- Joined
- May 13, 2008
- Messages
- 669 (0.11/day)
| System Name | HTPC whhaaaat? |
|---|---|
| Processor | 2600k @ 4500mhz |
| Motherboard | Asus Maximus IV gene-z gen3 |
| Cooling | Noctua NH-C14 |
| Memory | Gskill Ripjaw 2x4gb |
| Video Card(s) | EVGA 1080 FTW @ 2037/11016 |
| Storage | 2x512GB MX100/1x Agility 3 128gb ssds, Seagate 3TB HDD |
| Display(s) | Vizio P 65'' 4k tv |
| Case | Lian Li pc-c50b |
| Audio Device(s) | Denon 3311 |
| Power Supply | Corsair 620HX |
So, I was watching this video from MLID. Pretty much backs up what I've been saying for an extremely long time. Always figured it would be 256-bit/GDDR6, but other sources made me have my doubts, because they *could've* done 192-bit GDDR7, although I agree it would have been ill-advised. 4080-like (if OC), 16GB, and cheap is name of the game and 192-bit would've meant 12GB (cold pourage) or 24GB (too hot, expensive pourage).
18GB I don't *think* would've been possible, as I think AMD will stick to Samsung/Hynix (which is shooting for [faster] 16Gb; 2GB chips wrt GDDR7) for the long haul. nVIDIA might as well buy Micron, as they appear pretty-well married.
His sources say 3.0-3.3ghz in the labs, right on target with my belief of ~3200mhz (round number + similar perf to 7900xt) to ~3264-3265mhz (which if the same cache config of N32, would be the bandwidth limitation of 24gbps RAM).
I also think die sizes will be on the smaller side of what his sources imply (I especially think N33 redux will be tiny; almost too small to make sense [like AD107], but there needs to be a cheap discrete/laptop 16GB 'ps5/XSX-like' option).
There's a weird thing to explain about clocks being better than more units (clocks scale perf at ~99% while units typically ~96-97% or so) if power permits. Maybe worth explaining; mb not. Anyway, his perf target is very accurate imo.
Where we diverge is RAM speed/availability. Some have speculated 24gbps (at least at first) is a long shot bc of yields, etc, and I'll explain why it isn't. That's not to say AMD couldn't change cache and use slower mem, but FWIW:
First of all, chips generally say "sampling" until a customer releases a part with it, and sometimes long afterwards. There was a time this wasn't the case, but I can only assume one or more company became angry when people would look at datasheets to extrapolate the next products coming from ATi/AMD/nVIDIA. Hynix went so far as to HIDE (available only by business request) their datasheets/databook/catalogue (which was publically available and used to show dates/processes/densities of upcoming chips), which is unfortunate, and it was very obviously because of stuff like that. It's also worth noting when parts are generally made FOR a customer/product spec, it doesn't really make sense to list it before the BOM is ready for such a product. Usually "sampling" just means they *can* make it to order, but currently there is not enough demand to list it. This would obviously change if AMD built a card BOM around it.
Second of all, Hynix (widely used on most AMD cards because decent-performing and often cheaper than Samsung) uses 1.35v, the industry standard, for all parts up to and including their (not) new/'sampling' 20gbps.
Samsung, OTOH, uses 1.1v for 20gbps, as stated in this press release about 24gbps memory (fifth body paragraph). It certainly does not say 24gbps uses 1.1v; obviously the extra VDD (.25v is a lot) is used to make it feasible.
Some likely know where this is going, but the end may blow your mind (or not, because I've seen nVIDIA do this so many times before. While I'm just a weirdly-devoted hobbyist, without a doubt nVIDIA have people like me on payroll).
If you extrapolate 1.35v/1.1, you end up with something like 22-23%. That alone covers the difference between 20->24gbps if it scales, which it almost-surely does (to industry spec). While I don't have any of those cards in front of me atm, looking at how Hynix performs on a 7700xt/7800xt/7900 series, I would assume they are running different voltages (something like 1.25->1.3->1.35v?), but still the industry spec.
Furthermore, let's talk about binning.
With Hynix, every (7900 series) card you see can clock to 2750mhz (22gbps) @ 1.35v. This is likely where they are binned. Some companies (like Asus) probably bin them further (2800mhz @ 1.35v, 2650 @ 1.3v, etc).
With the sole Samsung card on TPU, you see 2720mhz, which has lead me to believe Samsung bins their parts at 2700mhz, or 200mhz above their rating.
Now, let's extrapolate. If (3ghz) 24gbps RAM is binned 200mhz above it's rating, they would work until 25.6gbps, which again I have said *should* scale to ~3482mhz core. This coincides very well with AMD's love of ~6% overclocking.
