Nvidia’s next-gen GeForce graphics playing cards stay a scorching matter on the rumor mill, and a brand new publish on X is attracting a whole lot of consideration, having spilled a number of the potential specs of those Blackwell GPUs.
This comes from common leaker Kopite7kimi (through VideoCardz), who offers us the rumored – add seasoning – loadout for Streaming Multiprocessors (SMs, which instantly mirror the all-important core depend), in addition to reminiscence bus and kind down the product vary.
GB202 12*8 512-bit GDDR7GB203 7*6 256-bit GDDR7GB205 5*5 192-bit GDDR7GB206 3*6 128-bit GDDR7GB207 2*5 128-bit GDDR6June 11, 2024
As you’ll be able to see, the leaker states that GB202 – the top-end chip for Blackwell that’ll energy the RTX 5090 – will purportedly provide a most depend of 192 SMs.
Because the leaker notes, that’s labored out utilizing the sum of 12 x 8 – which is the Graphics Processing Clusters or GPCs occasions the Texture Processing Clusters or TPCs inside – which is 96, after which doubling (which assumes Nvidia sticks with 2 SMs for each TPC) to get the whole of 192 SMs.
The reminiscence bus of the RTX 5090 might hit 512-bit, as we’ve been informed earlier than, however Nvidia doesn’t should push to the utmost capabilities of the spec for GB202 outlined right here – we’ll come again to debate that in second.
So, trying down the vary, subsequent now we have GB203 (RTX 5080) geared up with (as much as) 84 SMs and a 256-bit bus, after which GB205 (presumably the RTX 5070 – however not essentially) which is 50 SMs plus 192-bit.
With GB206 (possibly the RTX 5060 Ti?) Nvidia is plumping for 36 SMs and 128-bit, and eventually GB207 (RTX 5060?) is about to run with 20 SMs. As a ultimate twist, the latter chip might stick on GDDR6, moderately than advancing to GDDR7 video reminiscence as the remainder of the Blackwell vary will do.
Evaluation: Too early to worry but
In some methods, efficiency is absolutely not the central problem right here – moderately next-gen pricing, and we actually don’t have the inside track on that from the rumor mill but.
Let’s dig into the small print a bit extra, and whereas the maths of GPCs and TPCs might look advanced and appear complicated, it doesn’t matter as a lot because the comparability of SM counts to the current-gen Lovelace graphics playing cards. And these are roughly alongside the identical strains, with the notable exceptions of the top-end GB202 (RTX 5090), which has a a lot heavier loadout of SMs (192 for Blackwell vs 144 for Lovelace), and what may be the RTX 5070 or GB205 (50 for Blackwell vs 60 for Lovelace).
Nevertheless, we should stress that the SM figures offered listed here are the total counts obtainable for the given chip, so whereas GB202, for instance, could also be geared up with 192 SMs, the RTX 5090 might use a cut-down model of that chip. Certainly, that is precisely what’s rumored, and that the reminiscence bus will likely be trimmed considerably, as nicely (to 448-bit – and probably 28GB of VRAM because of this, not the total 32GB Nvidia might squeeze on if it needed to, no less than in concept).
Why lower GB202 for the RTX 5090? As a result of Nvidia doesn’t have to run the total configuration, and certain that the majority highly effective, totally enabled GB202 chip will likely be utilized in skilled AI graphics playing cards – or maybe reserved for a future RTX 5090 Ti (although that feels unlikely, frankly, however you by no means know).
Concerning the RTX 5090’s potential efficiency, earlier rumors have actually steered a really highly effective graphics card – within the order of fifty% quicker than the RTX 4090, or much more maybe – so we wouldn’t fear on that rating.
What some people have taken from this specific leak, although, are considerations that the RTX 5070 appears to be like a bit weak – though keep in mind the RTX 4070 makes use of a really cut-down chip itself (46 of 60 SMs enabled).
We must also be very cautious right here as a result of on prime of not figuring out precisely the place Nvidia would possibly draw the strains for reducing down SM counts, we additionally aren’t conscious of the architectural good points Blackwell delivers. So guessing about efficiency at this stage – based mostly on mere rumored specs – isn’t actually a really fruitful train proper now.
Briefly, sit tight and watch for extra concrete leaks than this, however worries are being aired, and loads of them, not simply in regards to the RTX 5070 as talked about, however the RTX 5080 trying a great way off the tempo in comparison with the flagship. And certainly the lower-end Blackwell graphics playing cards not trying all that tempting or able to making our checklist of the finest GPUs. Once more, all of that is leaping to untimely conclusions, so let’s hold that leaping to a minimal in the meanwhile.
You may also like
RelatedPosts
