 4070 although the latter can be viewed as a minor upgrade depending on what you compare it to.
This may be more a result of them launching the first SKUs later than originally planned than accelerating the launch of lower end cards. My point is that this in itself isn't a sign of anything about the competition.
That doesn’t negate that they’re new cards! I don’t disagree with your characterization of their value but saying they don’t count as releases is a stretch. It’s not like they’re minor refreshes of the old cards (new node, new functionality, etc).
It might not be - just one possible explanation, even if we may disagree on how likely it is.
Interesting development. The change being mulled here sounds like it would just expand the restrictions to the Middle East. Depending on the election results all GPUs could get a lot more expensive to boot.
www.tomshardware.com
US govt reportedly mulls capping AI GPU exports from Nvidia and AMD to the Middle East
The goal here is to regulate the dissemination of advanced AI technology.
DegustatoR
Legend
Nvidia to ship a billion of RISC-V cores in 2024
Nvidia quietly adopts RISC-V and replaces proprietary microcontrollers.
Intel's former CEO reportedly wanted to buy Nvidia for $20 billion in 2005 — Nvidia is worth over $3 trillion today
A purchase could have drastically changed GPU and AI tech.
DavidGraham
Veteran
So the proposal happened around the time AMD bought ATi? Interesting.Intel's former CEO reportedly wanted to buy Nvidia for $20 billion in 2005 — Nvidia is worth over $3 trillion today
A purchase could have drastically changed GPU and AI tech.www.tomshardware.com
I am guessing the deal would have fell through due to Jensen insisting to be the CEO, NVIDIA was approached multiple times before, but the deals fell through due to this exact reason.
On other news, India is building a gigawat data center and NVIDIA is supplying Blackwell GPUs for it. India is also building another data center which will house tens of thousands of H100s.
I find it interesting because about a decade ago when people were talking about “exascale” (meaning supercomputers at least 1000 PFLOPS, or 1 EFLOPS), the question was that no one thought it’s practical to have a supercomputer requiring more than like 100MW power (otherwise you just pile more GPU, and the largest supercomputer at that time was around 60MW). Now people seem to talk about numbers like “1 GW” as it’s nothing, all just because of “AI”.
Yep the only thing that changed is $$$$$$$.
Two things I would like to point out here:
Ten years ago, a singular computer which could consume an entire gigawatt of power would've been monumental in size, to the point where it might as well be several hundred thousand square feet -- equal to the typical wing of a modern colocation datacenter itself. At such physical scale, it likely would've ended up being a clustered ecosystem of supercomputers, rather than "one supercomputer." Maybe the clustering comment is just semantics, "supercomputer" work isn't single threaded anyway. Today, with the newest commercial GPUs coming in at ~1KW each, we can do 150KW or more in single 42U rack without accounting for cooling. Napkin maths says around 5000 racks packed with GPUs should consume roughly 750MW and sit on ~40,000 sqft of raised floor, then add in aisleways between those rack rows, a campus-wide water loop, the bigass pumps for the chiller system, the thermal plates for "cooling capacitance" in the chiller loop, some quantity of DX cooling for non-loop-chilled workloads (standard CRAC stuff), a room for an operations center, the necessary electrical rooms, UPS gear, and a pair of reactor rooms, and you've probably hit the 1GW mark in a datacenter smaller than 100,000sqft. There's no way we were building this sort of facility a decade ago.
Secondly, a decade ago we had no reasonable way to send an entire gigawatt of power to a single building; there are moderately sized cities that don't run on a gigawatt of power. As a simple example: the Tennessee Valley Authority (TVA) generates and distributes power to all of the state of TN, a substantial portion of Mississippi, the northernmost ~20% of Alabama, the southwestern-most ~20% of Kentucky, and several hundred square miles of Georgia, Virginia and the two Carolinas. As I'm typing this, the weather in this area of the country is around 78*F and sunny, it's about 5:30pm, and TVA is generating 19.2GW for all of their coverage. (Link: https://www.eia.gov/electricity/gridmonitor/dashboard/electric_overview/balancing_authority/TVA) Imagine sending 5% of that power to just a single building? Even today that sounds ludicrous.
Yet today, and for the past several years thanks to EPA authorizations and certifications, it's now commercially possible to drop one (or more) modular fission reactor(s) into a datacenter campus and achieve GW-class baseload at less than $0.20 kWh. This is crazy cheap for self-generation, especially when calculated over the decades-long lifespan of the unit even after considering fueling and maintenance costs.
Money isn't the only thing which has changed, so too has technology localized energy production.
