Monday, March 11th 2024
NVIDIA GeForce RTX 50-series "Blackwell" to use 28 Gbps GDDR7 Memory Speed
The first round of NVIDIA GeForce RTX 50-series "Blackwell" graphics cards that implement GDDR7 memory are rumored to come with a memory speed of 28 Gbps, according to kopite7kimi, a reliable source with NVIDIA leaks. This is despite the fact that the first GDDR7 memory chips will be capable of 32 Gbps speeds. NVIDIA will also stick with 16 Gbit densities for the GDDR7 memory chips, which means memory sizes could remain largely unchanged for the next generation; with the 28 Gbps GDDR7 memory chips providing 55% higher bandwidth over 18 Gbps GDDR6 and 33% higher bandwidth than 21 Gbps GDDR6X. It remains to be seen what memory bus widths NVIDIA chooses for its individual SKUs.
NVIDIA's decision to use 28 Gbps as its memory speeds has some precedent in recent history. The company's first GPUs to implement GDDR6, the RTX 20-series "Turing," opted for 14 Gbps speeds despite 16 Gbps GDDR6 chips being available. 28 Gbps is exactly double that speed. Future generations of GeForce RTX GPUs, or even refreshes within the RTX 50-series could see NVIDIA opt for higher memory speeds such as 32 Gbps. When the standard debuts, companies like Samsung even plan to put up fast 36 Gbps chips. Besides a generational doubling in speeds, GDDR7 is more energy-efficient as it operates at lower voltages than GDDR6. It also uses a more advanced PAM3 physical layer signaling compared to NRZ for JEDEC-standard GDDR6.
Sources:
kopite7kimi (Twitter), VideoCardz
NVIDIA's decision to use 28 Gbps as its memory speeds has some precedent in recent history. The company's first GPUs to implement GDDR6, the RTX 20-series "Turing," opted for 14 Gbps speeds despite 16 Gbps GDDR6 chips being available. 28 Gbps is exactly double that speed. Future generations of GeForce RTX GPUs, or even refreshes within the RTX 50-series could see NVIDIA opt for higher memory speeds such as 32 Gbps. When the standard debuts, companies like Samsung even plan to put up fast 36 Gbps chips. Besides a generational doubling in speeds, GDDR7 is more energy-efficient as it operates at lower voltages than GDDR6. It also uses a more advanced PAM3 physical layer signaling compared to NRZ for JEDEC-standard GDDR6.
47 Comments on NVIDIA GeForce RTX 50-series "Blackwell" to use 28 Gbps GDDR7 Memory Speed
If true, that makes perfect business sense. Nvidia will start low and have the ability to do a mid-cycle refresh with faster memory.
You keep mentioning CPUs because for GPUs it's just not better at this point in time and the main reason is that it's still cheap enough to make monolithic dies. Like I said before even for CPUs AMD still uses monolithic designs. They seem to have forgotten about that paper with Blackwell still being monolithic almost 10 years later.
Clearly for Nvidia it's not better at the moment as either the cost to make it work negates the savings or the performance/efficiency hit is too big. It's so weird to me how people on here act as if they know better than the hardware engineers at Nvidia and that you can just Google the anwer or read some papers. Funnily enough the 4090 was actually cheaper than the 3090 if you take into account inflation. What's AMD's excuse for the increasing prices? They've moved to chiplets now.
379mm2 is not what I'd call "barely bigger" than 304mm2 either. That's a significant difference in terms of yield.
What you don't seem to understand when it comes to chiplets is that total die area is not as direct an indicator of costs as it is with monolithic. The GCD is only 304mm2 while the MCDs are a mere 34mm2. The yield for the MCDs is going to be insanely high due to the very small size and you will get a ton per wafer, making them extremely cheap. Yield decreases exponentially as die size increases so in fact by having more smaller chiplets, regardless of whether the total die area is greater, means the total chip can be cheaper to manufacture. This is why AMD is able to produce server CPUs at a lower cost than Intel's server CPUs while also scaling higher. It's a first generation GPU chiplet design. Zen 1 was not better right out of the gate either so it's stands to reason that we apply logic evenly here.
