Unfortunately advances in nodes are too small nowdays to cover that much more performance improvements, and with MCM design we will see even higher performance jump but at the huge cost of power consumption. Unless they discover something new that will massively decrease power consumption without affecting performance, 1000w PSU will be standard.
-
Competitor rules
Please remember that any mention of competitors, hinting at competitors or offering to provide details of competitors will result in an account suspension. The full rules can be found under the 'Terms and Rules' link in the bottom right corner of your screen. Just don't mention competitors in any way, shape or form and you'll be OK.
AMD Loses Value Crown - Best CPUs of 2021, December Edition
- Thread starter Purgatory
- Start date
Soldato
The top end EPYC CPUs use 9 chiplets and have a TDP of 280W, the new v cache Milan X CPUs also have 9 chiplets with even more cache on them and still have a TDP of 280W. Same with the 3990X Threadripper (although that will draw more power if you tell it to).
I know that TDP and power draw are only loosely analogous, but I think we're a long way away from needing 1000W PSU just because CPU design is moving towards lots of little chips on a substrate.
Not only cpu, but gpu too. With MCM design, gpu will have even higher performance jump between generations than before because yield will be much better so they can put more cores, tranzistors or whatever, but advances in nodes are too small to cover that so power consumption will be massive.
Soldato
I don't understand what you're saying. Smaller nodes generally mean lower power consumption. MCM design also doesn't mean massive power consumption.
Like I said before, AMD have CPUs which link together 9 chiplets and don't break 300W. That's because they are very proficient at creating an interconnect that doesn't require lots of power. In fact, the biggest power draw in Zen 2 and Zen 3 is the IO die which still uses GloFo 12nm. Zen 4 will have TSMC 5nm chiplets and a TSMC 7nm IO die, that alone will see a massive power reduction in the overall design.
Intel, however, have significant issues keeping their power requirements down. That is not a failing of the technology, that is a failing of Intel's implementation.
MCM for graphics cards will be exactly the same. AMD can make great interconnects, so theirs won't draw much power. If AMD have a low power interconnect, but Nvidia have a high power one, then the failing is Nvidia, not MCM. If Nvidia and Intel have low power interconnects in their graphics cards, and AMD as a high power one, the failing is AMD's, not MCM.
And don't think the current power requirements of graphics cards is representative of where things are going, they're not. This is a blip born from Nvidia's arrogance. If Nvidia played the game a little nicer they would've had Ampere gaming cards built at TSMC, and the power requirements would be much lower. Instead, they screwed themselves over and got lumped into using Samsung's objectively worse 10nm node. That gave AMD the favourable optics to push RDNA 2 to be more competitive at the expense of power requirements higher than perhaps originally planned.
Nvidia's Ampere isn't hot and hungry because the design says so, it's hot and hungry because of how it's made. Just like Intel's Rocket Lake too. This doesn't mean everything is always going to be hot and hungry needing 1000W PSUs in the future.
Thats what i said, but problem is we dont have big improvements in nodes like before, for example not so long ago TSMC migrated from 40nm to 28nm, that's huge improvement, today it is from 7nm to 5nm, from 5nm to 3nm, small improvements and very costly, while performance grows equally as before so node improvements can't decrease power consumption equally. And with MCM design they can increase performance even more because yield is better, but more performance you increase, and if nodes improvements can't cover that, power consumption will grow significantly, i hope you understand now.
Soldato
I getcha. Yes, we are in the realm of diminishing returns now, but at least with TSMC their power reductions are still not-insignificant. This is where clever design and Total Board Power is going to come into play, which is why I highlighted the current Zen IO die getting a big shrink (oxymoron?) and AMD's leadership in interconnect design.
Caporegime
TSMC 5nm is 1.8X as dense, so an 82mm^2 Zen 3 direct shrink would be 45.5mm^2, or a direct doubling of cores and cache the die would be 91mm^2.
Power consumption is reduced by 30%, so 105 watt 5950X chiplets would be 73 watts. +15 watts for the IO die would be 88 watts vs 120 watts, or 88 watts + 37 watt chiplet = 125 watts for a 24 core.
You're right to point at the IO die, its a high power consumption part, being GloFo 14nm, moving that to TSMC 6nm they could reduce its size by 70% and bring the 15 watt power down to 5 watts.
Overall you can see just how much a 5nm + 6nm Zen 4 could be over Zen 3, a direct shrink of a 5950X would be 78 watts, now with them moving to an LGA socket and rumoured to bring the power up to about 175 watts you can see how much more CPU AMD are packaging in to that.
We would have to define screwed. Unable to compete on technical merit with AMD and offer a solution to Zen in the next 5 years?
Kick wholesale arse in the graphics card market?
Soldato
Erm.. budget £260 lol. It's the 10400/11400. Weird.
The 1600AF cost 1/3 of that. Just let that sink in.