You guys came up with the limits, I want it as an option on all. Ideally with the chiplet system I'd like to see a point where we just order one to four CCXs, one IO die to fit our system and whatever is left that will fit on the package, APU, GPU, HBM, etc...
Imagine if AMD made a shield competitor with there APU and it had HBM stacked on die 😱 Nvidia would be 💩ing 🧱's as they couldn't compete the Tegra has been refreshed more then Intel's 14nm ++++++++++ node.
@@shadowarez1337I won't go as far as trashing other companies, but... Imagine the 8700G has at least 4GB of 2Gbps HBM2 via 1024-bit memory bus. The Radeon 780M as well as the built-in NPU might have a massive boost as long as the built-in HBM didn't fill up. But it's just my own guess and I could be wrong.
@@timytimotius4679 it'd have to be a custom SoC the APU might cost too much but the addition of HBM and a powerful CPU compared the Sabertooth Tegra SoC
Bring these to laptop APUs please! No more dedicated GPU on laptops for AAA gaming is the future. We need a 32-40CU APU with integrated 16GB HBM. Now that would finally bring AAA gaming on thin and "on the go" laptops.
I see your point, but doing this will simply destroy repairability: no more replacing a cpu, gpu, or even ram when a computer breaks. Plus, 16GB of RAM is starting to not be sufficient for many use cases, but 16GB HBM don’t come cheap. At that point, you may be better off buying a gaming laptop with a much faster CPU and GPU for the same price as this supposedly “thin and light” device that somehow have to manage cooling 32~40CU (~100W of power, referring to data collected from my undervolted 7900xtx at 2.3GHz) of GPU power
They don't strictly NEED hbm2, MLID has talked (a lot) about Strix Halo coming in early 2025 (delayed from mid to late 2024) up to 16 Zen 5 cores 40 RDNA 3.5 CUs and an XDNA 2 NPU. Should be a solid chip even with DDR5 memory.
although a powerful APU would be cool; a dedicated GPU makes a lot of sense, since it allows more flexibility and is better for thermals bc it is in a physically different location from your APU, reducing stress on the cooling
@@the_spicy_orange Yeah strix Halo will have 40CUs. I am curious to see how it will perform. But the bandwidth of ddr5 is a limiting factor. Its still much lower than gddr6 even
Most consumers in those AMD dev kit video were just checking out if they should buy more stocks or not 😂 They have no idea what HBM is. Even no idea what this board is used for.
@@user-f1ff23 Off course! I play at 2k always. Is not the best card in the world, but is strong. But i have 60hz limit, so it's good. Play at 60hz, because my rig is from 2016-2017. But i only playing Skyrim modded and Fallout 4 modded. i7 8700k, 32gb ddr 3000 and RX Vega64 8GB HMB2. It's old, but powerfull.
Since RX 5700 onwards, Vega is not even a good used buy except below 100 USD. RX5700 can be had for about that, draws half the power for more fps and has ongoing driver support. Thats coming from someone who screwed the biggest air cooler on his Vega56 and made it run stockV64 clocks, undervolted.
I am not sure this is done with 2 channels PCs in mind, maybe it is more suitable for Threadripper Epyc or Xeon machines. However, Thanks Very Very Much for sharing this with me! 5950x + 6800xt (with custom loop!) 😂
@@zepandza8479you better don't, becasue after RDNA2 cutting costs strategy on consumer products just tell your sensible brain and sentimental hearth to go buy RTX 40 and some DDR4 intel 12th gen or Zen3 (5500 to 5700x) instead of whatever new unfinished product with difficult name to decode is being launch, that or 7800x3d (or 7950x for wrokstation like or 7500f/7600x if you want to feel faster than 5800x3d but no need to spend accordingly for ultra fast gpu, as they're since rx 6800XT or 3080, too much for current cpus and gaming)
Tackle Memory Bottlenecks with the quad chanel, octa channel and dodeca channel DDR4 and DDR5 starting at consumer level.... boom 200%/400%/600% bandwith over 64 bit and lower latency (¿-50%?f easy) or cheap, way cheaper than ridicoluos XPO ram costing more than the cpu or the gpu...
After learning to work with FPGA's at campus, i wanted to try making a GPU with one as a hobby. UA-cam recommended me *this* video, and it really looked like something i could use for that. Then i looked at the price... I'm a student, not a millionaire! 😭
AMD Time Machine went back to 2017 in 2024 after launching most bottlenecked APUs for 7 years and according to roadmaps more years to go... a bit confused what year is in AMD R+D+i labs 2016 Nvidia Turing server and computing, 2020 Nvidia Ampere and server Tesla, 2022 Grace Hopper without ARM cpus... they need to remember IBM invented "Infinity cache" used in RDNA2 and CDNA2 below levels suggested of IBM (256mb) and diminissing in RDNA3/CDNA3.... shrinking bus bit while using expensive overclocked ram no jedec spec'ed...Does time goes in several directions back there? Hello, 2024 here, you're Welcome. I'm talking to you the people who invented X3D and Vega GH (gpu with hbm put in the same board of a socket cpu)...