Ten years ago, a singular computer which could consume an entire gigawatt of power would've been monumental in size, to the point where it might as well be several hundred thousand square feet -- equal to the typical wing of a modern colocation datacenter itself. At such physical scale, it likely would've ended up being a clustered ecosystem of supercomputers, rather than "one supercomputer." Maybe the clustering comment is just semantics, "supercomputer" work isn't single threaded anyway. Today, with the newest commercial GPUs coming in at ~1KW each, we can do 150KW or more in single 42U rack without accounting for cooling. Napkin maths says around 5000 racks packed with GPUs should consume roughly 750MW and sit on ~40,000 sqft of raised floor, then add in aisleways between those rack rows, a campus-wide water loop, the bigass pumps for the chiller system, the thermal plates for "cooling capacitance" in the chiller loop, some quantity of DX cooling for non-loop-chilled workloads (standard CRAC stuff), a room for an operations center, the necessary electrical rooms, UPS gear, and a pair of reactor rooms, and you've probably hit the 1GW mark in a datacenter smaller than 100,000sqft. There's no way we were building this sort of facility a decade ago.
Secondly, a decade ago we had no reasonable way to send an entire gigawatt of power to a single building; there are moderately sized cities that don't run on a gigawatt of power. As a simple example: the Tennessee Valley Authority (TVA) generates and distributes power to all of the state of TN, a substantial portion of Mississippi, the northernmost ~20% of Alabama, the southwestern-most ~20% of Kentucky, and several hundred square miles of Georgia, Virginia and the two Carolinas. As I'm typing this, the weather in this area of the country is around 78*F and sunny, it's about 5:30pm, and TVA is generating 19.2GW for all of their coverage. (Link: https://www.eia.gov/electricity/gridmonitor/dashboard/electric_overview/balancing_authority/TVA) Imagine sending 5% of that power to just a single building? Even today that sounds ludicrous.
Yet today, and for the past several years thanks to EPA authorizations and certifications, it's now commercially possible to drop one (or more) modular fission reactor(s) into a datacenter campus and achieve GW-class baseload at less than $0.20 kWh. This is crazy cheap for self-generation, especially when calculated over the decades-long lifespan of the unit even after considering fueling and maintenance costs.
Money isn't the only thing which has changed, so too has technology localized energy production.
The EPA only certified the very first one 18 months ago, so it's gonna be just a bit longer before they find themselves in use.
However, I'm direclty familiar with two datacenters being built inside US shores which are designed around local, modular fission reactors. One of them is using the GE/Hitachi unit; I'm unsure of the other.
However, I'm direclty familiar with two datacenters being built inside US shores which are designed around local, modular fission reactors. One of them is using the GE/Hitachi unit; I'm unsure of the other.
DavidGraham
Veteran
It's official, xAI is expanding their super cluster from a 100K H100s to 200K H100s + H200s, making it the absolute biggest and most powerful supercomputer in the world, for now.
wccftech.com
xAI Is Now In The Process Of Doubling The Size Of Its Colossus Supercluster To 200,000 NVIDIA Hopper GPUs
xAI and NVIDIA brought Colossus online in just 122 days when it would ordinarily take "many months to years" to operationalize such a system.
wccftech.com
Last edited:
Cappuccino
Newcomer
I wouldn’t exactly call this feasible, has this been actually done at all? I’ve seen proposals on paper and endless hype about modular reactors but they all end up in development hell essentially.
Supplying a GW from the grid to a building is incredibly difficult at the distribution level. I don’t want to claim any professional standing here since I’m a relatively new engineer but I am an EE and do something like this for a living, and at best a distribution feeder has a rating of ~550A (or ~13MW at the voltage we use here, 13.8kV, which is more or less common throughout the country give or take a few kV, call it 15kV class). So serving this from a distribution standpoint is essentially impossible, our substation xfmrs are rated at like 50 MVA (for laypeople just consider this MW, which is fairly accurate anyways as power factor is pretty close to .9 at the distribution level anyways due to cap banks) so even if you built out dedicated feeders you would need dedicated transformers.
Of course, you can serve this at the transmission level, which obviously goes a lot higher than a 15kV class distribution feeder. I imagine a lot of these gigantic data centers are fed like this, although frankly I have no idea as I don’t work in transmission.
Oh as far as i've been seeing discussed online , there seem to be nothing at all especially challenging with SMRs.
Their main issue is about bringing the costs down - SMRs still need to comply to all of the ( too exagerated, btw ) safety regulations . So in most practical cases SMRs cost more per MW the the equivalent big reactors.
I guess the few deployments we'll get will happen due to the SMRs getting the adequate amount of subsidies and being able to fit the requirements of a given contract at the right time. My country is going to install on such reactor in a not so distant future
