So is there a N5A for AMD? No. Maybe for Apple? No as well and Apple is TSMC's biggest customer.
So is there a N5A for AMD? No. Maybe for Apple? No as well and Apple is TSMC's biggest customer.N at the beginning is TSMC nomenclature.
N at the end is NVidia specific.
Anand mention of the 12nm FFN:
"built on TSMC’s still green 12nm “FFN” process (the ‘n’ stands for NVIDIA; it’s a customized higher perf version of 12nm for NVIDIA)."NVIDIA Volta Unveiled: GV100 GPU and Tesla V100 Accelerator Announced
www.anandtech.com
Though turns out it's just tweaked 16nm.
So is there a N5A for AMD? No. Maybe for Apple? No as well and Apple is TSMC's biggest customer.
Someone hijacked this account? I can't remember ever reading such from this for do long. Weird.You lie Tuna fish. you said all 5nm variants were the same. Even worse, you wrongly assumed customized processes were named for the companies who used them. They were created by TSMC and customers could choose which process they wanted to purchase on the 5nm silicon node. As the 5nm matured, they release newer more efficient processes on the node.
Nvidia GPU's are ultra efficient because of the 5nm process they are on. If AMD had RDNA3 on 4N, you would see power consumption numbers dramatically drop and probably increased performance as well. It's the cost for 4N that keeps AMD on the standard N5.
I cannot understand what you are trying to say.Someone hijacked this account? I can't remember ever reading such from this for do long. Weird.
This is one of the most stupid things people do when trying to OC their rdna 2/3 gpus, especially among those just switched from nvidia.Method seems to be set Max to 5000, Min to 2800
I got a 15% gain out of Timespy vs stock by setting a max frequency of 2900MHz, undervolting down to 950mv and OCing memory to 2700MHz. I can probably squeeze out a little more still, but I'm very power limited by the 314W TBP, so gains are going to be tied to mem OC and undervolts.TPU got nearly 3Ghz out of their nitro+
Method seems to be set Max to 5000, Min to 2800, max out the power slider and then keep undervolting until you are unstable then add a few more mv. Once the core is dialed you can clock the ram up. If core clocks seem to stick up the minimum clock.
With that they saw a 15% gain in timespy. Would love to see more games tested to see if that holds up in actual games.
Yeah, the "limited power limit" is the main issue here. Although 950 mv and +15% looks impressive (not sure if it's common for all rdna 3 sku's or it's more like carefully calibrated VF curve) and unlimited frequency ceiling really helps, my n21xt with the PPT and TDC throttlers completely turned off bumping in 2800 mhz hard ceiling at 410W ppt that is equivalent of (omg) ~ 500W TBP lolI got a 15% gain out of Timespy vs stock by setting a max frequency of 2900MHz, undervolting down to 950mv and OCing memory to 2700MHz. I can probably squeeze out a little more still, but I'm very power limited by the 314W TBP, so gains are going to be tied to mem OC and undervolts.
Even after this video , RTX4070 will sell 2-3x times the volume of 7800XT.
But that is because pricing isn't the same everywhere. Maybe it will change but for now I only found 2 listed 7800 xt here and they are more expensive than the cheapest 4070. That will likley change but still the difference in price is not >$100 everywhere. And we already know power cost will go up by 50% next year so the lower power use of the 4070 will make it cheaper if you keep it a couple years and include power costs.
But that is because pricing isn't the same everywhere. Maybe it will change but for now I only found 2 listed 7800 xt here and they are more expensive than the cheapest 4070. That will likley change but still the difference in price is not >$100 everywhere. And we already know power cost will go up by 50% next year so the lower power use of the 4070 will make it cheaper if you keep it a couple years and include power costs.
That's rich.You lie Tuna fish.
No, I didn't. Excerpt from my first post in this thread:you said all 5nm variants were the same.
(emphasis added)The GCDs on N31 and N32 use basically the same process as NV does.
Yes, all the big clients customize their processes a little, as in, in addition to compiling their designs into the TSMC-provided standard cells they also design and add a few cells of their own. The N4 line of processes has more, additional changes over that.This is not true. They both customize their processes, all the big clients do.
Even worse, you wrongly assumed customized processes were named for the companies who used them.
No, as I said, only nVidia. They insist on naming whatever they touch, probably as a branding thing. Others may point out that they customize their processes a little, but don't insist on new names.The only difference is that NV named theirs, while AMD uses the standard TSMC naming.
