Review AMD Radeon RX 7600 XT review: double the memory and higher clocks, still Navi 33

[Admin]

We tested Sapphire's AMD RX 7600 XT Pulse graphics card, which doubles the memory and boosts the clocks and power limits compared to the vanilla 7600, but still utilizes the same Navi 33 GPU. Performance has improved, but there are better alternatives at this price.

AMD Radeon RX 7600 XT review: double the memory and higher clocks, still Navi 33 : Read more

 

[Amdlova]

Just buy a rx 6700xt same price way better...
With some undervolt will have same TBP.

 

[Roland Of Gilead]

AMD's version of the 4060ti! Disappointing it comes behind even the Arc770.

 

[evdjj3j]

I'm not sure why anyone would chose this over a 4060.

 

[Gururu]

Thank you been waiting on this review. Certainly helps with the decision paralysis.

 

[btmedic04]

This would have been a much more compelling product had it used a further cut down Navi 32 with 3x mcds, 12gb of vram and a $350ish price point. Disappointing

 

[King_V]

AMD's version of the 4060ti! Disappointing it comes behind even the Arc770.

Well, AMD's version of the 4060ti 16GB, I would say... although, to be fair, it does give SOME performance benefit over the 7600 non-XT. The 4060 Ti 16GB gave just about nothing over the 4060 Ti 8GB (for gaming purposes, an edge case or two notwithstanding)

 

[JarredWaltonGPU]

This would have been a much more compelling product had it used a further cut down Navi 32 with 3x mcds, 12gb of vram and a $350ish price point. Disappointing

I do wonder if AMD may eventually try doing an RX 7700 non-XT with those specs, but realistically I think the total cost of 3xMCD plus the Navi 32 GCD means making such a card and selling it at $350 is a losing proposition.

 

[btmedic04]

I do wonder if AMD may eventually try doing an RX 7700 non-XT with those specs, but realistically I think the total cost of 3xMCD plus the Navi 32 GCD means making such a card and selling it at $350 is a losing proposition.

Yeah, it's more than likely too expensive to hit a $350 price point and still have any kind of profit margin. One can still dream at least lol.

 

[Roland Of Gilead]

Well, AMD's version of the 4060ti 16GB, I would say... although, to be fair, it does give SOME performance benefit over the 7600 non-XT. The 4060 Ti 16GB gave just about nothing over the 4060 Ti 8GB (for gaming purposes, an edge case or two notwithstanding)

True

 

[usertests]

+$60 is not a huge premium over the 7600 for the extra memory, but the 7600 is overpriced in the first place.

It's not much slower than the 6700 XT, but the 6700 XT is clearly better and 12 GB VRAM is fine. If you want the 7600 XT to play with AI, good luck.

Get it down to $270-280, what the 6700 10 GB used to cost, and it looks more interesting. Although 6700 XT was as low as $295-300 (new) in Nov-Dec. Something tells me they will run out before they return to that price point.

 

[DavidLejdar]

"...keep in mind that ray tracing support is becoming pretty common from the major engines and games."

Yeah. But at not a low price point, e.g. with Alan Wake 2 the RTX 4070 Ti does 17.4 fps at 4K with RT (and 82.2 when DLSS 3.5, with upscaling, is used).
https://www.tomshardware.com/news/alan-wake-2-4090-130-fps-4k-maxed-out-ray-tracing

I don't know... Sounds like a quite a performance hog, and I don't need every game to be photorealistic, as I do enjoy different art styles - see e.g. Two Point games, where photorealism would seem out of place. (And new major games don't always step much beyond the top console capabilities.)

I mean, maybe it is cool, and if e.g. a retailer would show side-by-side how it looks in "4K native without RT" and "upscaled 4K with RT", I could change my mind. And for someone mostly interested in 1080p gaming, as to help with aiming (and not needing 200+ fps), ray-tracing sure sounds like a nice boost.

But as it is, in my case, I rather first go with a 10-bit screen (and HDR), which adds quite some to visual output.

 

[Colif]

No article about fact AFMF went live today with 24.1.1.?

https://www.amd.com/en/support/kb/release-notes/rn-rad-win-24-1-1

I might need to find a game to play now.

 

[Notton]

This is one of those cards where 10~12GB would have been plenty, were there a 3GB memory chip.

 

[JarredWaltonGPU]

No article about fact AFMF went live today with 24.1.1.?

https://www.amd.com/en/support/kb/release-notes/rn-rad-win-24-1-1

https://www.tomshardware.com/pc-com...-at-1080p-in-first-full-radeon-driver-release

That's also linked in this review, and I mentioned giving AFMF a try in a couple of games. It... didn't feel very nice. Lots of stuttering, worse minimum FPS than without AFMF. Maybe some games will respond better, but I had intentionally picked a couple of demanding RT games to see if it could make them more viable.

 

[usertests]

This is one of those cards where 10~12GB would have been plenty, were there a 3GB memory chip.

Imagine if we see 9 GB GPUs (96-bit) replacing the 8 GB GPUs next gen.

 

[eepok]

Hi Jared - Are any of the benchmark values done for the 7600 XT or new 4070s compatible with the geometric mean fps scores in the GPU Hierarchy or are you using a different set of benches?

 

[Eximo]

Imagine if we see 9 GB GPUs (96-bit) replacing the 8 GB GPUs next gen.

That would end up being a 12GB or 6GB card, which is what we see today with the A380. 6GB across 96 bits.

 

[JarredWaltonGPU]

Hi Jared - Are any of the benchmark values done for the 7600 XT or new 4070s compatible with the geometric mean fps scores in the GPU Hierarchy or are you using a different set of benches?

The GPU hierarchy tables are currently based on 12900K testing, which I have done on the new GPUs... and discovered things have changed with Windows (or maybe the games, but probably Windows 11) and I may need to retest a bunch of cards. About half of the games seem to be more CPU limited now.

Anyway, long story short, I'm going to be working to switch all the hierarchy over to the new test bed in the coming days. I still have several older GPUs to test, but I may do the switch without them (and leave the current table as "legacy" status).

But the review Geomean uses 9 rasterization and 6 DXR, while the hierarchy has 8 rasterization and 5 DXR, so the geomeans are not directly comparable. Also, the hierarchy scores are actually a combination of average and 1% low fps (I believe it's the geomean of four average and one minimum, so that the average has heavier weighting, but I'd need to double check to be sure), so the table FPS values don't directly match even older reviews that used the 12900K test PC.

Oh what a tangled web we weave.... 🕷️

 

[JarredWaltonGPU]

That would end up being a 12GB or 6GB card, which is what we see today with the A380. 6GB across 96 bits.

I'm pretty sure he's talking about non-binary VRAM capacities. We have those on DDR5 DIMMs (24GB and 48GB), and it's possible to do this for other memory types. I doubt it will happen in the next cycle, though. There's extra work for the non-binary DIMMs that causes them to run a bit hotter, and I suspect GPUs would prefer to not add more heat if possible.

But if it does happen, I'm mostly okay with the principle. 96-bit interfaces would be a problem, but 12GB using four chips on a 128-bit interface would be better than 8GB and easier than eight chips and 16GB like the 7600 XT. The 192-bit designs would be 18GB instead of 12GB, which would also be great. Then 256-bit would have 24GB and 384-bit would jump to 36GB.

Nvidia's RTX 4070 Ti Super proves that the bandwidth limitation isn't a serious problem in most workloads, and some further tuning of the memory subsystem and cache could reduce that even more. Clock the VRAM higher as well and I think high-end cards with 18GB and a 128-bit interface makes plenty of sense.

But I don't know that anyone is seriously talking about 3GB GDDR6/GDDR7 chips right now.

 

[Eximo]

Wouldn't having two sets of chips raise the BOM costs as well, and complicate the board design somewhat? Doesn't seem like a wise choice for a lower end product.

 

[JarredWaltonGPU]

Wouldn't having two sets of chips raise the BOM costs as well, and complicate the board design somewhat? Doesn't seem like a wise choice for a lower end product.

Only a bit... we used to have graphics cards with memory on both sides all the time about a decade ago. It was only with larger 1GB and now 2GB chips that AMD and Nvidia have been able to stop doing that — and save the clamshell designs for the professional GPUs.

Best guess, having double sided VRAM probably costs maybe $10 extra at most, but the cost of the extra memory is going to be another $20–$30 for 8GB of GDDR6.

 

[baboma]

>...[RX 7600 XT] basically ties the RX 6700 XT.

>Current prices put the 7600 XT about $60 higher than the RX 7600. That's not a terrible price to pay, but neither does it make for an amazing deal.

It's not proper to compare 7600XT's MSRP to 6700XT's clearance price and 7600's street price and make a value judgment on that. Whatever tech merits your benchmarks show, their credibility is diminished by the slanted price valuations.

In effect, your value assessment is valid today--on day one or week one of release, but not when the card's pricing inevitably drops to probably ~$300, or when the 6700XT is no longer available, as I doubt you will go back and revise your summation to take the altered pricing/value into account some days or weeks from now. That makes your review pertinent to the regular readers here, but not to the wider audience at large.

The effect is more pronounced given that the 7600XT is a value product. Value buyers don't rush out to buy products on day 1 of release, and they don't buy at MSRP if at all possible. The 7600 quickly dropped from $300 to $250 after release, and that's what value buyers would expect from 7600XT, to drop to at least $300 if not lower.

I understand that as a reviewer, you can only make value assessment based on present pricing, and not on guesses of future prices even if reasonable. But the above factors need to be taken into account as well. My suggestion is that, rather than predicating value on price so much, that you leave the "value" portion more open-ended, and simply present the tech merits as the main valuation. Readers don't need to be led by the nose toward which card to buy; they can figure the pricing part out by themselves.

 

[thisisaname]

I wonder how warm it is going to get 190w in a single slot could get a little toasty?

 

[suryasans]

AMD's biggest mistake in Navi based GPU generation is AMD Navi 33 GPU. It does not get the same smaller process fab node like its pricey siblings.