AMD Confirms Radeon RX 7900 Series Clocks, Direct Competition with RTX 4080

[btarunr]

AMD in its technical presentation confirmed the reference clock speeds of the Radeon RX 7900 XTX and RX 7900 XT RDNA3 graphics cards. The company also made its first reference to a GeForce RTX 40-series "Ada" product, the RTX 4080 (16 GB), which is going to launch later today. The RX 7900 XTX maxes out the "Navi 31" silicon, featuring all 96 RDNA3 compute units or 6,144 stream processors; while the RX 7900 XT is configured with 84 compute units, or 5,376 stream processors. The two cards also differ with memory configuration. While the RX 7900 XTX gets 24 GB of 20 Gbps GDDR6 across a 384-bit memory interface (960 GB/s); the RX 7900 XT gets 20 GB of 20 Gbps GDDR6 across 320-bit (800 GB/s).

The RX 7900 XTX comes with a Game Clocks frequency of 2300 MHz, and 2500 MHz boost clocks, whereas the RX 7900 XT comes with 2000 MHz Game Clocks, and 2400 MHz boost clocks. The Game Clocks frequency is more relevant between the two. AMD achieves 20 GB memory on the RX 7900 XT by using ten 16 Gbit GDDR6 memory chips across a 320-bit wide memory bus created by disabling one of the six 64-bit MCDs, which also subtracts 16 MB from the GPU's 96 MB Infinity Cache memory, leaving the RX 7900 XT with 80 MB of it. The slide describing the specs of the two cards compares them to the GeForce RTX 4080, which is what the two could compete more against, especially given their pricing. The RX 7900 XTX is 16% cheaper than the RTX 4080, and the RX 7900 XT is 25% cheaper.



View at TechPowerUp Main Site

 

[Deleted member 185088]

It will all depend on the real pricing and availability, apparently the 4080 in addition to be overpriced won't be available.

 

[wolf]

Need to see the actual 3rd party verified performance of all 3 products, 7900XT/XTX and RTX4080. With performance being more nuanced than ever these days with things like RT, accelerated upscaling, differing performance delta's depending on resolution etc, I don't think this will be as simple as "buy X and avoid Y", people will need to carefully consider multiple things, like features, target res, whether you loathe the company, power needs etc. Hopefully they've been doing it all along, but circa Q4 2022 it feels like there are more considerations and possible avenues for let-down than ever.

 

[Richards]

The ray tracing performance looks horrid .. the 4080 will land the knockout punch

 

[Hxx]

The ray tracing performance looks horrid .. the 4080 will land the knockout punch

I don’t think anyone expects better RT out of AMD when compared to nvidia . However in terms of raster performance, that’s where AMD will likely win so it will come down to what matters to the buyer

 

[Toss]

WHO CARES ABOUT RAY TRACING. ARE OU OUT YO MIND?

 

[WhoDecidedThat]

The ray tracing performance looks horrid .. the 4080 will land the knockout punch

Which website has the benchmarks for 7900XTX?

 

[the54thvoid]

Personally quite excited to see these cards. The lower RT performance isn't an issue to me if it's Ampere level. The price is important.

 

[Towelie00]

WHO CARES ABOUT RAY TRACING. ARE OU OUT YO MIND?

reality is who care about RT below 60fps at 4K like 4090

 

[spnidel]

The ray tracing performance looks horrid .. the 4080 will land the knockout punch

if only ray tracing made games good

 

[Valantar]

I'm a bit surprised at the drop in clocks and CUs for the XT v. the XTX, to be honest - the drop in power seems a bit small compared to that difference, with 10% less CUs and 10% lower clocks for just ~15% less power. Makes me wonder whether the XT will either often boost higher than spec, or if it'll just have a ton of OC headroom - or if it's explicitly configured to allow essentially any silicon to pass binning for that SKU.

Either way, very much looking forward to benchmarks of these!

Which website has the benchmarks for 7900XTX?

None, since it doesn't launch till December 3rd?

 

[ratirt]

RT is cool but we are not there yet. Showcasing and marketing can use RT angle and they do but buying a card solely for RT purposes and low res with DLSS or whatnot just to play it? No thanks.
Everything will look good if the price is right. I'm curious how those cards will stack against NVidia. If the 7900XTX can reach 4090 and can beat 4080 comfortable that is a huge win for consumers with the price which could have been lower though but it is not bad. For NV pricing? Well, I'm kinda losing words how to call this since it's insanely crazy?

 

[pat-roner]

Couldn't care less about RT

 

[btk2k2]

I'm a bit surprised at the drop in clocks and CUs for the XT v. the XTX, to be honest - the drop in power seems a bit small compared to that difference, with 10% less CUs and 10% lower clocks for just ~15% less power. Makes me wonder whether the XT will either often boost higher than spec, or if it'll just have a ton of OC headroom - or if it's explicitly configured to allow essentially any silicon to pass binning for that SKU.

I expect the yields (die itself, parametric and MCD bonding) are actually very good so AMD want to sell as many parts as the XTX as possible hence the pricing difference. This lowers demand for the XT and allows AMD to avoid using perfectly good dies in the XT part to meet demand.

