Monday, August 14th 2023
NVIDIA Blackwell Graphics Architecture GPU Codenames Revealed, AD104 Has No Successor
The next-generation GeForce RTX 50-series graphics cards will be powered by the Blackwell graphics architecture, named after American mathematician David Blackwell. kopite7kimi, a reliable source with NVIDIA leaks revealed what the lineup of GPUs behind the series could look like. It reportedly will be led by the GB202, followed by the GB203, and then the GB205 and GB206, followed by the GB207 at the entry level. What's surprising here, is the lack of a "GB204" succeeding the AD104, GA104, TU104, and a long line of successful performance-segment GPUs by NVIDIA.
The GeForce Blackwell ASIC series begins with "GB" (GeForce Blackwell) followed by a 200-series number. The last time NVIDIA used a 200-series ASIC number for GeForce GPUs was with "Maxwell," as the GPUs ended up being built on a more advanced node, and with a few more advanced features, than what the architecture was originally conceived for. For "Blackwell," the GB202 logically succeeds the AD102, GA102, TU102, and a long line of "big chips" that have powered the company's flagship client graphics cards. The GB103 succeeds AD103, as a high SIMD count GPU with a narrower memory bus than the GB202, powering the #2 and #3 SKUs in the series. There is curiously the lack of a "GB104."
NVIDIA's xx04 ASICs have powered a long line of successful performance-thru-high end SKUs, such as the TU104 powering the RTX 2080, and the GP104 powering the immensely popular GTX 1080 and GTX 1070 series. The denominator has been missing the mark for the past two generations. The "Ampere" based GA104 powering the RTX 3070 may have sold in volumes, but a its maxed out RTX 3070 Ti hasn't quite sold in numbers, and missed the mark against the Radeon RX 6800 (similar price). Even with Ada, while the AD104 powering the RTX 4070 may be selling in numbers, the maxed out chip powering the RTX 4070 Ti, misses the mark against the RX 7900 XT with a similar price. This has caused NVIDIA to introduce the AD103 in the desktop segment—a high CUDA core-count silicon with a mainstream memory bus width of 256-bit—out to justify high-end pricing, which will continue in the GeForce Blackwell generation with the GB203.
As with AD103, NVIDIA will leverage the high SIMD power of GB203 to power high-end mobile SKUs. The introduction of the GB205 ASIC could be an indication that NVIDIA's performance-segment GPU will come with a feature-set that would avoid the kind of controversy NVIDIA faced when trying to carve out the original "RTX 4080 12 GB" using the AD104 and its narrow 192-bit memory interface.
Given NVIDIA's 2-year cadence for new client graphics architectures, one can expect Blackwell to debut toward Q4-2024, to align with mass-production availability of the 3 nm foundry node.
Source:
VideoCardz
The GeForce Blackwell ASIC series begins with "GB" (GeForce Blackwell) followed by a 200-series number. The last time NVIDIA used a 200-series ASIC number for GeForce GPUs was with "Maxwell," as the GPUs ended up being built on a more advanced node, and with a few more advanced features, than what the architecture was originally conceived for. For "Blackwell," the GB202 logically succeeds the AD102, GA102, TU102, and a long line of "big chips" that have powered the company's flagship client graphics cards. The GB103 succeeds AD103, as a high SIMD count GPU with a narrower memory bus than the GB202, powering the #2 and #3 SKUs in the series. There is curiously the lack of a "GB104."
As with AD103, NVIDIA will leverage the high SIMD power of GB203 to power high-end mobile SKUs. The introduction of the GB205 ASIC could be an indication that NVIDIA's performance-segment GPU will come with a feature-set that would avoid the kind of controversy NVIDIA faced when trying to carve out the original "RTX 4080 12 GB" using the AD104 and its narrow 192-bit memory interface.
Given NVIDIA's 2-year cadence for new client graphics architectures, one can expect Blackwell to debut toward Q4-2024, to align with mass-production availability of the 3 nm foundry node.
