DegustatoR
Veteran
Do we consider HBM 2.5D integration "chiplets"?
Nvidia GeForce RTX 4080 Reviews
- Thread starter BRiT
- Start date
Jawed
Legend
Is such a thing a GPU?
DegustatoR
Veteran
A GPU can't work without memory and even has it on die.
Jawed
Legend
Yep, eventually NVidia will put its IP across multiple chiplets to construct a GPU.
DegustatoR
Veteran
So it has to be made with IP from one company? Then it's chiplets?
Jawed
Legend
I'm talking about a GPU (for consumers, by the way), you appear to be trying to redefine a GPU to be something imaginary - good luck with that
DegustatoR
Veteran
I'm asking if HBM being integrated through an interposer (which is a more complex and advanced form of die to die interconnect) and itself being a 3D design made of several stacked dies is considered a "chiplet" design in general. It's not clear to me why it shouldn't be such really.I'm talking about a GPU (for consumers, by the way), you appear to be trying to redefine a GPU to be something imaginary - good luck with that
And the point I'm trying to make is that there are more than one way to do something as "chiplets". This can also be an area where one company would get an upper hand simply through choosing a better design for it's "chiplet" product.
Probably but will it happen soon? According to AMD they were motivated by the diminishing returns of manufacturing cache and interfaces on advanced nodes. The question though is what did they have to sacrifice? Infinity cache latency benchmarks will be interesting.
RTX 4080 unbeliebt: Alternate unterbietet sogar die UVP im Tages-Deal [Update]
Jetzt aktualisiert: Während die RTX 4090 deutlich über der UVP liegt, verkauft Alternate die Geforce RTX 4080 derzeit sogar unterhalb der UVP.
RTX 4080 has dropped slightly under MSRP in Germany
Jawed
Legend
Yes, agreed there's a variety of ways for a product to be a chiplet based product. e.g. Xenos can be described as a chiplet product, it's also a chiplet GPU. The pair of chips implements GPU functionality. You can say that Voodoo2:
Voodoo2 - Wikipedia
is a chiplet GPU, if you want, but integration/packaging isn't in the same mould as current GPUs (and "GPU" was later coined as a term that includes geometry processing, not just pixel processing).
"Chiplet", you can argue is just a marketing term. The history of systems, and system integration is pretty colourful:
and "chiplet" arguably stretches back decades...
AMD's claims about latency:
It's a cost-performance trade-off in the end.
For "data centre GPUs" we've seen NVidia contemplating chiplets as a way to specialise its designs based on sector:
NVIDIA COPA - Composable On-Package Architecture
This seems to be purely for data centre compute: 2104.02188.pdf (arxiv.org) "In this work, we demonstrate that diverging architectural requirements between the HPC and DL application domains put converged GPU designs on a trajectory to become significantly under-provisioned for DL and...
forum.beyond3d.com
With L2 cache in Ada being "finely" distributed across the GPU, it's hard to extract that as a function point that can be separated out into one or more chiplets. The stacking seen for 3D V-cache in Ryzen might be a way to provide localised pools of cache so that NVidia can retain the locality/scale seen in Ada?
It'll be interesting to see whether cache (on consumer GPUs) gets substantially larger than ~100MB as the next generations unfold. It might turn out to be the sweet spot. Similar to the way that 8/12/16/24GB appears to be the segmentation for current performance tiers while it looks unlikely that entry level cards will be 16GB within a decade, if ever.
Scalpers are struggling to sell the RTX 4080 above MSRP, but retailers won't let them return the cards
We're used to seeing graphics cards arrive with such high demand that they're quickly bought in bulk by scalpers and sold on auction sites at hugely inflated...
www.techspot.com
Cry me a river.
Last edited:
It's a weird situation in the UK at the moment. Seems to be loads of really well priced stock of 3070Ti and below, then nothing at all in the 3080/3090 range. Then tons of 4080's at around MRSP, then no 4090's.
Looks like people can see it for the rip off it is.
Looks like people can see it for the rip off it is.
The 4090 benefits from the Cialis effect.
DavidGraham
Veteran
The Newegg screenshot is accurate but the editorialized comment is misleading. A specific Gigabyte 4080 variant is the top-selling model (in whatever time scale they use, I'm not sure). To claim that the 4080 is the best selling GPU you would have to calculate the summation across all brands/models. Still, I suppose the fact that there are 3 4080 models in the top 10 means the card is competing well vs. popular midrange offerings (3060, 6600 etc.).
Amazon sales are a little different. 3060s seem to top the list, followed by a specific 7900XTX. But 4080s do show up in the top 10.
Given that the cheapest 4090 you can buy has been at least 40% higher than MSRP since not long after launch I could see this being true.
DegustatoR
Veteran
NVIDIA RTX 4080 price drops below MSRP in Europe thanks to cheaper US Dollar - VideoCardz.com
US Dollar value is falling and so is the price of RTX 4080 in Europe NVIDIA RTX 4080 is slowly appearing with better prices here and there. However, those discounts limited to certain shops and countries are not part of the official promotion, but rather an impact of US Dollar exchange rates...
videocardz.com
Similar threads
