Oh man this brings back some memories. I remember doing this with an AMD card a long time ago as well, but the actual model eludes me.
Probably ruins the per/W though.

Note - just increased power and frequency. You can not enable more internal calculation units.

Note - Idle is not stable yet.
Note - was done with an ASUS graphic card
Source: german text

NostrasOh man this brings back some memories. I remember doing this with an AMD card a long time ago as well, but the actual model eludes me.
Probably ruins the per/W though.

It most certainly does, X800 Pro (12 pixel pipes) flashed to X800 XT for full 16 pixel pipes and clocks to boot. I remember those days well!

HD6950 flashed to 6970 brings back fond memories also.

This was quite common with RX 5700 and going with a RX 5700XT vbios. Unfortunately for me, the Pulse cooler on my RX 5700 was a bit too weak. The fans turned out to be a bit too noisy for my taste and the temperatures weren't to my liking

While not a GPU, i remember being able to enable the extra core on my Phenom X3 720. Always Good Fun!

There's no point in making modifications that could invalidate your warranty when you can already get an extra 10-15% of performance from OC.

DenverThere's no point in making modifications that could invalidate your warranty when you can already get an extra 10-15% of performance from OC.

Well this is probably not giving you only 10% gain, but maybe extra 20% which should make it worth while.

NostrasOh man this brings back some memories. I remember doing this with an AMD card a long time ago as well, but the actual model eludes me.
Probably ruins the per/W though.

6950 to 6970. :)

I just noticed something. N48 doesn't even scale with clocks(18%), let alone throughput (about 34% for shaders and TMU), And since most games leave at least an entire GB of RAM free(from what I saw most games sit at 12-14GB on the 9070XT at 4K), I'm guessing the chip is bandwidth starved. Meaning I'm sure a 9070XTX(since they abandoned the GDC + IO design that means a brand new chip) with 28 Gbps GDDR7 should be an par with the 5080. Hell maybe even returning to 128MB of L3 should close the gap

FishymachineI'm guessing the chip is bandwidth starved.

That was obvious from the very beginning. You can try cheating by using fast cache and whatnot but you can't fool the math. 20 GT/s over a 256-bit bus just isn't enough.

NostrasOh man this brings back some memories. I remember doing this with an AMD card a long time ago as well, but the actual model eludes me.
Probably ruins the per/W though.

The earliest I can recall was the Radeon 9800 SE to 9800 Pro back in 2003. At least that's the only time I had any luck.

NostrasOh man this brings back some memories. I remember doing this with an AMD card a long time ago as well, but the actual model eludes me.
Probably ruins the per/W though.

Would that be the ATI Radeon 9500 Pro? Some of them can be flashed with the bios from the Radeon 9700 Pro to become a working Radeon 9700 Pro. There were some lower tier cards in the X800 and X850 series that can also be flashed with higher tier cards' bios to unlock additional rendering pipelines.

It was certainly way more fun back in the day when ATI was around, then AMD bought them and ruined all the fun. :rolleyes:

Outback BronzeIt most certainly does, X800 Pro (12 pixel pipes) flashed to X800 XT for full 16 pixel pipes and clocks to boot. I remember those days well!

HD6950 flashed to 6970 brings back fond memories also.

Had those, as well as the Vega 56 which you could flash to a Vega 64, and pump up the memory by 33% and undervolt the gpu part so it runs at ~200W with nonstop turbo clocks and massive performance boost from the memory overclock.

I literally called this before it was even released (due to it's price/specs).

I knew this had to be true. I wanted to ask W1z to try it/flaunt it, but couldn't find the nerve. I also didn't have the guts to tell people the thing below, but I will now bc why not.

FishymachineI just noticed something. N48 doesn't even scale with clocks(18%), let alone throughput (about 34% for shaders and TMU), And since most games leave at least an entire GB of RAM free(from what I saw most games sit at 12-14GB on the 9070XT at 4K), I'm guessing the chip is bandwidth starved. Meaning I'm sure a 9070XTX(since they abandoned the GDC + IO design that means a brand new chip) with 28 Gbps GDDR7 should be an par with the 5080. Hell maybe even returning to 128MB of L3 should close the gap

Accurate about bandwidth, which is true of almost every AMD design. Ram would be 24gbps GDDR6 though, and with overclocking shouldn't really even need bigger L3. Just a whole fuck ton of power.
It should really only need ~3850mhz or so for next tier. Hmm. They could probably do that within 525w (or 3x8-pin), maybe even with 32GB. 4ghz...well, the Alienware Poppa Edition would certainly require ~600w.
Probably. If it were to exist. Which certain people say does not.

Want some homework, fellas? Maybe I can teach you a little bit about electrical engineering.

Count the MLCCs on N48.

