Radeon RX 6700, 6700 XT & 6750 XT users club

[Ruru]

all speculating:
maybe cuz the min (imposed by amd for reference board) power delivery is just enough for the actual max frequency and that you'd need an overbuilt powerdelivery.
so red devil, toxic and nitro+ would probably be the only ones to be able to do that, and it would need a hell of a lot o' work to do that properly wich and would prefer to invest this workload into rnda4/5/zen 6/zen5x3d

Hard to say. Pure guess but overclocking the VRAM without touching voltages shouldn't increase the power draw that much?

 

[datnoz]

Hard to say. Pure guess but overclocking the VRAM without touching voltages shouldn't increase the power draw that much?

power is W(att), voltage is V, amp is I
P=V*I
so the voltage may not fluctuate, but the amperage surelly will cuz you'r asking them more than what they're supposed to do.

 

[Ruru]

power is W(att), voltage is V, amp is I
P=V*I
so the voltage may not fluctuate, but the amperage surelly will cuz you'r asking them more than what they're supposed to do.

I know, I know. I still doubt that just pure OC without raising the voltage is so dramatic.

 

[datnoz]

I know, I know. I still doubt that just pure OC without raising the voltage is so dramatic.

You already know that but oc without raising voltage is a deal to get less stability

 

[Macro Device]

Just wondering that why AMD doesn't release drivers which would increase the maximum memory overclocking. It's not anymore like it would cannibalize the sales of RX 6800 or something.

It's much more simple than @datnoz said: AMD just don't need it. RDNA2 is near "nobody cares" so the only way to change it is to lower the pricing. Altering OC limits isn't worth an effort.
Also to note: RX 6800 is a whole third faster than 6750 XT so I don't know how you cannibalise that with barely existent performance boost from VRAM OC.

 

[Ruru]

You already know that but oc without raising voltage is a deal to get less stability

Of course, that's why a card without voltage control doesn't usually OC as much as a tune-able card. Or any component in overall.

It's much more simple than @datnoz said: AMD just don't need it. RDNA2 is near "nobody cares" so the only way to change it is to lower the pricing. Altering OC limits isn't worth an effort.
Also to note: RX 6800 is a whole third faster than 6750 XT so I don't know how you cannibalise that with barely existent performance boost from VRAM OC.

Is that so? But 6750 XT is not much but still noticeable faster and isn't that because of higher VRAM bandwith?

I may be wrong so I'm fine if I'm corrected here.

 

[Macro Device]

Is that so? But 6750 XT is not much but still noticeable faster and isn't that because of higher VRAM bandwith?

Most of that is thanks to higher GPU clock maintained due to a considerably higher power limit. VRAM helps but doesn't help THAT much. You just can't deny the power of 60 CUs of RX 6800, albeit they're at ~2200 MHz, unlike the 40 of 67x0 XTs, which are closer to 2550/2650 MHz.

 

[Nordic]

Just wondering that why AMD doesn't release drivers which would increase the maximum memory overclocking. It's not anymore like it would cannibalize the sales of RX 6800 or something.

Is it even useful? Overclocking or underclocking the VRAM on my 6750xt doesn't seem to have much of a measurable difference. I am not exactly doing scientific testing but it doesn't seem to matter much. Regardless, I run my VRAM at 2312 mhz which is as high as it will go.

 

[Lew Zealand]

Is that so? But 6750 XT is not much but still noticeable faster and isn't that because of higher VRAM bandwith?

I may be wrong so I'm fine if I'm corrected here.

Not sure about default clocks on the 6750 XT but I suspect as both GPUs were sold at the same time that the better binned GPU chips all went to the 6750 XT so it would clock the cores higher.

Which is needed as the 67x0 XT cards are the most VRAM-heavy and GPU core-constrained of all RDNA 2 GPUs. What the 67x0 XTs needed most to get more performance was more cores, not VRAM. So my suspicion is VRAM OC on the 67x0 XTs is of less consequence than other RDNA2 GPUs. OTOH the 6900 XT was core-heavy with the 6800 XT and 6800 less so but still more core-heavy than the 6700 XT.

So I suspect the 6750 XT's increased performance is more due to its 2489 to 2752MHz (10.6%) increase in GPU core clock speed rather than its 12.5% VRAM speed increase (numbers taken from reviews here).

 

[Macro Device]

the 67x0 XT cards are the most VRAM-heavy and GPU core-constrained of all RDNA 2 GPUs.

