Following the rumors of the month, the new NVIDIA Blackwell has been revealed to me at the end of the self-Jensen Huang, and will affect, significantly, not only in the market of Artificial Intelligence, with the çipat the most powerful of the building, sometimes for this purpose, but the one video game, based on the options the customer (let's be clear, it is substantially different, at the level of the structural, which is born of him later on in the coming months.
The first three products to dominate the market
In the case of the GTC 2024, the CEO, with the xhaketën of the skin revealed the more recent creations of his own house, and the two GPUS, and a “superçip”, which combines a CPU's Grace, with the two GPU Blackwell on a platform of single.
There are three names that need to be taken into consideration, namely, B100, B200, and GB200, with the last two, which represents the pinnacle of the pyramid of the performance of the new generation of çipave with the color green.
Blackwell, represents the state of the art on the accelerator, and HE and introduces the technologies, current manufacturer tajvanez the çipave TSMC in a limit order to achieve the beauty of the 208 billion transistors. GPU B200, in particular, would in fact be an accelerator of the çipeve, and I will see my face on the two machines for the GPU to be one, to be united by a bond 10TB/s s s s s NV-HBI chip-to-chip.
As it suggests, tipsteri of the well-known kopite7kimi, and the number of transistorësh, it was thanks to the nyjës “order” TSMC 4NP, upon which it is based and architecture-Blackwell, which represent only a upgrade of the nyjes 4N 5 nanometers to use now on Ada Lovelace. but with the improvement of 30% in density.
The efficiency of the new generation, so it would depend largely by factors such as architectural, and not by a change of the nyjës, the properties of which may have been the only influence marginal in comparison with what has already been observed in the Hopper, and the Ada.
It is exactly in contrast with the Hopper and, in particular, with the H100, the greatest benefit in the progression to a state of the chiplet is the very change of the significant rate of between 80 billion tranzistorëve of the “old” and 208, 104 current on any machine. B200.
A system with a lot of chipsets allow you also to go further in the front part of the memory the sun is connected with four module to the memory 24 GB HBM3E, for a total of 192 GB on a bus (8192-bit.
NVIDIA does not shtjelloi a lot of the numbers, providing an overview, not a complete, but no less interesting to the supply of the new generation to databazat/HPC. The options that will come up on the market and will have a TDP of 700W and 1000W (i.e. B100 and B200), taking into account that the H100 stopped at 700W.
In terms of performance, the estimates of the NVIDIA suggests that the architecture of the cloud will be at least two times more powerful than the Hopper, and now there are clients of the calibre of the Spot on the door, and what awaits çipat, the first by the end of 2024.
In general, the Blackwell could hit 20 petaflops in FP4, and the 10-petaflops on the FP8. Taking into account that it is not possible to be operated with a precision FP4 in the Hopper, and now the FP8 Blackwell provides more than twice the performance.
Member of the improvements, in addition to the very design and architecture, and is the second generation of the Transformer the Engine, which adds support for 4-bit, increasing the capacity of the calculator, and the size of models that can be managed.
In the end, it will be interesting to see how the market would accept the possibility of operating with the precision is so low.
Discussion about this post