>>109927620
>LPDDR5x
no matter how many letters they shove into it it still won't make it GDDR7
>LPDDR5X offers peak bandwidth of up to 10.7 Gbps
>GDDR7 offers a total system bandwidth greater than 1.5 terabytes per second (TB/s)
>HBM3e can achieve bandwidths of up to 4.8 TB/s

File: poweringrobotslaves.jpg (139 KB, 1024x1024)

139 KB JPG

>>109927620
I think I want the bifurcation x8 dual 5060 tis.

>350W
>LPDDR5x
And they have the fucking nerve to call it """inference optimized"""
LMAO
WTF is Intlel doing

>>109927632
why can't they just use gooder7 or whatever.

>>109927620
Does it support CUDA?

intel desktop gpu is dead.

>>109927620
what games need fucking 480 gigs wtf is this
maybe use smaller textures

>>109927620
>32 Xe (trans) cores

>>109927632
Not disagreeing with you but why are you comparing lp5x per pin bandwidth to a fully implemented gddr7 and hbm3e setup

>>109927646
Morer expensiver

The newer x7 intel chips are near unobtanium at 14" for under $2k. I just want a good 14" laptop that isn't ewaste

>>109927620
That overly complicated GPU with tiles? It does have the VRAM but it's not going to be very fast.

>>109927844
at max 480GB AI says it should be maybe closer to 500GB/s so less than 16GB GDDR6 but still better yeah. are there official figures?

>>109927620
That's a lot of RAM

>>109928160
With 480gb of lpddr5x it should fall between 512-856GB/s depending on how wide the memory bus on the CPU/GPU is

>>109927620
>fp4
Literally and unironically what is the use case? Is this shit 1 bit sign, 1 bit exponent, 2 bits significand? What the fuck?

>>109928160
>>109928307
>500GB/s so less than 16GB GDDR6
no way.

>>109927632
>>109927645
>LPDDR5X > GDDR7 > HBM3e
When anon understands less than half of the factors.

>>109928259
>That's a lot of RAM
M5 Ultra 512GB preorders opening next month.

>>109928415
>fp4
A block of 32 fp4 values, with an fp8 header. Probably.

>Intel says the design removes 3D and ray-tracing hardware to dedicate more silicon area to Xe cores and AI processing.

The datacenter gpu can't do graphics, it's strictly AI calculations on the 16-deep systolic array XMX

>>109928431
5060ti 16gb has less than 500gb/s bandwidth

>>109928702
It sure is neat that they're using all the fabs in the world solely to produce chips which can only be used for token prediction and nothing else.
I'm sure that won't all turn into e-waste in one year when the AI bubble bursts.
Plus all the fabs will go bankrupt since all their current customers will go bankrupt, so we won't ever get GPUs again.
Really cool!

>>109927620
How to decide if Intel is the right GPU for you!
1. download the onnx source and try to compile it
2. was it fun?
3a. yes? OK! Buy an Intel GPU
3b. No? Smart choice! Stay the fuck away from Intel GPU!

also...
>32 xir cores
lol

>>109927711
GPUs are used for a lot more than just games now.

>>109929853
yea i know stuff like zed editor and some terminals render on the GPU but its trivial

>>109928755
you can have this card make any game you want and then have it available for play back as a video.
3D rendering is obsolete

>>109929913
damn, i wish i was this much living under a rock too

>>109927620
>480 GiB of memory
big if true.

>>109929913
this anon knows.

>>109928755
>Plus all the fabs will go bankrupt since all their current customers will go bankrupt, so we won't ever get GPUs again.
>Really cool!
god please let it happen. we need total death.

>>109927620
Price?

>>109930047
it won't matter because it will be unobtainium if it is even remotely competitive with mac minis with same amount of ram, since this thing would (on paper) shit all over that and I can't fathom what other use you'd have for this meme card if not hobbyist tier LLM shit.

File: 1786464697703277.jpg (151 KB, 1559x731)

151 KB JPG

>>109930025
even with a gpu browser, terminal and editor im sub 2 gig. certainly current games are more demanding but theyre more processing-demanding not memory-volume-demanding
>>109930035
yes

>>109930068
what meme htop fork is this?

>>109930061
It's relatively hard to get an intel flex so I doubt these cards are going to be available for most hobbyists

>>109930074
nvtop

>>109927620
It can work with multiple users LLM inference, or diffusion inference if the compute core is strong enough, well let see

>>109928415
And what's the use case for FP64 on a GPU? Is FP32 not enough?