True that. +105 MHz on GPU (4% uplift) provide with ~3% framerate boost, whereas +150 MHz on VRAM (7.5% uplift) do only amount to about 0.8% additional FPS. Measured on my RX 6700 XT. 48 CU instead of 40 would've made this a cost efficient purchase, given the clock doesn't decrease.

 

[Lew Zealand]

True that. +105 MHz on GPU (4% uplift) provide with ~3% framerate boost, whereas +150 MHz on VRAM (7.5% uplift) do only amount to about 0.8% additional FPS. Measured on my RX 6700 XT. 48 CU instead of 40 would've made this a cost efficient purchase, given the clock doesn't decrease.

100%. These cards should have had 3072 cores to match the rest of the product stack. It bothers me that they don't so I try not to think about it.

Um.

Well also I bought a 7700 XT to some extent because they directly addressed this issue: 12GB VRAM/192-bit bus with the 6750 XT's bandwidth, but 3456 cores. That's a much better balance.

But... but... freakin' AMD. They cut the Infinity Cache in half. It's 50% faster so that's something I guess.

Still, I like the balance much better with the 7700XT though the price... well it wasn't too bad as I waited for a decent one.

 

[Ruru]

100%. These cards should have had 3072 cores to match the rest of the product stack. It bothers me that they don't so I try not to think about it.

Um.

Well also I bought a 7700 XT to some extent because they directly addressed this issue: 12GB VRAM/192-bit bus with the 6750 XT's bandwidth, but 3456 cores. That's a much better balance.

But... but... freakin' AMD. They cut the Infinity Cache in half. It's 50% faster so that's something I guess.

Still, I like the balance much better with the 7700XT though the price... well it wasn't too bad as I waited for a decent one.

Would've needed a cut-down 6800 die as 67x0 XT has a full chip.

Evem though I have a Greedhorse GPU, at least they don't have million SKUs with the same GPU...

 

[Lew Zealand]

Would've needed a cut-down 6800 die as 67x0 XT has a full chip.

Yeah this is where I wonder what the original plan was. Why is there a 2560 core and a 2048 core die? And the surrounding ones are 5120 and 1024. This seems nonsensical, I need an explanation. Changing the 2560 to 3072 seems to make more sense for market stratification, even with keeping the 6800 (non-XT) in the mix at 3840 cores.

My best guess is the 2048 is a simple doubled 1024 design and the 5120 the same doubling of the 2560. It could have cut down on design costs?

 

[Ruru]

Yeah this is where I wonder what the original plan was. Why is there a 2560 core and a 2048 core die? And the surrounding ones are 5120 and 1024. This seems nonsensical, I need an explanation. Changing the 2560 to 3072 seems to make more sense for market stratification, even with keeping the 6800 (non-XT) in the mix at 3840 cores.

My best guess is the 2048 is a simple doubled 1024 design and the 5120 the same doubling of the 2560. It could have cut down on design costs?

What makes me wondering is that the difference between 6700 and XT is minor when going against 6800 XT.

 

[RJARRRPCGP]

6700 non-XT, is suspected of being binned poorly. So don't be surprised if it OCs like dogwater! Or that it has less cores.

Of course, a major contribution, can very well be the card builder.

 

[Ruru]

6700 non-XT, is suspected of being binned poorly. So don't be surprised if it OCs like dogwater! Or that it has less cores.

Of course, a major contribution, can very well be the card builder.

Even that amount of disabled memory bandwith does things.

 

[RJARRRPCGP]

Even that amount of disabled memory bandwith does things.

Looks like the 675x has the better VRAM bandwidth. But believe it or not, it looks like even my 6750 XT, possibly can't hold a candle to my Arc A770, especially my Sparkle Titan! The Arc A770 is clocked lower, at only 2.5, and it looks like it can spank a 67x0 @ 2.9! The Arc A770, acts very much like an RX 6800 family GPU!

Just in case there's a driver bug that is causing slowness, I plan to do more benches with Adrenalin 24.6.x or later, when released.

 

[AusWolf]

Is it even useful? Overclocking or underclocking the VRAM on my 6750xt doesn't seem to have much of a measurable difference. I am not exactly doing scientific testing but it doesn't seem to matter much. Regardless, I run my VRAM at 2312 mhz which is as high as it will go.

Exactly my thoughts. VRAM overclocking is useless 95% of the time.

 

[Ruru]

I'm honest an I say that I'm drunk and I'm not participating for this anymore today

plz don't go too much offtopic, please?

