NVIDIA is busy finalizing the development of its new Blackwell architecture, which is the successor to the current Ada Lovelace and which will give life to the RTX 50. But, apparently, things haven’t been going as well as expected, as a new report suggests that doubling the performance as predicted may be more complicated than just putting the numbers in a box. And that is, for now, both GB102 and GB100 are not in compliance. What happens to the greens? What’s your problem with RTX 50?
It was logical that there would not be that big increase in performance as expected. And it is that the figures, if we remember, were leaked with nothing less 2.6x Regarding the current graphics card. A never-before-seen jump that appears to be weak from the latest information available. What’s new now?
NVIDIA is having trouble delivering a significant performance boost on the RTX 50
Since it seems that raster performance and ray tracing performance already need to be separated, the information is confusing. and it’s precisely because of the much better performance talk with ray tracing where the intent is to make a new pipeline ends with noise problem In both this and super sampling.
Similarly, earlier engineering samples confirm that the expected performance has not been achieved since the architecture, this time, it will be brand new, The jump forward seen between Pascal (GTX 1000) and Turing (RTX 20) is closer to what we’ve seen in the latter series, so it’s far from doubling the performance or even coming close . ,
For now, NVIDIA is working on two of its main chips, which will reportedly be different, mainly because GB100
There will be an architecture “MCM“(dual chip, ie two GB102 in one interposer) and GB102 shall be monolithic traditional.
Few benefits away from theory
And it is that reality always wins the game. The GB102 with the Blackwell architecture is said to have a design between the 144SM and 142SM, with 12 VRAM controllers with 384 bus bits, 96MB L2 and support for PCIe 5.0 x16, not to mention the new tensor core. and RT Cor.
Also, the frequencies would be around 3 ghz as standard thanks to node 3 nm d tsmc In its high performance version. In the absence of seeing architecture changes to implement the new graphics pipeline, nothing suggests that performance will be at least double that of the RTX 4090 with the current AD102.
rather Jumping over +30% If possible, it doesn’t seem feasible to add more muscles to each chip individually. not surprisingly, he GB100 as predicted 256 sm Thief 5.120 bits For hbm3 why 128 MB D L2So there isn’t much room for improvement in the minimum units, possibly because the density jump doesn’t leave much scope at the moment.
In any case, we’re talking about first engineering prototypes, so there’s still plenty of time for improvement between now and next year (at least, 2025 and beyond), but we’re very pleased with what we’ve seen. Not optimistic.