>>109930333
>FP64
Simulation of shockwaves, apparently.
https://www.youtube.com/watch?v=HlsyZAWnOu8

>>109928702
So it can't be used for 3D rendering and it's basically unusable for render farms? That's a shame.

>>109927632
It depends on how many channels they use, with 384bit lpddr5x they can reach 300GB/s, with 512bit it's 400GB/s. Still four times slower than GDDR7, but with GDDR7 that card would cost three times as much.

>>109927844
he is very obviously a genuine retard whose only familiarity with how memory works is the chatgpt conversation he is regurgitating back to you.

>>109927620
from what I recall, the upcoming AMD 10600 and 10700 will use the same LPDDR5x setup, except they'll have 24CU and 48CU, respectively. only their 10800 or 10900 will have GDDR7.

prices are fucked for that apparently.

>>109927620
Source for this image? I thought the celestial dGPU line-up was confirmed to be canned?

>>109928747
that's 128bit bus I think. bus dictates speed, not memory size.

>>109931556
>the upcoming AMD 10600 and 10700 will use the same LPDDR5x setup
doesn't make sense. those are most likely used for gaming which needs fast memory. if anything I'd expect it the other way around.
linux kernel driver was recently updated for GDDR7, but seen no lpddr5x in there, only some older lpddr5 (not a recent update), which I doubt they'll use.
if I had to guess I'd say they'll use GDDR7 on all of them. upper end ones might have monster bus width, 512bit or even more. at 512bit they can fit 96GB with 6GB chips (next year release) and I saw there might even be 8GB chips but no release date for those, which means a 512bit bus could fit 128GB GDDR7

I bought a used b580 for 190€ recently and like it so far, but it's still hard to recommend with the amount of troubleshooting involved sometimes. CS2 still runs like shit and they will never fix it, and current drivers look more like life support for whenever the newest AAA garbage title comes out.
Really intel just has to fix the software side of things and maybe give an update to xess. But realistically the gpu department is probably lucky to even be alive rn. The only saving grace is that all the other current gen budget gpus are overpriced garbage.
>>109932292
It is canned according to leaks, but allegedly we will see the "Druid" lineup next year. A D770 with a new xess upscaling upgrade would be kino.

>>109927632
That's "10.7Gbps" as in a 10.7GT/s "clock rate", per chip. LPDDR can have full 64-bit interface size or 32 or 16.
Intel saying 480GB card capacity would be 30 16GB chips.
Assuming the 32 bit lpddr standard that's 1.28TB/s card bandwidth which is enough for 15-20tk/s of a 4-bit 600-700B param, 6-10x10B active MoE model w/ 256k context. 25+ tk/s with modern MTP and other nice things.

It's most definitely not bad hardware if you consider AI model improvements are still making rapid advances and you could buy a card like that, and it would still be highly usable 10 years from now.

==powergap==
On a side note if you read just the wiki page for LPDDR it's amazing how we're seeing the interface slowly return to what is essentially a serial connection with caching as a buffer.
>lpddr6
>4 bit CA bus
>bus width 12 bit per channel
>burst rate transfer window method
>256 bit optimal data per frame

DDR8 is going to be a 1 bit CA point to point opti-link serial 40Ghz SERDES on CAMM3 form factor with 2380 pins, calling it now.

>>109932430
>doesn't make sense.
the new Alpha Trion range is used in the upcoming xbox and playstation paired with Zen 6 cores, so they just re purpose that at different core configs as Medusa Halo AI boxes, and also as discrete GPUs. that's one design covering low-range cards, mid-range cards, AI boxes, both consoles, and high-end cards too (3 or 4 different silicon that only differs in CU count). so to be so multipurpose, it has to support both GDDR7 and LPDDR5X.

at which point the choice between gddr or lpddr comes down to whichever they can buy the most of, which is kind of a big deal right now with availability being in the pits.

and there's no way they'll put 128GB on a dGPU, that's MI Instinct card range.

>>109932564
>that's MI Instinct card range.
*older one
newer ones will probably get crazy boost in size. but yeah I do not expect 128GB on consumer shit. at most 96GB, and even that I doubt. most likely it will be 48GB max. and 24GB as replacement for what is now 16GB. 4K gaming can do with more than 16GB these days, in some cases.

>>109932496
>DDR8 is going to be a 1 bit CA point to point opti-link serial 40Ghz SERDES on CAMM3 form factor with 2380 pins, calling it now.

