Nvidia’s next-gen RTX 3090, 3080 and 3070 GPUs are coming. So if you’ve been keenly waiting for a new GPU to beef up your PC ahead of Cyberpunk 2077, or you’ve just been holding off for a long-awaited upgrade, here’s what we know about Nvidia’s new RTX GPUs.
This post has been updated with new info and photos.
Update 2/09: Nvidia has officially announced the RTX 3090, 3080 and 3070, with the first two cards available from September 14 and 22, and the RTX 3070 sometime in October.
Here’s the official specs:
NVIDIA GeForce RTX 3090, 3080 and 3070 | ||||
---|---|---|---|---|
RTX 3090 | RTX 3080 | RTX 3070 | RTX 2080 Ti | |
CUDA Cores | 10496 | 8704 | 5888 | 4352 |
Boost Clock | 1.7GHz | 1.71GHz | 1.73GHz | 1545MHz |
Memory Clock | 19.5Gbps GDDR6X | 19Gbps GDDR6X | 16Gbps GDDR6 | 14Gbps GDDR6 |
Memory Bus Width | 384-bit | 320-bit | 256-bit | 352-bit |
VRAM | 24GB | 10GB | 8GB | 11GB |
Single Precision Perf. | 35.7 TFLOPs | 29.8 TFLOPs | 20.4 TFLOPs | 13.4 TFLOPs |
Tensor Perf. (FP16) | 285 TFLOPs | 238 TFLOPs | 163 TFLOPs | 114 TFLOPs |
Ray Perf. | 69 TFLOPs | 58 TFLOPs | 40 TFLOPs | ? |
TDP | 350W | 320W | 220W | 250W |
GPU | GA102? | GA102? | GA104? | TU102 |
Transistor Count | 28B | 28B | ? | 18.6B |
Architecture | Ampere | Ampere | Ampere | Turing |
Manufacturing Process | Samsung 8nm | Samsung 8nm | Samsung 8nm | TSMC 12nm “FFN” |
Launch Date | 09/24/2020 | 09/17/2020 | 10/2020 | 09/20/2018 |
Launch Price | MSRP: $1499 | MSRP: $699 | MSRP: $499 | MSRP: $999 Founders $1199 |
The RTX 3090 will have 24GB of GDDR6X RAM, while the RTX 3080 will have 10GB.
Videocardz got themselves hold of spec sheets from AIB partners Gainward, revealing the upcoming specs for their RTX 3090 and RTX 3080 models. It’s likely that Nvidia’s own GPUs, which will be the first ones available for pre-order, will vary slightly. The official specs for the RTX 2080 Ti/2080/2070 Founders Edition and reference models had their own tweaks, which is largely because companies like Gainward — and ASUS, MSI, Gigabyte etc. — will all want to pre-overclock and tune the cards themselves.
But some things don’t change — like the amount of VRAM cards will have. As seen on the box shot above, the RTX 3090 model will have a whopping 24GB of memory, compared to 10GB on the RTX 3080. Both cards will have boosted tensor and ray-tracing cores, as well as substantially more CUDA cores than the 20-series GPUs.
This leaves an obvious upgrade path for Nvidia in the future too, letting them sell a 3080 Ti or some kind of 3080 SUPER alternative later on with more VRAM.
The RTX 3090 will have a new NVLink bridge for SLI support — but not the RTX 3080.
This one’s another revelation from the Gainward leak, which revealed a new NVLink bridge on the RTX 3090 model. It means the RTX 3090 will be your only effective choice if you still want a dual-GPU setup — although with the amount of support from developers, it’s still very, very hard to recommend any gamers seriously invest in a dual-GPU system.
There’ll be a new 12-pin connector that requires an adapter, but third-party cards aren’t required to use them.
A rumour in the works for months was that Nvidia was working on a new power connector for their flagship 3000-series GPU. This started doing the works around mid-March, with the logic being that a single 12-pin connector was capable of delivering much more power than multiple 8-pin or 6-pin connectors. Two 8-pin connectors, according to the specification, can deliver a maximum of 375 watts of power. A 12-pin, on the other hand, can deliver up to 600 watts. (That doesn’t mean an RTX 3090 needs 600W of power, FYI.)
Nvidia eventually confirmed the rumours in a recent video, showing off a shot of what the new 12-pin connector would look like. They added that any RTX cards direct from Nvidia that use the 12-pin connector would ship with an adapter.