It's RAM integrated in the motherboard along the CPU/GPU, so they can put them directly next to it and with more connecting lines than regular RAM. It also saves space. The new thing here is the way they are connected. They have higher density of connections and each controller has more connections to the memory chips. All this is for data-centers, heavy AI workload and the like. They said a few more things, but I didn't understand them hehe.
This is fan****ingtastic! We need more HBM
Good job AMD. I'm happy to see you growing. Always in my heart
Thanks!🙌❤️
@@AMDcan we have a machine learning up scale like other side? I mean amd will have matrix in their gpu why not use them for ai upscale
Let's have some of that as an L4 for our CPUs.
*laptop CPUs
Or desktop APU.
You guys came up with the limits, I want it as an option on all.
Ideally with the chiplet system I'd like to see a point where we just order one to four CCXs, one IO die to fit our system and whatever is left that will fit on the package, APU, GPU, HBM, etc...
@@timytimotius4679Just APU's as a whole. They NEED it.
@@guidedorphas10any cpu you Neanderthal
I wonder will it be used in consumer space again eventually, though a much faster desktop APU would be awesome.
Imagine if AMD made a shield competitor with there APU and it had HBM stacked on die 😱 Nvidia would be 💩ing 🧱's as they couldn't compete the Tegra has been refreshed more then Intel's 14nm ++++++++++ node.
@@shadowarez1337I won't go as far as trashing other companies, but... Imagine the 8700G has at least 4GB of 2Gbps HBM2 via 1024-bit memory bus. The Radeon 780M as well as the built-in NPU might have a massive boost as long as the built-in HBM didn't fill up.
But it's just my own guess and I could be wrong.
@@timytimotius4679 it'd have to be a custom SoC the APU might cost too much but the addition of HBM and a powerful CPU compared the Sabertooth Tegra SoC
Bring these to laptop APUs please! No more dedicated GPU on laptops for AAA gaming is the future. We need a 32-40CU APU with integrated 16GB HBM. Now that would finally bring AAA gaming on thin and "on the go" laptops.
I see your point, but doing this will simply destroy repairability: no more replacing a cpu, gpu, or even ram when a computer breaks. Plus, 16GB of RAM is starting to not be sufficient for many use cases, but 16GB HBM don’t come cheap. At that point, you may be better off buying a gaming laptop with a much faster CPU and GPU for the same price as this supposedly “thin and light” device that somehow have to manage cooling 32~40CU (~100W of power, referring to data collected from my undervolted 7900xtx at 2.3GHz) of GPU power
you can only fit so many CU's on a apu. the 780m is a start to it but i doubt you'll see more then 20+ cu's on a apu soon.
They don't strictly NEED hbm2, MLID has talked (a lot) about Strix Halo coming in early 2025 (delayed from mid to late 2024) up to 16 Zen 5 cores 40 RDNA 3.5 CUs and an XDNA 2 NPU. Should be a solid chip even with DDR5 memory.
although a powerful APU would be cool; a dedicated GPU makes a lot of sense, since it allows more flexibility and is better for thermals bc it is in a physically different location from your APU, reducing stress on the cooling
@@the_spicy_orange Yeah strix Halo will have 40CUs. I am curious to see how it will perform. But the bandwidth of ddr5 is a limiting factor. Its still much lower than gddr6 even
Consumers in this comment section won't understand that using HBM is expensive stuff, and they will end up not buying such products
They can make an Halo product much like how the 4090 is or how the 5090 will be.
Most consumers in those AMD dev kit video were just checking out if they should buy more stocks or not 😂
They have no idea what HBM is.
Even no idea what this board is used for.
RX Vega 64 here :)
Are you gaming on it? Is it still good?
@@user-f1ff23im using a refurbished rx vega 64, it can still game
@@user-f1ff23 Off course! I play at 2k always. Is not the best card in the world, but is strong.
But i have 60hz limit, so it's good. Play at 60hz, because my rig is from 2016-2017.
But i only playing Skyrim modded and Fallout 4 modded. i7 8700k, 32gb ddr 3000 and RX Vega64 8GB HMB2. It's old, but powerfull.
Since RX 5700 onwards, Vega is not even a good used buy except below 100 USD. RX5700 can be had for about that, draws half the power for more fps and has ongoing driver support. Thats coming from someone who screwed the biggest air cooler on his Vega56 and made it run stockV64 clocks, undervolted.
Vega 64 memory almost as fast as my 4090s... HBM is just better
Would be cool if they can add this to GPUs in some type of way.
Do you mean like they (and others) used to do in 2017 (other 2016)?
@@InternetListener to a certain extent. Take advantage of the technological changes and progress.
Too expensive.. had it in my Fury and Vega64. We don't need this capacity in gaming.
add what exactly?