Let's say, for whatever reason, we take 1.35/1.1v, but then take that times this (albeit singular) overclocking result of 2720mhz, which would leave us with ~26.7gbps.
Which is to say that 27gbps likely is a bad yield, and perhaps true to the statements made about 24gbps. You'll notice Samsung pulled back on 27gbps, and that's likely why. Perhaps they *could* make it, but it's probably not feasible.
As I've said before, if we simply take 375w and extrapolate from 7800XT, it's very possible core could hit ~3500mhz. They don't need to though, as they really only need to plan for ~3482/25600 (or ~3200/24000, if you prefer stock).
BUT. BUT. And this is the fun part: What if AMD let partners go wild, add that second 8-pin they don't need, and don't clamp core (voltage/power)/memory clocks?
Well, it's interesting you (probably didn't) ask. If the core/ram could scale to this hypothetical 3632-3633mhz/(3338) 26704mhz... and remembering the stock clock is probably ~3200mhz and 7900xt perf....Well...I'll let you do the math:
Yes. nVIDIA truly does figure this stuff out in advance. I can't prove it, but as someone whom crunches these numbers like almost-nobody else (that doesn't do it for a living), I can say I've seen this EXACT thing before...many times.
nVIDIA knew what AMD was planning; they know the max they're capable of doing. They plan stock perf of cards (like 4080) so they can say, 'on avg', even at stock vs an overclock, they are a 'better' part at something (this time 4k).
AMD could go all-out OC vs OC and similar at 1440p, with perhaps 4080 squeaking out another of the most meaningless of victories, but again, it makes more sense they'd simply shoot for an OC catching up to 4080 at stock (1440p).
AMD likely won't go balls-out, as I doubt they want (most) cards to go over 375w, there-by adding another power connector and looking even less power-efficient. Let partners do that, if they want, and some surely will.
That lack of (AMD-winning) OC overlap allows nVIDIA to keep a large margin on products, especially for those that don't consider overclocking (or the effect it has on performance), ntm pay devs to target (example) 60fps on their cards that AMD couldn't quite hit (without lowering some non-important settting one notch).
And they can do that. That's their right. It's then my right to show up and say it's actually 4070/Super/Ti competition that will perform very damn close to a 4080 (esp if you OC); probably good-enough for a lot of people for less $.
That has to happen so we can have the next fight: The one where nVIDIA undoubtably would like to release 5070 with 12GB of RAM (for probably $600) while slightly outperforming Navi 4/7900xt in raw raster.
I will fight until we hear that part will have 18GB, and if it doesn't, I will then fight the same contrived planned-obsolescence fight I have for years, and either way watch AMD's price become even cheaper, then probably buy one.
As was said, hats off to people that have/will just buy a 7900xt and be done with it, even if spending a little more right now. For the rest of us, the wait continues for reasonably-priced high-end (compared to console) performance.
18GB I don't *think* would've been possible, as I think AMD will stick to Samsung/Hynix (which is shooting for [faster] 16Gb; 2GB chips wrt GDDR7) for the long haul. nVIDIA might as well buy Micron, as they appear pretty-well married.
His sources say 3.0-3.3ghz in the labs, right on target with my belief of ~3200mhz (round number + similar perf to 7900xt) to ~3264-3265mhz (which if the same cache config of N32, would be the bandwidth limitation of 24gbps RAM).
I also think die sizes will be on the smaller side of what his sources imply (I especially think N33 redux will be tiny; almost too small to make sense [like AD107], but there needs to be a cheap discrete/laptop 16GB 'ps5/XSX-like' option).
There's a weird thing to explain about clocks being better than more units (clocks scale perf at ~99% while units typically ~96-97% or so) if power permits. Maybe worth explaining; mb not. Anyway, his perf target is very accurate imo.
Where we diverge is RAM speed/availability. Some have speculated 24gbps (at least at first) is a long shot bc of yields, etc, and I'll explain why it isn't. That's not to say AMD couldn't change cache and use slower mem, but FWIW:
First of all, chips generally say "sampling" until a customer releases a part with it, and sometimes long afterwards. There was a time this wasn't the case, but I can only assume one or more company became angry when people would look at datasheets to extrapolate the next products coming from ATi/AMD/nVIDIA. Hynix went so far as to HIDE (available only by business request) their datasheets/databook/catalogue (which was publically available and used to show dates/processes/densities of upcoming chips), which is unfortunate, and it was very obviously because of stuff like that. It's also worth noting when parts are generally made FOR a customer/product spec, it doesn't really make sense to list it before the BOM is ready for such a product. Usually "sampling" just means they *can* make it to order, but currently there is not enough demand to list it. This would obviously change if AMD built a card BOM around it.