There are 98 big caps (X2Y?), and 26 smaller caps. The small caps are (probably) rated for stabilizing 16mhz. The large caps (I think) 2.2x that.
There is space for 10 more small caps (left-hand middle) that are not attached to the pads. I don't know if N4X respin, that's the bw limit of 24gbps GDDR6 (even overclocked), or any other reason they did that.
Maybe they were hoping 27gbps GDDR6 would actually exist, or 24gbps would get close...and they ended up using what fit the other limitations.

What does that tell you, though? :p
Certainly they're doing something.

Like the bios thing, I kind of explained this a long time ago as well. Not inside info; I have no idea what they will actually productize. Like everything else I write, just know I am just that big of a nerd.

FishymachineI just noticed something. N48 doesn't even scale with clocks(18%), let alone throughput (about 34% for shaders and TMU), And since most games leave at least an entire GB of RAM free(from what I saw most games sit at 12-14GB on the 9070XT at 4K), I'm guessing the chip is bandwidth starved. Meaning I'm sure a 9070XTX(since they abandoned the GDC + IO design that means a brand new chip) with 28 Gbps GDDR7 should be an par with the 5080. Hell maybe even returning to 128MB of L3 should close the gap

I would say that it doesn't even need GDDR7, just faster GDDR6.
This is old news, but it seems that everyone is content with using 20 Gbps GDDR6. Hey, whatever is cost-effective right?
semiconductor.samsung.com/news-events/news/samsung-electronics-develops-industry-first-24-gbps-gddr6-dram-to-power-next-generation-technologies-needing-ultra-high-speed-and-large-processing-capacity/
If it had 24 Gbps GDDR6 we would surely see some improvements.

Macro DeviceThat was obvious from the very beginning. You can try cheating by using fast cache and whatnot but you can't fool the math. 20 GT/s over a 256-bit bus just isn't enough.

I'm not saying it isn't bottlenecked to some degree but that fast cache voodoo actually works, the proof is that the 9070 XT, a card with 645 GB/s (slighty higher than the 7800 XT, lol) bandwidth is able to get (quite) close to cards that have (sometimes significantly) higher bandwidths:
5070 Ti - 896 GB/s
4080 - 717 GB/s
4080 Super - 736 GB/s
7900 XTX - 960 GB/s
5080 - 960 GB/s (this one is clearly stronger but not due to bandwidth alone)

Darmok N JaladThe earliest I can recall was the Radeon 9800 SE to 9800 Pro back in 2003. At least that's the only time I had any luck.

Same went from 9800 Pro to XT. Did that to mine.

Sol_Badguythat fast cache voodoo actually works

It works but it isn't always enough. Some games demand pure and distilled performance of VRAM chips alone and there we only got just a little more than crumbs. I know it would've made a GPU significantly more expensive but going for 20 GB over a 320-bit bus, even if it stays 20 GT/s, should've made all the difference and we could've had a GPU that's faster than 5080 in more than just one scenario. The main problem of 9070 series is that it doesn't bring enough cost efficiency to the table, it's still a horribly expensive thing to buy because it's way over MSRP in 95% of the countries. I've seen 9070 XT selling for $1000 (and that's the cheapest retail SKU in the country, also without VAT that's 20% in this particular case!) last week and it tells how bad it is. Not a bad product, just... beyond expensive.

HereticbarWould that be the ATI Radeon 9500 Pro? Some of them can be flashed with the bios from the Radeon 9700 Pro to become a working Radeon 9700 Pro. There were some lower tier cards in the X800 and X850 series that can also be flashed with higher tier cards' bios to unlock additional rendering pipelines.

That's how I got involved with the hardware scene, good memories :)

In this case, all you are getting is the increased power limit and that's it. Not a procedure worth risking your card for, given that there are safety measures employed to prevent the user from attempting to do this and the problems it introduces - by the original poster's own recount, the hardware is not stable and cannot be reliably operated. The 9070 and 9070 XT use the same memory, so unlike with the Vega 56 (and this was exclusive to Samsung HBM models) where you could flash the 64's BIOS, you won't get better internal memory timings, and just like with Vega, no extra compute units are unlocked.

Useful for extreme enthusiasts attempting to benchmark - nothing to see here for anyone else.

It's been awhile since I've seen interesting OC's like this. 20% extra performance on a 9070 is pretty nutty. I miss old OC's man... but they're probably gone (besides here) for a good reason.

Sol_BadguyI would say that it doesn't even need GDDR7, just faster GDDR6.
This is old news, but it seems that everyone is content with using 20 Gbps GDDR6. Hey, whatever is cost-effective right?
semiconductor.samsung.com/news-events/news/samsung-electronics-develops-industry-first-24-gbps-gddr6-dram-to-power-next-generation-technologies-needing-ultra-high-speed-and-large-processing-capacity/
If it had 24 Gbps GDDR6 we would surely see some improvements.

