AMD 7900XT's CHIPLETS are a bit different

I now stream at:​​

Footage from various sources including official youtube channels from AMD, Intel, NVidia, Samsung, etc, as well as other creators are used for educational purposes, in a transformative manner. If you'd like to be credited please contact me

#AMD #CHIPLETS #NAVI31

[schmitzi]

Daughter doing tasks from Mother to help her with her work. How sweet.

[anonymous-kb]

Chiplet-based designs are very interesting

[TechWechSech]

You're hands down my favorite tech YouTuber. Thanks for sharing the knowledge and possible theories about future developments with us!

[lemisorae]

Oh, i waited for a video about this from you c:!

[Pushing_Pixels]

The scaling for compute oriented workloads would be pretty linear I'd imagine, as long as there was full utilisation of the extra cores. CDNA will be the biggest beneficiary of MCM tech. For gaming, though, it's anyone's guess what the uplift would be. The biggest obstacle might be introduced latency from the bridge die. That die would also consume power too, which would limit the size and number of the chiplets. It's worth remembering that Crossfire used to get pretty close to 100% scaling in some games. It also got zero percent in some games, but that was usually a result of it not being programmed for very well, or at all. That suggests that, with the coherent cache and the absence of any need for special programming, the scaling could be pretty damn good.

[Siuolplex]

Considering how at this rate I might have to wait til the next generation to upgrade, this is very interesting to know.

Even if I dont upgrade next generation, this is still cool.

[tru-bo]

Love your content. Can you do more content on ARM and RISC-V. Recently I read an article about a tiny company called Micro Magic making World’s Fastest and Most Efficient RISC-V.

[opteronprimext]

Thank you for this news, Been a AMD and Intel user for creating a while. Love this

[markhackett]

Making a low end GPU chiplet that is about the size of a CPU chiplet could allow an IO Chip to access, for example, up to four modules, mixing CPU and GPU, allowing a dual chiplet CPU to have one or two GPUs at about the same power as a their lower end dGPU or double (given 100% efficient parallelism), or allowing a low end CPU single chip to be added to a single low end GPU, and for the previous generation or earlier CPUs with this layout or without, a discrete GPU even from the low end would be a gaming improvement, using the same silicon as they would currently be using for their discrete line. Additionally, they could initially roll out the low end smaller chiplet with a CPU chiplet to get APUs out in markets, then roll out the discrete version after their halo and midrange products have been released, then mix around how they allocate their lowest tier GPU chiplet according to product shelf residency.

[bhuvaneshs.k]

Please do a video on CEREBRAS Wafer Scale Integration Technology. They recently released Chip with 2.2 Trillions transistors...

[doctorfresh]

My guy finally upgraded to 4K. Finally you realized that it was 2021.

[jjdizzl]

This has been another fantastic presentation by Coreteks!

[alibizzle]

This makes a good compliment to GN Steve's latest vid. They could do all of this but they probably just give us a 20% performance increase for a 50% price increase

[mohamedelshazly]

AMD have been very creative lately with design for CPU and GPU, utilizing technology to deliver more efficient performance.
I wouldn't be at all surprised if they lead both CPU and GPU markets soon.

[erkkiboy]

One day I will hire Coretex to read a hypnosis script for me. Such a mesmerizing voice!

[zacktrever]

I gave up on AMD and Nvidia

Can't wait for Intel's GPU lineup

[peturstefansson]

The advances in technology are good. But if there is trouble producing sufficient amounts for users, it does tiny goods.

[johnknightiii]

An active interposer is a likely option instead of that bridge chiplet.

[starofeden]

i love that collection in tha glass case

[Eternalduoae]

I'm very skeptical about the rumoured 2.5x perf. Moving to a new node doesn't help with perf per clock, only power draw at a given clock... or achievable maximum frequency.

TSMC's 6nm process is just a refinement of 7nm so i wouldn't expect more than a couple hundred MHz increase possible but looking at the OC perf of 5000 and 6000 series GPUs, we don't see much of an uplift for that, a couple of fps at most.

Secondly, I'm not sure about the economics of this with an 80 CU die +active bridge. I can only see this being achievable with another increase in price points given the increased prices of TSMC wafers.

Like you said, Coreteks, CPU chiplets are small...