Second of all, Hynix (widely used on most AMD cards because decent-performing and often cheaper than Samsung) uses 1.35v, the industry standard, for all parts up to and including their (not) new/'sampling' 20gbps.
Samsung, OTOH, uses 1.1v for 20gbps, as stated in this press release about 24gbps memory (fifth body paragraph). It certainly does not say 24gbps uses 1.1v; obviously the extra VDD (.25v is a lot) is used to make it feasible.
Some likely know where this is going, but the end may blow your mind (or not, because I've seen nVIDIA do this so many times before. While I'm just a weirdly-devoted hobbyist, without a doubt nVIDIA have people like me on payroll).
If you extrapolate 1.35v/1.1, you end up with something like 22-23%. That alone covers the difference between 20->24gbps if it scales, which it almost-surely does (to industry spec). While I don't have any of those cards in front of me atm, looking at how Hynix performs on a 7700xt/7800xt/7900 series, I would assume they are running different voltages (something like 1.25->1.3->1.35v?), but still the industry spec.
Furthermore, let's talk about binning.
With Hynix, every (7900 series) card you see can clock to 2750mhz (22gbps) @ 1.35v. This is likely where they are binned. Some companies (like Asus) probably bin them further (2800mhz @ 1.35v, 2650 @ 1.3v, etc).
With the sole Samsung card on TPU, you see 2720mhz, which has lead me to believe Samsung bins their parts at 2700mhz, or 200mhz above their rating.
Now, let's extrapolate. If (3ghz) 24gbps RAM is binned 200mhz above it's rating, they would work until 25.6gbps, which again I have said *should* scale to ~3482mhz core. This coincides very well with AMD's love of ~6% overclocking.
Let's say, for whatever reason, we take 1.35/1.1v, but then take that times this (albeit singular) overclocking result of 2720mhz, which would leave us with ~26.7gbps.
Which is to say that 27gbps likely is a bad yield, and perhaps true to the statements made about 24gbps. You'll notice Samsung pulled back on 27gbps, and that's likely why. Perhaps they *could* make it, but it's probably not feasible.
As I've said before, if we simply take 375w and extrapolate from 7800XT, it's very possible core could hit ~3500mhz. They don't need to though, as they really only need to plan for ~3482/25600 (or ~3200/24000, if you prefer stock).
BUT. BUT. And this is the fun part: What if AMD let partners go wild, add that second 8-pin they don't need, and don't clamp core (voltage/power)/memory clocks?
Well, it's interesting you (probably didn't) ask. If the core/ram could scale to this hypothetical 3632-3633mhz/(3338) 26704mhz... and remembering the stock clock is probably ~3200mhz and 7900xt perf....Well...I'll let you do the math:
Yes. nVIDIA truly does figure this stuff out in advance. I can't prove it, but as someone whom crunches these numbers like almost-nobody else (that doesn't do it for a living), I can say I've seen this EXACT thing before...many times.
nVIDIA knew what AMD was planning; they know the max they're capable of doing. They plan stock perf of cards (like 4080) so they can say, 'on avg', even at stock vs an overclock, they are a 'better' part at something (this time 4k).
AMD could go all-out OC vs OC and similar at 1440p, with perhaps 4080 squeaking out another of the most meaningless of victories, but again, it makes more sense they'd simply shoot for an OC catching up to 4080 at stock (1440p).
AMD likely won't go balls-out, as I doubt they want (most) cards to go over 375w, there-by adding another power connector and looking even less power-efficient. Let partners do that, if they want, and some surely will.
That lack of (AMD-winning) OC overlap allows nVIDIA to keep a large margin on products, especially for those that don't consider overclocking (or the effect it has on performance), ntm pay devs to target (example) 60fps on their cards that AMD couldn't quite hit (without lowering some non-important settting one notch).
And they can do that. That's their right. It's then my right to show up and say it's actually 4070/Super/Ti competition that will perform very damn close to a 4080 (esp if you OC); probably good-enough for a lot of people for less $.
That has to happen so we can have the next fight: The one where nVIDIA undoubtably would like to release 5070 with 12GB of RAM (for probably $600) while slightly outperforming Navi 4/7900xt in raw raster.
I will fight until we hear that part will have 18GB, and if it doesn't, I will then fight the same contrived planned-obsolescence fight I have for years, and either way watch AMD's price become even cheaper, then probably buy one.
As was said, hats off to people that have/will just buy a 7900xt and be done with it, even if spending a little more right now. For the rest of us, the wait continues for reasonably-priced high-end (compared to console) performance.
Last edited:
