>432GB VRAM
Why not AMD?

>1TB VRAM
>100pf fp4
amd btfoed once again

>>105578808
>1 GPU vs 4-14 GPUs
nJudea saar…. you must into the reading

I don't think that's one GPU, it's a cluster similar to an H200 cluster or whatever

You're gonna have to win the lottery if you want one, and rewire your house
Also it'll be very loud so you'll need it to be in its own soundproof room with dedicated AC

>>105579042
Try using google, nJeetia bitch
> In addition to the compute capability, AMD is also going to leverage HBM4 memory for its Instinct MI400 series. The new chip will offer a 50% memory capacity uplift from 288GB HBM3e to 432GB HBM4. The HBM4 standard will offer a massive 19.6 TB/s bandwidth, more than double that of the 8 TB/s for the MI350 series. The GPU will also feature a 300 GB/s scale-out bandwidth/per GPU.
https://wccftech.com/amd-instinct-mi400-accelerator-doubles-compute-40-pflops-432-gb-hbm4-memory-2026-launch/amp/

>>105579042
>Rewire your house. You're going to need to talk with the city and your local energy company. Pretty sure these things are full height rack units with power draw in the hundreds of kilowatts.

>>105579042
1600W is nothing on an EU/UK socket kek
I can run this shit on a fucking powerstrip rated up to 3000W

>>105579042
>rewire your house
Our outlets are rated for continuous 10 A, so 2.3 kW should be fine.

>>105579042
lol @ mutts
ima just plug this into an ikea extension cord

8%
>8%
8%
>8%
8%
>8%
8%
>8%

>>105579117
we’re not talking about desktop Rakshit

>>105579060
crazy, glad to be wrong
I thought single gpus hadn't cracked 200gb yet

>>105579270
now u know
research before u reply, bitch

fp4 performance, why not fp64? What's the performance of that? Even that's fairly low precision.

>>105578764 (OP)
>>105578808
how much $$$?

>>105578764 (OP)
>Nvidia minus 500 bucks

>>105579042
American, please do not include us into your sick humiliation fetish.

>>105578764 (OP)
okay-okay, but can it run native 4k raster in 60 fps without framegen and upscaling tho?

>gaming bad
>ai gooning good

>>105578764 (OP)
I don't want zluda

>>105580513
Probably. 8x32 CU is immensely powerful.

>>105581342
They're putting 8 of these things side by side per tray too. Fucking crazy.

>>105578764 (OP)
>amd
>getting on the ai train
not happening

>>105581590
Based

>>105580513
no rops
can't rop

>>105579032
The more you buy... the more you save!

sneed

>>105579104
>>105579073
>>105579089
>>105580512
Ah, so THAT is why EU countries keep having outages ...
And they are as smug as ever about it

>>105578808
>second half 2027
yeah heh tough luck AMD but it appears that adding an extra generation of development has resulted in a more powerful design

>>105579060
>>105579032
AMJeet, you must be retarded to think that MI400X is going back to a single GPU when MI355X is already 2x.

>>105584136
what are you blubbering on about?
my last powercut was a thunderstorm in 2020

>>105586555
Don't sabotage the cope of your american brethren.

>>105578764 (OP)
>>105578808
But can it run crysis

>>105588459

>>105578764 (OP)
>Why not AMD?
How to tell the world you have never actually worked with ML workloads in three simple words.

>>105590848
but he uses amd

>>105578808
You only need 400gb to run deepseek

>>105591456
Is this a bait?

>>105578764 (OP)
I need more
UE6 will need 433GB VRAM