I just hope we'll get a new form factor that will make things change around, ATX is a fucking blight and the main limiting factor in performance and cooling. LPCAMM or SOCAMM would be so nice to have on boards, and 12VO would improve things a lot too. Hell I wouldn't even be against putting memory right next to the CPU die, if the speed is good enough, but HBM is not affordable yet.

motherboards using optical interconnects would be the stuff of dreams.

on the note of going serial, I'm not up to the spec, but isn't parallel sea-of-wires better? one of the reasons non-monolithic Ryzens suck at handling fast memory is because the IO die uses serial connections which uses too much power (distance + constant data encoding/decoding), and they are ditching that shit for parallel connections with Zen 6.

>>109932585
lolno, you'll not see that in consumer graphics cards, probably not in our life time. even if we did, it would cost a car or a house. we are lucky if we can get 24-32GB and that's on high end halo cards like the 5090.

>>109932612
Well the line between "serial" and "parallel" has kind of blurred since PCI-e. Also the in-chip/interchip network is a different thing from the chip-to-RAM interface. the zen IF being a limitation is an entirely diff problem from what the ram interface can/could be.

>>109932631
willing to bet they upgrade from 16GB to 24GB for what is now 9070 and 9070xt cards equivalent, and have a higher tier with at least 32GB at minimum, if not a top tier with 48GB for normie local models. they sell a lot of R9700 32GB which is a 9070xt with 32gb. now today there is a 5090 tier memory size from AMD you can buy it right now at the price of a 5080 or something. 32GB, GDDR6. they will most likely do 48GB, not even joking. 96GB for sure no, but they will do 48GB that's my guess now.

>>109932631
3GB gddr7 is now the production standard with 4-6GB chips "coming"
so 128 bit interfaces will be 12GB cards, 24GB is the new 16GB, etc.
By 2030 gddr7/x should feasible make 128 bit cards with 16GB (but then also by 2030 16GB is likely going to be the "minimum" for max quality 1080p)

>>109932697
Yeah, in datacenter cards.

There's a reason they are trying to use LPDDR5X in discrete gpus. GDDR7 is too hard to find.

>>109929943
>you can have this card make any game you want and then have it available for play back as a video.
not at a playable framerate no you can't

>>109932585
I guess it could happen that all this market pressure on RAM could spark an innovation or revolution in RAM manufacturing to the point that it becomes very cheap. That would be actually really cool.

>>109932739
>There's a reason they are trying to use LPDDR5X in discrete gpus
they aren't. they can't. they'd downgrade the performance, makes no fucking sense. 24GB of lpddr5x would have abismal performance, lower than GDDR6 on rdna3 and rdna4.
you cannot possibly think they'll do such a thing on dGPUs lol. maybe on consoles, steamdeck thing, whatever. they are not replacing 9070xt with a newer more powerful model but with 24gb lpddr5x lol wtf it would totally ruin it's gaming performance. gddr6 was already lower performance in some games compared to gddr7 which has the edge in game performance in some cases.
not sure where the lpddr5x conspiracy theory comes from, and why, or why you're trying to push it. some kind of anti-AMD campaign for upcoming cards or?

>>109927620
? I'm a gamer, I don't care about ai shit

>>109927620
Why are intel drivers still shit with modern llms? Can't they just vibe code drivers on par with nvidia's?

>>109932799
>they aren't. they can't.

Bitch, you'll be lucky if they don't end up shipping with DDR4. Did you not look at the markets? What you see on the consumer market is just a drip in the ocean, the supply for OEMs is way worse.

>>109932799
>24gb lpddr5x lol wtf it would totally ruin it's gaming performance

they can just turn on 8x frame generation.

>>109933046
They don't care to keep their shit up to date, it targets one stack with select refinements on a set date and they only update for needed stack revisions.

File: maxresdefault.jpg (118 KB, 1280x720)

118 KB JPG

>>109928573
>M5 Ultra 512GB preorders opening next month.
stop posting about it i don't want anyone else to know retard

>>109932465
>Playing games
Grow up

>>109932806
That's why you are overlooked by women

>>109936682
but my waifu is always there for me, and has been for years and years

>>109933077
plenty gddr6 chips, they put them in r9700. they'd ship with gddr6 and it would still be faster than ddr5
this seems like buthurt nvidiods trying to demoralize AMD chads

amd is gayer than (you)

>>109927620
It would probably be usable for the massive low active params MoE.
If it doesn't cost an arm and a leg (it will) and is actually available (it will not) it could be a good option.
But it won't.

>>109942677
checked