TopSpoiler
Regular
So is the GA102 only for the 4090 series?
4080 may be cheaper than expected.
- Status
So is the GA102 only for the 4090 series?
4080 may be cheaper than expected.
DegustatoR
Legend
AD102, and I find that to be very unlikely - unless they plan to use AD102 for workstation and server markets extensively. Wouldn't make much sense to have only one SKU on the chip with the most percentage of defects otherwise.
Product pricing rarely have anything to do with what chip a product is using.
But again with the wattage though.
4080 on AD103 having a "similar TGP to GA102" means that it should be 320-350W. And then we're supposed to have a AD104 based 4070 which is 300W? So there's like 20W of difference between a 4070 and 4080 based on two different chips?
Also if we assume that 4070 will be above 3090 (some sources are saying that 4060 should be about on par with 3090) then it will be a rather small perf/watt gain (350->300) from going down to an N5 (N4?) process AND presumably dropping from G6X to G6.
Looks fishy as hell.
Also AD104 with a 192 bit bus and G6? I dunno, these looks off.
AD104 is supposedly only 60 SMs. 300w on 5nm does seem high for that configuration unless Nvidia really screwed the pooch. Same for AD103. Doesn’t make sense for it to have the same number of SMs and less memory than the 3090 at the same
power on a much better process.
kopite7kimi has a pretty good track record of Nvidia leaks though so it might be true. Maybe Nvidia cranked clocks or there’s a lot more silicon in those chips than the SM count implies.
DegustatoR
Legend
Significant improvement in clocks would make chips with same SM numbers a lot faster though which in turn would mean that such 4070 wouldn't be just above 3090 but would in fact beat it by some 50%. Which again seems fishy for a supposedy 192 bit G6 chip.
TopSpoiler
Regular
From they learned from Ampere, Nvidia maybe decided not to put the largest chip into the not-top-end SKUs. So AD103 for 4080 makes sense for me.
Kopite said GA102, not the specific SKU. So the TGP he points could be ~450W, of 3090 Ti.
I agree it's too narrow. Maybe AD104 is not aiming 3090 in 4K gaming. Let's wait and see how large L2$ helps.
This is my estimation based on Kopite's tweet.
AD102 - 4090/4090Ti, TGP 600W~ w/ up to two 12VHPWRs.
AD103 - 4080/4080Ti, TGP ~450W w/ single 12VHPWR or up to three 8-pins.
AD104 - 4070/4070Ti, TGP ~350W w/ single 12VHPWR or two 8-pins.
DegustatoR
Legend
Or a 320W of 3080. Considering how pointless 3090Ti is from performance per watt scaling it makes even less sense to assume that he means 450W.
The higher we go in these wattage figures the less sense they make. His last tweet mentions a 900W board for AD102 - I mean, cmon now.
Right you would expect performance to be through the roof with these power numbers unless Lovelace is another Ampere situation where there was almost no improvement in perf/watt. That’s unlikely though.
The 3070 is a 220w card with 46 SMs. A 300w 4070 with ~58 SMs should be at least 40% faster which puts it right at 3090 performance. If Lovelace significantly improves on perf/watt then the 4070 should be quite a bit faster than that (or much lower power). If remains to be seen how much Nvidia will keep pushing up power consumption on mid range cards. The xx70 SKUs were consistently ~200w in the past 4 generations.
DegustatoR
Legend
Well one thing is for sure - a 4080 with 16GBs of VRAM at ~350W while being some +50-80% to 3080 would be a nice upgrade for all 3080/12/Ti owners.
These 900W 4090 and 300W 4070 look all sorts of weird though.
Worth noting that Kopite7kimi has already went from a 400W 4070 down to a 300W one over the last five or so tweets. So perf/watt seem to be improving fast there ¯\_(ツ)_/¯
TSMC states that N7 to N5 should result in 30% power reduction and general expectation for 8N to be considerably worse than N7 in power (if we look at Ampere vs Turing then it's not that much better than TSMC's 16FF+ to which N7 should be -60%; so about -80% or so?). Lovelace should have some very sizeable gains in perf/watt compared to Ampere even without much improvements on the architectural side. Unless something weird has happened of course, in the same veins as Vega 10 -> Vega 20 or something.
These 900W 4090 and 300W 4070 look all sorts of weird though.