Yes, TSMC does this. All the various N5 and N4 processes are examples of this. But 4N is not one of these more advanced processes made by TSMC, it's the minor nVidia tweaks of N5.They were created by TSMC and customers could choose which process they wanted to purchase on the 5nm silicon node. As the 5nm matured, they release newer more efficient processes on the node.
This is very unlikely, because 4N in production predates N4P.There are probably more power efficiency gains in 4N over N4P.
A lot of forum members here lift data and pass it off as facts without citing sources or where they get their info. They assume that I google everything and come to conclusions because of what wikipedia says. I go by memory. Either something that I heard or something that I read. It goes into my long term memory or my medium recall memory.He's saying you're acting unusually, irrationally intransigent about this N4 / 4N thing in a manner that does not fit your usual MO @Hans Gruber .
You normally come off as a pretty even keel guy but for some reason this topic has you accusing others of *lying* and otherwise throwing down a kind of wildcat defense of this idea that Nvidia has some sort of vastly superior node under its belt.
Everyone thought or hoped AMD was delaying N32 to fix it. It turns out they did it... they fixed the price!Hell the 7800XT is the best release of 2023 in relationship to the previous product at the same MSRP.
Maybe it's just me, but citing MSRP for cards launched years ago as a benchmark comparison seems like a poor comparison to make, given the whole inflation thing. $500 in Jul 2019 would be equivalent to ~$610 today. It seems like you'd also need to take into account relative positioning of the cards (like a performance vs highest end/lowest end cards at the time) and the costs for those jumps, then adjust it all for inflation to get some more equal comparison across generations.OK looking at the TPU numbers, the RTX7800XT is the true RTX3070 successor.
RTX2070 Super - MSRP 500 USD
RTX3070 - MSRP 500 USD
RX7800XT - MSRP 500 USD
3070 vs 2070 Super = 32% faster at 1440p
7800XT vs 3070 = 32% faster at 1440p
NVIDIA GeForce RTX 3070 Founders Edition Review - Disruptive Price-Performance
NVIDIA's GeForce RTX 3070 is faster than the RTX 2080 Ti, for $500. The new FE cooler is greatly improved, runs very quietly, and even has fan stop. In our RTX 3070 Founders Edition review, we'll also take a look at RTX performance, energy efficiency, frametimes, and overclocking.www.techpowerup.com
AMD Radeon RX 7800 XT Review
With the Radeon RX 7800 XT, AMD is going after the GeForce RTX 4070. Our review confirms that AMD has achieved performance parity in rasterization and is pretty close in ray tracing, at a much better price point, and they are giving you 16 GB VRAM, instead of just 12 GB.www.techpowerup.com
I can buy a new 6800XT for £20-30 less than the cheapest 7800XT in the UK, it's actually cheaper long term too as the power costs in the UK are vastly greater than the US.Maybe it's just me, but citing MSRP for cards launched years ago as a benchmark comparison seems like a poor comparison to make, given the whole inflation thing. $500 in Jul 2019 would be equivalent to ~$610 today. It seems like you'd also need to take into account relative positioning of the cards (like a performance vs highest end/lowest end cards at the time) and the costs for those jumps, then adjust it all for inflation to get some more equal comparison across generations.
I can buy a new 6800XT for £20-30 less than the cheapest 7800XT in the UK, it's actually cheaper long term too as the power costs in the UK are vastly greater than the US.
Even $20-30 premium hardly makes 7800xt of less value than 6800XT (imo)What about next month ??? 6800XT is EOL, it will not be on the shelves for much longer
6800XT 300W TBP.I can buy a new 6800XT for £20-30 less than the cheapest 7800XT in the UK, it's actually cheaper long term too as the power costs in the UK are vastly greater than the US.
There is very little literature on the process optimizations that TSMC makes for each customer because they are minor tweaks in the process and TSMC doesn't share that information publicly. They only share the fundamental aspects of the process publicly. The general public, or typical Semicon news organizations are picking through the PDK to examine in detail the specifics of the design parameters that fabless customers must stay within. The small variations in process parameters, because company A wants a bit high performance for some type of functional block in their design matters not to another company that has different goals.I assume the reason why there is no real literature on the 4N process that Nvidia uses is due to a couple of reasons. Nvidia is probably the only customer using it because it's expensive. The other TSMC customers are waiting for 3nm. If the Zen 5 CPU's are released on a N4 variant. Keep the 22% power efficiency in mind when Zen 5 arrives. Power efficiency means power reduction/savings to do the same process.
I can buy a new 6800XT for £20-30 less than the cheapest 7800XT in the UK, it's actually cheaper long term too as the power costs in the UK are vastly greater than the US.