NVidia CHIPLETS Strategy explained

Urcdkeys Black Friday
25% off code: C25

Affiliate links (I get a commission):

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

#nvidia #chiplets #gpu

[autarchprinceps]

AMD's chief engineer who introduced them to chiplets recently did an interview with GamersNexus about chiplets, and as complementary as he was of course, he also said they realised that chiplets for the compute elements of a GPU do not make sense, because they are far more interconnected than CPU cores. That is why RDNA 3 only has one compute die, no matter its size, unlike Zen which has potentially multiple CCX dies.

[gameguy]

You have consistently suggested a separate RT chip and it has never really made sense even now as you present the industry move towards chiplet based domain specialization. Not that it isn’t desirable, the problem is that it isn’t feasible for the foreseeable future. An RT core is an accelerator that is delegated portions of the RT work by the shaders primarily BVH traversal and intersection testing, when that delegated work is done the shaders then execute the hit shaders, what is critical to this being performant and efficient is that the RT core and the shaders can work on the same data in a shared cache without any need for data movement. Going off die completely destroys this fundamental. RT acceleration even at its most advanced form and in a raytracing only game, RT would be a full round trip of geometry set up -> trace rays -> shader work. Data movement of any kind is going to be an order of magnitude more power draw and latency, and going off die is another order of magnitude ontop of that. You have made no attempt acknowledge or explain away this problem.

[JohnnyWednesday]

When 3D stacking is mature - wouldn't the direct stacking of a GPU onto a CPU offer performance advantages that discreet chips and long length busses can't compete with?

[tringuyen]

NV’s next gen Blackwell GPUs for 2024 will also be monolithic. Intel tried to implement chiplets on Alchemist and Battlemage and failed utterly! AMD is the only one who understands how to partition the chiplets to work as a GPU!

[ATestamentToTech]

Nv will have to do something.. I genuinely believe this will be the last generation of gpu's where they have a clear market lead

[christophermullins]

The way that AMD did 7000 xt is really the only way chiplets can work. L3 cache is the only part of a gpu that can be disaggregated from the main die. Stacking logic circuits is far more difficult than stacking cache like zen3x3d. I suspect we have a few years and a couple of clever techniques to come along before logic stacking is viable. I bet AMD wins next gen hands down while Nvidia invests and develops these new technologies. I suppose Nvidia has more money to dump into stacking so Nvidia will remain the expensive brand and AMD will remain the budget brand.

[anonymoususer]

Coreteks, when used as an adjective, "chiplet" remains singular. For example, "chiplet strategy" and "chiplet GPU".

[MoldyMcdonut]

can't wait for light-based processors. They actually get bigger nanometer-wise but are far more energy efficient and faster, spificially for ai tasks

[kennyj]

I love learning about all the new Technologies but I can't afford the high-end GPU's anymore. My last new 80Ti was the GTX 1080TI. Now I'm picking used last or older generation 80Ti GPU's.

[dalkgamler]

How are you saying that People are ready to pay high prices when it was only bitcoin miners last year that drove the prices?
Now the decrease in the new GPU market is 60% rtx 4080s are nowhere to be sold out (aka the worst launch in 10 years). Scalpers even trying to return huge amounts of these cards.
I don't know how you miss such critical points when doing a presentation about those things.

Edit: why the hell do you have to explain how to use a windows key to your audience, a tech interested one in fact. Do you think we are stupid?

[DragonOfTheMortalKombat]

If nvidia makes chiplets, they'll make even more profits, while we pay big prices as usual.

[tuckerhiggins]

Coreteks, you've been blowing a lot of smoke lately

[Kratoseum]

This has been looming over for a long time. With the advancements in code translators, hopefully legacy support will be good enough for power user's software.
And we're seeing all the big players betting on SOCs, looks like the shake-up is finally starting to materialise.

[Gattberserk]

So we will be seeing monolithic design for at least 2 generation down? Blackwell and Blackwell-Next will be mono? Then BW-Next-Next will be the start of chiplet?

[DanielAirthNelson]

Voice changes are not a direct symptom of chronic obstructive pulmonary disease (COPD), but many COPD patients experience voice changes. Although air acts as a sound barrier when using ultrasound, subcutaneous emphysema may demonstrate by hyperechoic scattered densities.

[Macho_Man_Randy_Savage]

Jensen: "The more you buy, the more I get!"

[granatengeorg]

6:34 guess we now know the secret ingredient of tsmc's success :^) (top left)

[anslicht]

Great analysis as usual. Just to note: A long-term oddity of your narratives is having no gap between sentences. It's weird to listen to. Surely you hear that as well. I understand that you're stitching them together but just leave a little audio space and it will sound much more natural.

[OrjonZ]

3nm GPU will not be here until late 2024 at very earliest. Apple does not have 3nm yet which means 23 will be iPhone 3nm and u know Nvidia or AMD cant compete with Apple for silicon allocation.

[stevensmith]

Nvidia: That's a nice minor advantage you have there. It'd be a shame if someone were to take advantage of it....