@@Mottbox since AMD is adding more AI to their chips the memory bandwidth may come in handy.
Could you please consider developing GPUs that are compact enough to fit in NVMe or M.2 slots? That would be epic.
Advanced Micro Devices for a reason. 🤷🏻♂️ 👏🏻
So basically an FPGA with HBM, nice!
By "hardened HBM" switches, is this the same internal memory encryption feature in Epyc processors?
Going to be made in Malaysia?
Nice.
1275ns of memory latency (read) or 355ns of memory latency (write) - why so bad??? Consumer CPU's manage to stay
This is an FPGA, they run at far lower clock frequencies so the long latency doesnt matter.
was this video produced in 2013? what is going on at amd???
AMD should make an HBM add-on directly to their CPUs, integrating with the CPU similar to Intel's Omni-path interconnect.
NICE
+amd *Is 4096-bit memory qualifiable?*
I am not sure this is done with 2 channels PCs in mind, maybe it is more suitable for Threadripper Epyc or Xeon machines. However, Thanks Very Very Much for sharing this with me!
5950x + 6800xt (with custom loop!) 😂
Looks fun, but $15K! Can we please have a sub $5K PCIe card please AMD?
I would love an X3D zen 5 with HBM
Im fine paying a premium aa long as its got 64GB flat minimum
AMD YES!
Impressive
AMD we want strix halo in an SBC form factor
Hbm is back?
Never left, just not in consumer products
@@TheLoki1983 I never looked outside the consumer field
@@zepandza8479you better don't, becasue after RDNA2 cutting costs strategy on consumer products just tell your sensible brain and sentimental hearth to go buy RTX 40 and some DDR4 intel 12th gen or Zen3 (5500 to 5700x) instead of whatever new unfinished product with difficult name to decode is being launch, that or 7800x3d (or 7950x for wrokstation like or 7500f/7600x if you want to feel faster than 5800x3d but no need to spend accordingly for ultra fast gpu, as they're since rx 6800XT or 3080, too much for current cpus and gaming)
A handful of AMD instinct AI & HPC accelerators have them for quite a long time and still continue to have them even with the latest HBM3 standard.
Let's solder this onto our GPUs and run the fattest LLMs :)
nice
Tackle Memory Bottlenecks with the quad chanel, octa channel and dodeca channel DDR4 and DDR5 starting at consumer level.... boom 200%/400%/600% bandwith over 64 bit and lower latency (¿-50%?f easy) or cheap, way cheaper than ridicoluos XPO ram costing more than the cpu or the gpu...
After learning to work with FPGA's at campus, i wanted to try making a GPU with one as a hobby.
UA-cam recommended me *this* video, and it really looked like something i could use for that.
Then i looked at the price... I'm a student, not a millionaire! 😭
AMD Time Machine went back to 2017 in 2024 after launching most bottlenecked APUs for 7 years and according to roadmaps more years to go...
a bit confused what year is in AMD R+D+i labs 2016 Nvidia Turing server and computing, 2020 Nvidia Ampere and server Tesla, 2022 Grace Hopper without ARM cpus... they need to remember IBM invented "Infinity cache" used in RDNA2 and CDNA2 below levels suggested of IBM (256mb) and diminissing in RDNA3/CDNA3.... shrinking bus bit while using expensive overclocked ram no jedec spec'ed...Does time goes in several directions back there? Hello, 2024 here, you're Welcome. I'm talking to you the people who invented X3D and Vega GH (gpu with hbm put in the same board of a socket cpu)...
AMD® 9000 series and 10000 series processors has officially support of 16K resolution(15360 × 8640) and H.266(VVC) video codec.
should be used in consumer cards again,they actually need it now
It seems like the CPU should switch to this kinds of transistor.
When AFMF for rx 5700 xt?
My first gpu was an r9 380
rx 580 here
Bruh.... Last year you could make 716GB/s out of 8x GDDR chips, where are you getting 24 chips from?!?!?
Please mass production this one and we can buy one for fun😊
พูดได้คำเดียวเป็นไปได้ข้อข้างยากมากกดไปขึ้นเป็นผู้บริหารบริษัทความรับทางเทคโนโลยีอย่างนี้อยู่ในกรอบสุดๆไปเลย
What an expensive tech demo.
Xilinx payoff
can someone explain this in a more simple way?
It's RAM integrated in the motherboard along the CPU/GPU, so they can put them directly next to it and with more connecting lines than regular RAM. It also saves space.
The new thing here is the way they are connected. They have higher density of connections and each controller has more connections to the memory chips.
All this is for data-centers, heavy AI workload and the like. They said a few more things, but I didn't understand them hehe.
thanks!@@Mr.MasterOfTheMonsters
819 GB/s only? Meh.... I expected at least 1 TB/s
BAGUS.
NUSANTARA IBU KOTA INDONESIA.
SEGO JAGUNG INDONESIA.
TOKO CHINA INDONESIA.
🎉🎉🎉🎉🎉🎉🎉🎉.