It’s worth noting — not all the new RTX cards will use the new 12-pin adapter. It’s likely that it’s only necessary for the RTX 3090 at this point, but we’ll find out soon how many models are affected. Third-party boards don’t need them, as seen from this RTX 3090 Phoenix model from Gainward:
And more images have been posted to the Gainward website:
Vapour chamber cooling will be a thing.
Nvidia’s video into GPU design also spoke a fair bit about vapour chamber cooling. This has been used in consoles for a while now — it’ll be a feature of the PS5 and the Xbox Series X, but Microsoft used the cooling system for their Xbox One X refresh as well.
It alludes to a complete redesign for the flagship gaming GPUs, particularly the RTX 3090, although it’s not known if vapour chamber cooling will feature on the lower-end 30 series cards. AIB models might also go a different route — partners are still using three 8-pin connectors instead of the new 12-pin connector, and they might feel free to stick with a more traditional fan/heatsink setup as well depending on the card’s thermals.
The RTX 3090 is going to be a big card.
3 slots. 2000$. pic.twitter.com/ZgKsPUQ1U4
— James (@GarnetSunset) August 21, 2020
A shot of the RTX 3090 compared to the RTX 2080 leaked recently, showing that the flagship gaming GPU would be a triple-slot monster. That’s a problem for anyone wanting to build ultimate mini-ITX gaming rigs.
The RTX 3090 will have 24GB of GDDR6X RAM.
This one comes via a recent report from Videocardz, who learned from AIB sources that the flagship RTX 3090 would come with a whopping 24GB of VRAM. The RTX 3080, as we learned before, will only be shipping with 10GB GDDR6X memory, which should still be enough for 4K gaming particularly if the new cards utilise caching and memory advancements.
This leaves an obvious upgrade path for Nvidia in the future too, letting them sell a 3080 Ti or some kind of 3080 SUPER alternative later on with more VRAM.
The first 3000-series GPUs, including the RTX 3090, will be revealed on September 2.
A new banner image posted to the Nvidia GeForce Twitter account revealed the new GPUs would finally be unveiled on September 2 Australian time. It came after weeks of surprising leaks about the GPU — more than any Nvidia launch in recent history.
While most of those leaks were unverified, the biggest one came from Nvidia’s own partner, Micron. The memory manufacturer revealed in a white paper, released a month earlier than expected, that they were supplying GDDR6X modules for Nvidia’s RTX 3090 GPU.
In the paper, Micron revealed an interesting detail about the card’s memory bandwidth, which is just shy of 1GB/second. It’s a sizeable upgrade on the memory bandwidth available on other GDDR6 or GDDR5X cards (like the Titan RTX or AMD’s RX5700 XT), although what this translates to in real-world performance is unknown.
There’ll be an RTX 3090, 3080 and 3070 at launch.
Third party partners have helpfully confirmed the info on this one. PC Gamer spotted MSI entering multiple RTX 3090, 3080 and 3070 models into the Eurasian Economic Commission database, matching their naming convention from previous generations. It doesn’t tell us anything more about the cards individually, but it does mean we should expect to see some announcements around the RTX 3070 on September 2 — although the RTX 3080 and the 3090, or whatever the final names end up being, will be the featured products.
Some high-end RTX 3000 series cards will have a triple-fan design.
This isn’t a huge departure from the third-party RTX 2080 and RTX 2080 Ti designs, but it’s still nice to know for those curious about the card’s design. Videocardz received screenshots of internal design proofs of an ASUS ROG STRIX 3000-series GPU, sporting a design that’s basically identical to the ROG STRIX 2080 Ti.
It’s not final design, obviously. But with all the rumours that Nvidia’s reference RTX 3080 Ti (or RTX 3090) would have fans on the front and back of the GPU, it makes sense that AIB manufacturers would maintain a triple-fan design.
The GeForce RTX 3080 will be a 10GB card, with a clock speed of at least 2.1GHz.
You can’t have a GPU without GPU drivers, and eagle-eye viewers spotted new entries for the upcoming RTX 3000-series GPUs in a couple of places. The first was from T4CFantasy, the user responsible for maintaining the GPU database for TechPowerUP, and GPU leaker _rogame, who spotted new device IDs in a recent INF file from Microsoft’s automatic updates.