Worth noting that Kopite7kimi has already went from a 400W 4070 down to a 300W one over the last five or so tweets. So perf/watt seem to be improving fast there ¯\_(ツ)_/¯
3070 is 220W while performing like 2080Ti which is 250W so there was some improvement at the very least.
TSMC states that N7 to N5 should result in 30% power reduction and general expectation for 8N to be considerably worse than N7 in power (if we look at Ampere vs Turing then it's not that much better than TSMC's 16FF+ to which N7 should be -60%; so about -80% or so?). Lovelace should have some very sizeable gains in perf/watt compared to Ampere even without much improvements on the architectural side. Unless something weird has happened of course, in the same veins as Vega 10 -> Vega 20 or something.
Last edited:
300 W for GeForce RTX 4070 makes 350 W for GeForce RTX 4080 nearly impossible. It has ~40 % bigger die and 16 GB GDDR6X instead of 12 GB GDDR6 (look at the TDP difference between GDDR6 RTX 3070 and GDDR6X RTX 3070 Ti). Nvidia will need to use quite high clocks for GA103 to close the gap between GA103 and GA102 (71 % more TFLOPS at the same clock). ~400 W for GeForce RTX 4080 seems to be more likely.
Don't think only about die size but transistor count. 250mm2 on N4 has more transistors than 500mm2 on Samsung N8
DegustatoR
Legend
Nah, this does sound very logical for a bleeding edge process with very expensive wafers. The smaller a die can be the better for your margins here.
I'm not sure where the "~40% bigger die" comes from though. Do we know die sizes for AD104/103/102 from somewhere?
Jawed
Legend
Is it useful to start with A100 as a baseline for TSMC 7nm to 5nm (4nm)?
NVIDIA A100 PCIe 80 GB Specs | TechPowerUp GPU Database
HBM is saving a lot of power. Boost clock of 1410MHz is rather low, but board power is 250W.
Weirdly, I've noticed that the L2 cache is stated as 80MB for this card, whereas the SXM version has 40MB:
NVIDIA A100 SXM4 40 GB Specs | TechPowerUp GPU Database
What's going on there?
NVIDIA A100 PCIe 80 GB Specs | TechPowerUp GPU Database
HBM is saving a lot of power. Boost clock of 1410MHz is rather low, but board power is 250W.
Weirdly, I've noticed that the L2 cache is stated as 80MB for this card, whereas the SXM version has 40MB:
NVIDIA A100 SXM4 40 GB Specs | TechPowerUp GPU Database
What's going on there?
Maybe the dies aren’t that small. The SM count for AD104 is pretty low but maybe they’re much fatter SMs + lots of L2 cache.
I’m now realizing just how terrible a product the 3070 Ti is. 35% higher power for 5% more performance than the 3070. And that’s all due to 8GB of GDDR6X because it doesn’t clock any higher than the 3070 with 8GB GDDR6.
Exactly.
Just my guess. AD103 has 40 % more SMs than AD104, 40 % more ROPs, 33 % more LLC and 33 % wider memory bus. Maybe it will be just 30 % larger, but the point is that AD103 is larger than AD104 and its clock won't be (likely) lower. Maybe it will be, but resulting performance gap between GA103 and GA102 would be huge. Anyway, with the impact of GDDR6X, I can't imagine how could AD103 / RTX 4080 consume just 50 watts (17 %) over AD104 / RTX 4070.
TopSpoiler
Regular
That must be a typo. Same mistake in the GA100 page: NVIDIA GA100 GPU Specs | TechPowerUp GPU Database
DegustatoR
Legend
Hence why it would be more probable for 4070 to not actually consume 300W and stick to 200-250W range instead. Would also make the "3090+20%" more plausible IMO at such wattage. Cause at 300W it should be closer to 3090+50% which seems rather high for a 4070 card, and with a 192 bit bus to boot.
Why use 4nm at all for <400mm^2 dies? GA103 is ~450mm^2, 256bit, 16gbit/s, 16GB and delivers 3070 performance within 175W. 450mm^2 on 8nm Samsung should be cheaper than 250mm^2 on TSMC's 5nm...
DegustatoR
Legend
That's a different question. It's probably considerably cheaper to make all chips for one production line? But I wouldn't rule out them using Samsung again for some AD107 GPUs.
- Status
Similar threads