You are certainly mistaken if you think monolithic GPUs are cheap to make. High-end GPUs are several times larger than CPUs and by extension the yield and cost factors are vastly worse. This is why GPUs have historically been manufactured on a more mature node. You might only get a handful of good 600mm2 dies per wafer because 1) each die is large 2) each defect in the die wastes 600mm2 of space. Compare that to a 300mm2 die for example where each defect only impacts 300mm2 of space. You are wasting half the space per defect. No, it's just really hard to implement a chiplet based GPU. AMD has stated that they require magnitudes more bandwidth for inter GCD communication. This is why AMD introduced the fan-out links with the 7000 series and will likely further push what the infinity fabric is capable of handling on their GPUs. Clearly? You are assuming that whatever Nvidia has on the market now is what Nvidia thinks is the best possible product they will ever have, which is almost certainly false. There are so many other factors you are jumping over to try to force a conclusion that just isn't there. We don't know Nvidia's opinion on chiplets outside of the paper they published or what technical hurdles might be in the way preventing them from implementing a chiplet based architecture. You can't jump to the conslusion that no chiplet based products means that Nvidia doesn't think chiplets aren't a good approach. You are just completely guessing at that point. Both the 7900 XTX and 6900 XTX have an MSRP of $1,000. The 7700 XT has an MSRP of $409 while the 6700 XT had an MSRP of $479.
Not sure how this is relevant though. When I was discussing cost, I'm referring to the cost to produce. Not the amount Nvidia or AMD will charge. Those are two completely different things. A theoritical price increase to the customer says nothing of production costs.
AMD was first to introduce chiplet based GPUs to both consumer and enterprise market, it's a big advantage that we will need to see if they will use. Once they got chiplets figured out for GPUs completely they can just scale up like they did with Zen. And no monolithic GPU would be able to compete.
Either nvidia will also switch to chiplets by then or they will be outperformed. They are behind in packaging tech and ask Intel how that worked out for them, back when they were laughing that Zen 1 is just glued together CPU.
Why do you think AMD moved to chiplets in CPUs with Zen and achieved such success with this approach?
They have already moved to full chiplet approach i data center GPUs with Instinct MI300.
Next step is more chiplets in client GPUs. It will take a few years, segment by segment. Just watch it happen. How about you putting some effort to understand that transition to chiplets is complex, it takes time and it does not happen fully and miraculously over one generation of products? It's a multi generational effort of incremental improvements.
We do not know if and what happens with 'Navi 41'. It's rumours. Perhaps it takes more time to perfect it. They already have bigger chiplet designs for data center GPU Instinct MI300, so chiplets work 100% and MI300 sells like hot cakes right now. For client GPUs, it certainly takes a complex effort from multiple teams of engineers to perfect it because this is supposed to be the first multi compute chiplet, unlike Navi 31. It will just take more time to get it done. You do not need to like it. Just stop being prejudiced about it and stay tuned. You could have posted the same list of CPUs in 2017 and say that Zen was Beh and Meh and that they should drop chiplets. Nonsense.
Your "grand scheme of things" only looks into two generations of products. Short-sighted. You have no idea what it coming in next couple of years. That's how "grand" this "scheme of things" is. It is you who claims to know better than engineers in AMD.
TSMC 3N and GDDR7 will help bring on the typical large performance and efficiency gains for RTX 5000. Along with a couple new hardware features, IPC tweaks, and software improvements of course.
ignoring for a moment that they run a business and want to make money (shareholders), not to make "US" happy.
do "you" also go to a Porsche/Bentley (or similar) dealer and tell them you want their 4 door performance suv for the price of a VW or Toyota,
or that you should be able to get a 10 room mansion for the price of a 2 bedroom condo?
right.
short of having a +100K income/6 digit lottery win, i will never buy a (new) xx80Ti/xx90,
but that doesnt mean i will sour it for the folks that can and do.
instead of looking at a product and then "whine" how much it costs, buy the product that fits the wallet,
or dont if you dont like the offering, and have it (negatively) impact their sales, no one forces you to buy anything.
Like you, I have no problem with the existence of higher prices in the consumer space. But I do have a problem in not being able to pay $200-300 and getting a capable card in return. I mean, compare what a 460 could do, compared to a 480. And then think of how a 4060 compares to a 4080.
when i was kid my father had a porsche turbo, back then equivalent to what a house cost (germany, ~50K).
the car now goes for about 100-150K (decently equipped, not fully loaded stuff), most single family homes now START at 150K.
and its the same for everything else, as i dont pay for bread/milk, what i paid 40y ago.
a GTX460 was ~200. how many times faster is a 200$ card (e.g. 1660Ti) now?
and we havent even talked about bad (console) ports and/or missing optimization,
all contributing to the fact i need more "HP" to get similar FPS, but has NTOHING to do with the cost of the card.
(and to make sure before some complain: i never have made any statement, that im happy with what parts/gpus cost..)
also not taking any "changes" on your side into account, as i doubt you were running the same moni res/fps you do now,
and ignores what i already mentioned.
all contributing to the fact i need more raw power do do the same, but has nothing to do with the increase in perf, compared to its increase in price.