I'm not saying it isn't bottlenecked to some degree but that fast cache voodoo actually works, the proof is that the 9070 XT, a card with 645 GB/s (slighty higher than the 7800 XT, lol) bandwidth is able to get (quite) close to cards that have (sometimes significantly) higher bandwidths:
5070 Ti - 896 GB/s
4080 - 717 GB/s
4080 Super - 736 GB/s
7900 XTX - 960 GB/s
5080 - 960 GB/s (this one is clearly stronger but not due to bandwidth alone)

As I have explained many, many, times...nVIDIA does not need all that bandwidth. It's mostly a gimmick (although excess bw does help perf a small amount...double excess usually equals around 16% perf increase).

The humorous parts are that a stock 5080FE (2640mhz static) needs exactly 22gbps (given their cache). It's general overclocking potential, 3154mhz, would therefore require ~26.283_gbps.
What is the general limit of (overclocked) 24gbps GDDR6? I don't know, but I bet nVIDIA does. What is the general limit (especially imposed by AMD) of 20gbps (Answer: 21600mhz).

So I bet 24gbps is damn close to 5080's allowed speed. Also, let's just say for arguments sake the RT speed of 2.97ghz of 9070xt is the bw limitation of 20gbps GDDR6. That means 24gbps would be ~3.564ghz.
N48 also has a tendancy to boost about 6-8%, which you could probably stabilize with enough power.

Now, go pick your favorite TPU review and look at 1440p mins for something like Wukong, where something like a ~2800mhz 9070 xt is 44.6 and a 2640mhz 5080 is 52.6.
Where a general 5080 overclock will not keep 60fps (absolutely by design, as I keep telling people). This is because this fight is yet to come. Probably with 5080 actually having enough ram.

But I still think it's stupid, because even if these chips can pull off something like ~1440p60, which they may or may not, I bet they still won't keep 60 with 4k quality up-scaling. Maybe the goal will be 48.

I will buy the thing that can do both, and I still expect to be waiting until next generation. But it's a cool trick for small chips to achieve that market (potentially), although use (especially AMD) a lot of power.
AMD will likely be substantially cheaper because of that, though. I'm sure they *want* to directly compete, but I wonder if they'll end up facing off with 5070TI in price (where-as 9070/xt will drop in price a little).
To me that makes a lot of sense, especially when you figure the perf difference and the potential cost of faster and/or more GDDR6. Maybe 16/32GB (or 3.85 vs 4ghz), some kind of split, fight 5070ti/5080 24GB.

As I've said before, I feel both companies' models are refreshes once they absorb all the margins from people that wanted the early cards. Hence why I think 9070/xt will end up cheaper in the long run.

I think W1z thinks this too, at least in some respect. That's why you see estimates like perhaps 9070 being $350 by some point; everything needs to fit and make sense in the stack.

W1zzardThat's how I got involved with the hardware scene, good memories :)

Same. Well, mostly because of stuff like that.
I really don't think AMD understands how many people that brought into the community back in the day; hence my tirade about it a while back about them losing their soul since becoming more like nVIDIA.
Glad to see they've still not completely lost it.
:toast:

See, the thing *most* people don't realize is that a mod like this is (potentially) *just* enough to make 9070 not suck. It means where 9070 xt could be 60, 9070 could be 48fps...which is how it *should* be.
Those that don't understand, that's fine. Those that do really appreciate it. AMD needs that later group. To explain to people you can now run 1440p games with decent settings/rt/whatever and stay in VRR.
That it is now *actually* better than a 5070, which does in-fact suck, and this card (with this mod) would not suck as much in a very useful way.
Those are the people that have bought AMD parts forever, since BH-5, pin mods, and all the rest; just like they did ATi. Because they find, understand, and appreciate those things.
They can't afford to lose those people, in both financial and vocal support.

Outback BronzeIt most certainly does, X800 Pro (12 pixel pipes) flashed to X800 XT for full 16 pixel pipes and clocks to boot. I remember those days well!

HD6950 flashed to 6970 brings back fond memories also.

HL2 & I had a blast, speaking of... wonder when the full RTX version comes out, the demos are fun and all... but you can only play and oogle at the graphics so many times. :))

Reminds me vega56, when people flashed bios from vrga 64.

alwaysstsI really don't think AMD understands how many people that brought into the community back in the day; hence my tirade about it a while back about them losing their soul since becoming more like nVIDIA.

The big issue for the GPU vendors are fakes (not unlockable cards)

aah, i remember flashing my 5850 with a 5870 bios, ah the memories :toast: