Node shrink to 3nm or 2nm + 36-40 Gbps GDDR7 should be enough for a big uplift.
Node shrink to 3nm or 2nm + 36-40 Gbps GDDR7 should be enough for a big uplift.Any speculation about the gen after Blackwell? It seems (could still be wrong tho) that we will only get architecture iterations over Ada now, so what is NVidia aiming for in 2027? Surely there must be some new hardware feature? Or will there be enough node shrinking and memory tech improvements to get enough of a novelty out of sheer throughput and bandwidth?
Node shrink to 3nm or 2nm + 36-40 Gbps GDDR7 should be enough for a big uplift.
Too big an update for a refresh. No need (has nothing to do with AI focus or whatever), there is simply no need for a generational uplift in memory bandwidth when the normal strategy of using your salvaged chips at a slight hit to priceshas worked so well so farI feel like maybe 3 GB 36/40 gbps might be ready in time for the Blackwell Refresh. It is a bit of a stretch but possible.
Too big an update for a refresh. No need (has nothing to do with AI focus or whatever), there is simply no need for a generational uplift in memory bandwidth when the normal strategy of using your salvaged chips at a slight hit to priceshas worked so well so far
| Codename | Model | SRP | SM / CU | +/- SM/CU | Raichu's Perf Est | My TPU | Memory | BW | TDP | +/- TDP | ||
|---|---|---|---|---|---|---|---|---|---|---|---|---|
| GB202 | RTX 5090 | > $1599 | 170 | + 42 | + 60％ | 32GB 512-bit GDDR7 | 1.79 TB/s | 575 W | + 125W | |||
| GB203 | RTX 5080 | $999 ? | 84 | + 4 | + 30％ | ~20% | 16GB 256-bit GDDR7 | 960 GB/s | 360 W | + 40W | ||
| RTX 5070 Ti | $799 ? | 70 | + 4 | 16GB 256-bit GDDR7 | 896 GB/s | 300 W | + 15W | |||||
| GB205 | RTX 5070 | $599 ? | 48 | - 8 | + 20% | ~10% | 12GB 192-bit GDDR7 | 672 GB/s | 250 W | + 30W | ||
| RDNA4 N48 | RX 9070XT | $499 ? | 64 | + 4 | + 10% | 16GB 256-bit GDDR6 | 640 GB/s | 220 W | - 43W |
Its that 60+ percent perf increase for 5090 in gaming, or compute?
Codename Model SM / CU +/- SM/CU Raichu's Perf Est Memory TDP +/- TDP GB202 RTX 5090 170 + 42 + 60％ 32GB 512-bit GDDR7 600 W + 150W GB203 RTX 5080 84 + 4 + 30％ 16GB 256-bit GDDR7 400 W + 80W RTX 5070 Ti 70 + 4 16GB 256-bit GDDR7 285 W
350 W+ 0W
+ 65WGB205 RTX 5070 50 - 6 + 20% 12GB 192-bit GDDR7 250 W + 30W RTX 5060 Ti ? 12GB 192-bit GDDR6 ? RDNA4 N48 RX 8800XT 64 + 4 16GB 256-bit GDDR6 220 W - 43W
I have updated upcoming RTX5000 information with Raichu's performance estimate in the table above... will be updating if any leaks appear...
5090/5080/5070/5070Ti are all in Q1.Hmm, so maybe spring or summer for the 5070/70ti to show up for sale? Nice that they are now rumormilling that its 16Gb of gddr7 at least.
Makes sense since using current rumors as an indication, that's just GB202 and GB203 covering all these cards, although GB205 seems absent from rumors. Maybe 5060ti alone is on GB205 and 5060 on GB206. Or 5070 ends up on GB205 with 12GB after all5090/5080/5070/5070Ti are all in Q1.
5070 should be the full 205 tho.Makes sense since using current rumors as an indication, that's just GB202 and GB203 covering all these cards, although GB205 seems absent from rumors. Maybe 5060ti alone is on GB205 and 5060 on GB206. Or 5070 ends up on GB205 with 12GB after all
No reason for nvidia to release 16gb 5070Makes sense since using current rumors as an indication, that's just GB202 and GB203 covering all these cards, although GB205 seems absent from rumors. Maybe 5060ti alone is on GB205 and 5060 on GB206. Or 5070 ends up on GB205 with 12GB after all
No reason for nvidia to release 16gb 5070
I expect both 5070 & 5060 ti to be 12gb
I wouldn't give any weightage to wccftech5060ti seems 16GB
NVIDIA GeForce RTX 5060 Ti Gets 16 GB VRAM, RTX 5060 GPU Sticks With 8 GB GDDR7 Memory
NVIDIA's mainstream GeForce RTX 5060 Ti & RTX 5060 GPUs will retain memory capacities similar to those of the previous generation.wccftech.com
Plain 5060 16GB? Hell would freeze.But there could be 2 versions of 5060: 8gb & 16gb
8 GB thing is getting proper unsustainable for new launches - 60-level GPUs are bread and butter for NVIDIA that covers the market and makes devs optimise for them, they got burnt on "4080 12 GB" but very few people were going to buy it anyway at those prices, but this is far more commercially consequential market, plus both AMD and Intel are having a go at it.Plain 5060 16GB? Hell would freeze.
Or they could pair it with LPDDR5X-9600. They used to do this crap in the Geforce days with SDRAM laden cards, and had DDR for the better ones.If I was them I'd delay it till 3 GB modules become available.
I think that would be waaaaaaaaay too slow.Or they could pair it with LPDDR5X-9600
Codename Model SRP SM / CU +/- SM/CU Raichu's Perf Est My TPU Memory BW TDP +/- TDP GB202 RTX 5090 > $1599 170 + 42 + 60％ 32GB 512-bit GDDR7 1.79 TB/s 600 W + 150W GB203 RTX 5080 $999 ? 84 + 4 + 30％ ~20% 16GB 256-bit GDDR7 896 GB/s 400 W + 80W RTX 5070 Ti $799 ? 70 + 4 16GB 256-bit GDDR7 896 GB/s 285 W
350 W+ 0W
+ 65WGB205 RTX 5070 $599 ? 50 - 6 + 20% ~15% 12GB 192-bit GDDR7 672 GB/s 250 W + 30W RDNA4 N48 RX 8800XT $499 ? 64 + 4 + 10% 16GB 256-bit GDDR6 640 GB/s 220 W - 43W
I have updated upcoming RTX5000 information with Raichu's performance estimate in the table above... will be updating if any leaks appear...
512 bit bus and sadly not newer N3E/P node...The 5090 looks very impressive performance-wise but 600W??
The 5090 looks very impressive performance-wise but 600W?? If true, that means the uber-oced versions could be 700-750W.
I was thinking the same thing. At least if paired with an X3D chip, the rest of the system shouldn't draw a ton of power.There's a point where you would start to hit the practical hard limit of 1500 W. And that's for everything and also have to account for transient spikes too.