[Nvidia Ampere]🧐
PCI IDs for GA102 GPUs
GA102 12GB (RTX 3080 Ti ?)=> 0x2206
GA102 10GB (RTX 3080)=> 0x2206
GA102 24GB (RTX 3090)=> 0x2204— _rogame 🇵🇸 (@_rogame) August 16, 2020
An unnamed Nvidia GPU with the same device ID (10DE 2206) then popped up on Userbenchmark, although obviously pre-release drivers made it impossible to get a proper gauge of its performance.
Still, it fits with how Nvidia runs things. Their RTX 20 series launch featured three cards: The RTX 2080 Ti, RTX 2080 and the RTX 2070, although you could only pre-order the 2080 Ti and RTX 2080. At the time, those cards were going for $1899, $1199 and $899, which is a good expectation for what the 30 series cards might be priced at.
There’s also been plenty of chatter about three variants of the PG132 reference board, which is what will feature the chipset powering the RTX 3000 series GPUs. The PG132-10 board ID, for instance, will likely be the RTX 3090 (or whatever naming convention Nvidia opts to go for).
Nvidia’s likely to release a Titan RTX-version of their Ampere cards as well, one targeted at the workstation or super-enthusiast market.
The RTX 3000 series cards will support NVCache, a technology that uses your SSD and RAM to improve load times.
AMD already has a rival offering in the High-Bandwidth Cache Controller, a feature that could reportedly boost average FPS by up to 50 percent in certain scenarios.
In the real world, AMD’s HBCC’s performance improvements were mostly limited to synthetic benchmarks. But with more people playing games at higher and higher resolutions, and developers transitioning into the power of the next-gen consoles and their associated architecture, leveraging system memory and cache in different ways can become more useful.
The talk of the town right now is that Nvidia will come out of the gate with NVCache, which is designed to largely do two things. The first is to utilise your system’s SSD — not regular platter hard drives — to help with load times. It’s expected to be an answer to the next version of AMD’s caching technology, which is expected to launch with the Big Navi GPUs AMD are launching later this year (whenever that is).
In practical terms, the idea here is that NVCache will be able to decompress items stored in the GPU’s memory by using the ray-tracing tensor cores. It’s not meant to be used all the time, but when used it should be able to reduce the VRAM usage on the graphics card. That’s most helpful when playing at larger resolutions, or in instances where you’re playing games that max out the GPU’s memory through things like mods (like Skyrim).
Don’t expect the RTX 3080/3090 to be cheap.
There’s been suggestions over the last week that the flagship 3000 series GPU will hit $US1399, which would put it just shy of $2000 locally. That’s not much more than what the RTX 2080 Ti cost in Australia, with the Founders Edition boards retailing for $1899 and some AIB models going for $1999 or $2099 in later months, particularly as supply dried up.
Some leaked pricing also appeared for the cost of Colourful’s upcoming 3000 series AIB cards, which were priced at 13,999, 12,999 and 12,000 CNY (Chinese Yuan) respectively. That works out to about $US2,000, although obviously currency conversions, shipping and pricing across markets doesn’t always translate.
Still, given some of the talk scattered around — and the prospect of AMD about to compete with a proper high-end 4K capable card of their own — it’s no surprise that Nvidia would go all-0ut. Demand for the RTX 2080 Ti and RTX 2080 was massive at launch, despite complaints from users about the lack of games that actually supported ray-tracing at the time. With every AAA game for the next few years supporting ray-tracing in some fashion, it’s easier to charge a premium especially if it looks the part — as Cyberpunk 2077 has in so much of its 4K footage this year.
Expect more architectural updates in early October.
The coronavirus has postponed much of the industry’s usual cadence, and one such conference put on the backburner was the GPU Technology Conference (GTC). Typically held mid-year, it’s now running from October 5 to October 9, where Nvidia CEO Jensen Huang will give another keynote speech.
GTC doesn’t typically focus a lot on gaming, but there are often a lot of live demos, architectural talks and windows into the future of graphics computing. Final Fantasy 15 featured at a GTC event a few years ago, as did discussion on the use of AI in graphics, and applications for ray-tracing. Nvidia’s GTC 2018 keynote also showed off the famous Industrial Light and Magic Star Wars: Battlefront 2 ray-tracing demo.
GTC isn’t a place for announcing mid-range or entry-level RTX cards — Nvidia has always done those through separate briefings. But if you’re interested in the architectural improvements Ampere can offer, GTC typically has a couple of things in store.
Leave a Reply