 

[Lew Zealand]

Looks like the 675x has the better VRAM bandwidth. But believe it or not, it looks like even my 6750 XT, possibly can't hold a candle to my Arc A770, especially my Sparkle Titan! The Arc A770 is clocked lower, at only 2.5, and it looks like it can spank a 67x0 @ 2.9! The Arc A770, acts very much like an RX 6800 family GPU!

Just in case there's a driver bug that is causing slowness, I plan to do more benches with Adrenalin 24.6.x or later, when released.

Clocks are useless unless you also count cores and also IPC. My Radeon RX 6400 clocks to 2300 MHz but it's in no way faster than my 2060 Super which only manages about 2000 MHz. 768 vs 2176 cores matters.

Same for the 6700 XT with 2560 cores and the Arc A770 has 4096, 60% more. More cores usually better unless clocks are way different. However I see the 6700 XT beating the A770 on average so you must be playing a game that happens to work better than usual on the A770:

Exactly my thoughts. VRAM overclocking is useless 95% of the time.

This is true in most instances though a benefit of OC'ing VRAM is that it usually doesn't drive up power use by much more than a percent or two. So it's kinda like free performance and works well on some designs.

OK like compromised ones that are VRAM bandwidth-starved. My 3050 6GB (read: 3040 Ti) for instance with its 96-bit bus. I have one that OCs the VRAM by 26% and I get about 14% more performance from that and remember, this is a 70W design with no option for more power. UV the Cores to get about 10% more clock speed and total OC is about 17-18%, which is highly valuable in a lower performance power-limited card.

 

[datnoz]

Clocks are useless unless you also count cores and also IPC. My Radeon RX 6400 clocks to 2300 MHz but it's in no way faster than my 2060 Super which only manages about 2000 MHz. 768 vs 2176 cores matters.

Same for the 6700 XT with 2560 cores and the Arc A770 has 4096, 60% more. More cores usually better unless clocks are way different. However I see the 6700 XT beating the A770 on average so you must be playing a game that happens to work better than usual on the A770:

This is true in most instances though a benefit of OC'ing VRAM is that it usually doesn't drive up power use by much more than a percent or two. So it's kinda like free performance and works well on some designs.

OK like compromised ones that are VRAM bandwidth-starved. My 3050 6GB (read: 3040 Ti) for instance with its 96-bit bus. I have one that OCs the VRAM by 26% and I get about 14% more performance from that and remember, this is a 70W design with no option for more power. UV the Cores to get about 10% more clock speed and total OC is about 17-18%, which is highly valuable in a lower performance power-limited card.

totaly right, and (from my perspective) the most complete ansewer.

 

[AusWolf]

This is true in most instances though a benefit of OC'ing VRAM is that it usually doesn't drive up power use by much more than a percent or two. So it's kinda like free performance and works well on some designs.

OK like compromised ones that are VRAM bandwidth-starved. My 3050 6GB (read: 3040 Ti) for instance with its 96-bit bus. I have one that OCs the VRAM by 26% and I get about 14% more performance from that and remember, this is a 70W design with no option for more power. UV the Cores to get about 10% more clock speed and total OC is about 17-18%, which is highly valuable in a lower performance power-limited card.

Well, my 6500 XT overclocks like a dream. 2950 MHz on the core, 2400 on the memory with fast timings enabled. Unfortunately, +15% on top of not much is still not much (it's good for older games, though).

 

[datnoz]

talking about dream oc, i have a w2100 and i could put the slider all the way to the right (core and mem) is afterburner and it would still not crash and be prefectly stable.

 

[Macro Device]

it's good for older games

RX 590 is both cheaper and better. Of course you don't get anything resembling OK XeSS performance and it's more power hungry but still. RX 6500 XT and RX 6400 should be banned and AMD should be heavily fined for this crime against humanity.

 

[Lew Zealand]

Well, my 6500 XT overclocks like a dream. 2950 MHz on the core, 2400 on the memory with fast timings enabled. Unfortunately, +15% on top of not much is still not much (it's good for older games, though).

That's a nice overclock! My biggest annoyance about the RX 6400 is you get 2300 MHz Cores and 2000 VRAM. Locked, no fiddling. But I like fiddling! Same with my 5600XTs but at least you get VRAM OC which these need more than core OC.

RX 590 is both cheaper and better. Of course you don't get anything resembling OK XeSS performance and it's more power hungry but still. RX 6500 XT and RX 6400 should be banned and AMD should be heavily fined for this crime against humanity.

Sorry but no for the 6400. The HHSS versions are unique and ideal upgrades for space or power-constrained PCs. Niche, but a perfect fit. The full size models? I see no purpose for those. Of course price was always the real problem as both GPUs are fine and play games, just overpriced.