0 reactions
honoGaming Blackwell Specs: GB202 is 192 SM @ 512-bit and GB203 is 96 SM, according to Kopite7kimi
It is both interesting and irritating that the IT press has focused almost exclusively on the large memory interface with regard to the latest reports on "Blackwell" GB202 by leaker 'Kopite7kimi'. At 512 bits, this is certainly unusually wide for a consumer chip, but this chip specification does not mean that nVidia will actually use this memory interface for gaming solutions. It is also conceivable that the gaming solutions will remain with a 384-bit interface and that the full memory interface of the GB202 chip will only be available in the professional sector.
Even in the professional sector, this memory bandwidth may not be required, but the larger interface enables more memory (64 GB in clamshell mode with 2 GB chips) - and this is simply worth its weight in gold for today's tasks with locally running AIs. In addition to this certainly important point, however, it is forgotten that Kopite7kimi also made an important reference to the number of computing units of GB202 & GB203:
> I think **GB203 is half of GB202**, just like GB102 and GB100. But I don't know if GB202 has a multi chip package.
> Source: Kopite7kimi @ Twitter https://twitter.com/kopite7kimi/status/1767083262512615456 n March 11, 2024
> I am just thinking that **GB203 may only have 6 GPCs**.
> Source: Kopite7kimi @ Twitter https://twitter.com/kopite7kimi/status/176526987560876860 March 6, 2024
All you have to do is compare the leaker's latest statements with earlier reports about the GB202 chip's hardware: The GB202 chip was already described at the time as being equipped with 12 GPCs and a total of 192 shader clusters (SM) - a statement that has not yet been retracted, is often replicated and is therefore apparently still applies.
It is astonishing that the IT press has not drawn the simple conclusion that the GB203 chip should consequently compete with 6 GPCs and 96 SM - especially when the same 6 GPCs were explicitly mentioned by the leaker only last week. This would be much more important information, as it reveals the second-best Blackwell chip (in gaming segment). Of course, all of this could still be incorrect, but this also applies to the information on the memory interface of the GB202 chip.
> "Gaming Blackwell"
> **GB202: 12 GPC, 192 SM, 512-bit interface**
> **GB203: 6 GPC, 96 SM, probably 256-bit interface**
> Source: Kopite7kimi @ Twitter (in various tweets), only GB203's memory interface is not sure
> Note: Ada Lovelace's AD103 is 7 GPC, 80 SM, 256-bit and AD102 is 12 GPC, 144 SM, 384-bit
There was also a widespread misunderstanding of the difference between the graphics chip (GB20x) and the graphics solution based on it (GeForce RTX 50xx). The term "GPU" is often used for this, which is then used for both graphics chip and graphics card - and thus blurs this important difference. However, a GPU called "GeForce RTX 4090" is not actually a graphics chip, but simply a variant of the AD102 chip (AD102-300 to be precise) is used for the graphics card of the same name.
It is unfortunately all too common to read a formulation that (incorrectly) suggests that GeForce RTX 4070, 4070S, 4070Ti, 4070TiS, 4080, 4080S and 4090, for example, are all independent chips - which, as we know, is by no means the case. The distinction is particularly important when a new generation of chips emerges, where all the information revolves around the graphics chips themselves and concrete products are still a long way off.
|clear salvage variant|nearly full chip|100% full chip
|:--|:--:|:--:|:--:|
AD102|GeForce RTX 4090, RTX 5000 Ada, RTX 5880 Ada|RTX 6000 Ada, L40, L40G, L40 CNX|
AD103|GeForce RTX 4070 Ti Super|GeForce RTX 4080, GeForce RTX 4090 Laptop, RTX 5000 Ada Laptop|GeForce RTX 4080 Super
AD104|GeForce RTX 4070, RTX 3500 Ada Laptop, RTX 4000 Ada, RTX 4000 SFF Ada|GeForce RTX 4070 Super, GeForce RTX 4080 Laptop, RTX 4000 Ada Laptop|GeForce RTX 4070 Ti, RTX 4500 Ada, L4
AD106| |GeForce RTX 4060 Ti|GeForce RTX 4070 Laptop, RTX 3000 Ada Laptop
AD107|GeForce RTX 4050 Laptop|GeForce RTX 4060, RTX 2000 Ada Laptop, RTX 2000 Max-Q Ada Laptop|GeForce RTX 4060 Laptop
The data currently offered for "Gaming Blackwell" only shows the underlying graphics chips, from which nVidia will later form the corresponding graphics cards. They can appear in the full version (or close to it) of these graphics chips, but they don't have to - this is nVidia's free decision. At present, only maximum hardware data can be derived from these specifications, but the actual hardware data for the GeForce RTX 50 series is likely to be slightly lower.
Especially in the case of the GB202 chip, the expected "GeForce RTX 5090" could also come out significantly below the maximum specification of the graphics chip. On the one hand, this is due to the fact that nVidia does not have to increase their best gaming solutions very much in this generation due to a lack of competition at the top end. On the other hand, nVidia could reserve the best dies from the GB202 production for the professional sector, where the price ranges are even higher (also due to the AI hype).