71 Comments on NVIDIA Blackwell Graphics Architecture GPU Codenames Revealed, AD104 Has No Successor
we are left with 18432/ 256 bit and 7680/ 128 bit crystal. that way they can carve 5070 and 5080 and even the Ti models out of the 203 die. and have it coupled with 16 GB,
as for the 128 bit 12 Gb 3 gbit memory must be used similar to the 24 GB modules of RAM.
So what we get? A 4090 Ti with 512bit data bus as an RTX 5090, an upgraded 4080 Ti as an RTX 5080 and under that a bunch of 5070s and 5060s and 5050s based on Blackwell GPUs. While this might look as a very very very long shot, meaning the high end parts being of an older architecture than the mid and low end models, it does makes sense if we consider the possibility Nvidia to be keeping ALL the big dies for AI. Using Ada for the high end models, means that Nvidia will keep it's big gaming dies on the 5nm process instead of using valuable and more expensive 3nm wafers for gaming purposes. The performance gap between the two top Ada models and everything under them, does give Nvidia the chance to play with new GPUs for the mid range that will offer higher performance than existing RTX 4070s and 4060s while remaining way under RTX 4080. A small price increase, like RTX 5070 Ti at $900 could also help to make GB105 financially valid if it is small enough for use in a gaming GPU.
We can also add here the rumor that there is no high end models in AMD's RDNA4 series. Maybe AMD also sees that there is no financial logic into selling mid size or bigger 3nm dies for gaming GPUs, even with their chiplet approach.
Then again there is the much more possible scenario of Blackwell to be fast enough that a GB203 will end up faster than an AD102, so GB203 could be used for a few ultra expensive $1500+ graphics cards and we can forget anything bigger for the gaming market.
wouldt be the first time products were relaunched under a new name because tsmc was stuck tinkering with a node.
Although, this does suggest a step back from the "let's make everything more expensive" movement that we've seen in the last couple of years. *touch wood*
For the article: There actually was a GM107 chip, powering the 750 (Ti). The 950 and 960's GPUs were called GM206.
Only wafers above those contractually obligated are not guaranteed, much like what we saw during the pandemic. The silicon foundry business does not turn on a dime.
I'm not even sure Apple is taking all of the excess capacity, last I checked TSMC has spare capacity. In addition, Apple has saw a drop in sales over the last 3 quarters consecutively and has cut product shipments. TSMC also offers 2 3nm products, N3E and N3P. N3E being the mobile geared node, Apple is likely using that. It's how to tell how much that would impact AMD, who will almost certainly be using N3P for it's graphics cards and CPUs.
If AMD is late to 3nm, it's extremely likely that it down to AMD's own choice and not the fault of Apple. TSMC will have a glut of wafers in 2025 as that's when production volume that was pushed for during the pandemic starts coming online. On top of that Intel has been seeing success in it's foundry for hire business as well, which can only be good for the industry as a whole. The way I see it, capacity is set to increase while sales are set to decrease. Unless some large demand comes around again I see the potential for too much supply.
high end gpu's need 3nm to truly shine. Apple owns them all.
yep, I already said AMD will make a come back in late 2025 or early 2026.
2. It's said from numerous sources, that it comes at earliest 2025.
Intel is after AMD's lunch and they poached several former AMD employees who understand how AMD thinks and the limitations of their MO.
Arc cannot be ignored for much longer, especially with it providing the same Linux friendliness of Radeon, achieving a decent amount of performance and with ultra high end GeForce cards being priced ever out of reach.
It'd be a safe bet that the 5090 will be 2K, and we don't even have software that can take full advantage of Ada features yet.
en.wikipedia.org/wiki/3_nm_process
I think there is too much philosophy in the thread, negative thinking and lack of actual facts and solutions.
If indeed the new gens both from AMD and nvidia are delayed till 2025 | 2026, then you can safely assume that's the end of the graphics market as we know it.
2018 - RTX 2000
2020 - RTX 3000
2022 - RTX 4000
2025 | 2026 - RTX 5000
3 | 4 years to launch a new generation ?
Just wow o_O