Source: translated from 3DCenter.org ttps://www.3dcenter.org/news/news-des-11-maerz-2024 (my site btw)
#technology #hardware 4
source
It is both interesting and irritating that the IT press has focused almost exclusively on the large memory interface with regard to the latest reports on "Blackwell" GB202 by leaker 'Kopite7kimi'. At 512 bits, this is certainly unusually wide for a consumer chip, but this chip specification does not mean that nVidia will actually use this memory interface for gaming solutions. It is also conceivable that the gaming solutions will remain with a 384-bit interface and that the full memory interface of the GB202 chip will only be available in the professional sector.
Even in the professional sector, this memory bandwidth may not be required, but the larger interface enables more memory (64 GB in clamshell mode with 2 GB chips) - and this is simply worth its weight in gold for today's tasks with locally running AIs. In addition to this certainly important point, however, it is forgotten that Kopite7kimi also made an important reference to the number of computing units of GB202 & GB203:
> I think **GB203 is half of GB202**, just like GB102 and GB100. But I don't know if GB202 has a multi chip package.
> Source: Kopite7kimi @ Twitter https://twitter.com/kopite7kimi/status/1767083262512615456 n March 11, 2024
> I am just thinking that **GB203 may only have 6 GPCs**.
> Source: Kopite7kimi @ Twitter https://twitter.com/kopite7kimi/status/176526987560876860 March 6, 2024
All you have to do is compare the leaker's latest statements with earlier reports about the GB202 chip's hardware: The GB202 chip was already described at the time as being equipped with 12 GPCs and a total of 192 shader clusters (SM) - a statement that has not yet been retracted, is often replicated and is therefore apparently still applies.
It is astonishing that the IT press has not drawn the simple conclusion that the GB203 chip should consequently compete with 6 GPCs and 96 SM - especially when the same 6 GPCs were explicitly mentioned by the leaker only last week. This would be much more important information, as it reveals the second-best Blackwell chip (in gaming segment). Of course, all of this could still be incorrect, but this also applies to the information on the memory interface of the GB202 chip.
> "Gaming Blackwell"
> **GB202: 12 GPC, 192 SM, 512-bit interface**
> **GB203: 6 GPC, 96 SM, probably 256-bit interface**
> Source: Kopite7kimi @ Twitter (in various tweets), only GB203's memory interface is not sure
> Note: Ada Lovelace's AD103 is 7 GPC, 80 SM, 256-bit and AD102 is 12 GPC, 144 SM, 384-bit
There was also a widespread misunderstanding of the difference between the graphics chip (GB20x) and the graphics solution based on it (GeForce RTX 50xx). The term "GPU" is often used for this, which is then used for both graphics chip and graphics card - and thus blurs this important difference. However, a GPU called "GeForce RTX 4090" is not actually a graphics chip, but simply a variant of the AD102 chip (AD102-300 to be precise) is used for the graphics card of the same name.
It is unfortunately all too common to read a formulation that (incorrectly) suggests that GeForce RTX 4070, 4070S, 4070Ti, 4070TiS, 4080, 4080S and 4090, for example, are all independent chips - which, as we know, is by no means the case. The distinction is particularly important when a new generation of chips emerges, where all the information revolves around the graphics chips themselves and concrete products are still a long way off.
|clear salvage variant|nearly full chip|100% full chip
|:--|:--:|:--:|:--:|
AD102|GeForce RTX 4090, RTX 5000 Ada, RTX 5880 Ada|RTX 6000 Ada, L40, L40G, L40 CNX|
AD103|GeForce RTX 4070 Ti Super|GeForce RTX 4080, GeForce RTX 4090 Laptop, RTX 5000 Ada Laptop|GeForce RTX 4080 Super
AD104|GeForce RTX 4070, RTX 3500 Ada Laptop, RTX 4000 Ada, RTX 4000 SFF Ada|GeForce RTX 4070 Super, GeForce RTX 4080 Laptop, RTX 4000 Ada Laptop|GeForce RTX 4070 Ti, RTX 4500 Ada, L4
AD106| |GeForce RTX 4060 Ti|GeForce RTX 4070 Laptop, RTX 3000 Ada Laptop
AD107|GeForce RTX 4050 Laptop|GeForce RTX 4060, RTX 2000 Ada Laptop, RTX 2000 Max-Q Ada Laptop|GeForce RTX 4060 Laptop
The data currently offered for "Gaming Blackwell" only shows the underlying graphics chips, from which nVidia will later form the corresponding graphics cards. They can appear in the full version (or close to it) of these graphics chips, but they don't have to - this is nVidia's free decision. At present, only maximum hardware data can be derived from these specifications, but the actual hardware data for the GeForce RTX 50 series is likely to be slightly lower.
Especially in the case of the GB202 chip, the expected "GeForce RTX 5090" could also come out significantly below the maximum specification of the graphics chip. On the one hand, this is due to the fact that nVidia does not have to increase their best gaming solutions very much in this generation due to a lack of competition at the top end. On the other hand, nVidia could reserve the best dies from the GB202 production for the professional sector, where the price ranges are even higher (also due to the AI hype).
Source: translated from 3DCenter.org ttps://www.3dcenter.org/news/news-des-11-maerz-2024 (my site btw)
#technology #hardware 4
source
| 15,000 mlx | |
| 15 | |
| 0 |