I also expect the 7800XT to be a reasonable amount faster than the 6950XT (+15% to 20%) but in a much lower BOM so I would not be surprised if AMD have done what you said by making it so the XT can be made by anything that is either partially defective, does not meet the performance bin for the XTX or has a failed MCD bond.

If true though I suspect it will mean a lot of variance in how the XT overclocks because you could have a die that was defective (missing shaders) but clocks great, a die that failed the bin so clocks poorly or a die that is actually an XTX die but one of the MCD bonds didn't work properly so it needs to be sold as an XT. This leads me to think the driver power limits will be pretty restricted for the XT version because otherwise I can imagine people saving the $100, buying the XT and then overclocking the snot out of it (a bit like the 7950 vs 7970).

 

[AusWolf]

Where's all the people who cried at the launch event that the 7900 XTX doesn't match up against the 1.6x more expensive 4090?

 

[Valantar]

I expect the yields (die itself, parametric and MCD bonding) are actually very good so AMD want to sell as many parts as the XTX as possible hence the pricing difference. This lowers demand for the XT and allows AMD to avoid using perfectly good dies in the XT part to meet demand.

I also expect the 7800XT to be a reasonable amount faster than the 6950XT (+15% to 20%) but in a much lower BOM so I would not be surprised if AMD have done what you said by making it so the XT can be made by anything that is either partially defective, does not meet the performance bin for the XTX or has a failed MCD bond.

If true though I suspect it will mean a lot of variance in how the XT overclocks because you could have a die that was defective (missing shaders) but clocks great, a die that failed the bin so clocks poorly or a die that is actually an XTX die but one of the MCD bonds didn't work properly so it needs to be sold as an XT. This leads me to think the driver power limits will be pretty restricted for the XT version because otherwise I can imagine people saving the $100, buying the XT and then overclocking the snot out of it (a bit like the 7950 vs 7970).

Yeah, that's pretty much my thinking exactly. While yields/defect rates for TSMC N5 aren't public in the same way they were for N7, we still know that they are good - good on a level where they're churning out massive amounts of large-ish dice with very high yields. Combine that with even AMD's biggest GPU die now being <400mm² and, like you say, most likely the vast majority of dice will qualify for the XTX SKU. We saw the exact same thing with Navi 21 on N7, where the 6900 XT was always (relatively) plentiful, while 6800 XT and 6800 supplies were nonexistent at times, and even at the best of times very limited, simply because AMD would rather sell a fully functioning die as a 6900 XT than cut it down to sell as a cheaper SKU.

Of course they're not in the same supply-constrained market today, so they're going to need to be a bit more flexible and more willing to "unnecessarily" sell parts as a lower bin than they actually qualify for - this has been the norm in chipmaking forever, after all. But I still expect their first push to be plenty of XTXes, and notably fewer XTs. This also (assuming the XT PCB is also used for the 7800 XT) makes the XTX having its own PCB more understandable - it's likely supposed to have enough volume to recoup its own costs, while the XT is designed to be an in-between SKU with more cost optimization. Which is kind of crazy for a $900 GPU, but it's not like those cost optimizations are bad, it just looks slightly less over-the-top.

It will definitely be interesting to see what the power limits for the XT will be like - AMD has had some pretty strict power limits for lower end SKUs lately, like the RX 6600, but nothing like that for a high end, high power SKU. It also raises the question of what premium AIB models of the XT will be like, as AMD is making it pretty clear that there'll be heavily OC'd partner XTXes. That might also be part of the pricing strategy here - with a mere 10% difference, ultra-premium 3GHz XTs don't make as much sense, as they'd cost more than a base XTX - so the upsell to an equally ultra-premium XTX would be relatively easy. And AMD makes money on selling chips after all, not whole GPUs, so they'd always want to sell the more premium SKU.

Also definitely looking forward to seeing what the 7800 XT will be - a further cut down Navi 31? If Navi 32 has the rumored CU count, they'd need another in-between SKU (the poor competitive performance of the 6700 XT demonstrated how they can't leave gaps that big in their lineup), but with die sizes being this moderate and GPUs being relatively affordable and easy to tape out compared to most chips (being massive arrays of identical hardware helps!) I could see AMD launching more Navi 3X dice than 2X from that fact alone.

 

[AusWolf]

Yeah, that's pretty much my thinking exactly. While yields/defect rates for TSMC N5 aren't public in the same way they were for N7, we still know that they are good - good on a level where they're churning out massive amounts of large-ish dice with very high yields. Combine that with even AMD's biggest GPU die now being <400mm² and, like you say, most likely the vast majority of dice will qualify for the XTX SKU. We saw the exact same thing with Navi 21 on N7, where the 6900 XT was always (relatively) plentiful, while 6800 XT and 6800 supplies were nonexistent at times, and even at the best of times very limited, simply because AMD would rather sell a fully functioning die as a 6900 XT than cut it down to sell as a cheaper SKU.

That alone makes AMD's attitude to sales a lot more sympathetic than Nvidia's in my opinion.

• AMD tries to sell fully functional dies as much as possible, while also having cheaper versions at the same time - they play with open hands right from launch.
• Nvidia sells partially disabled dies as their high-end, and reserves fully functional ones to carve out an even higher-end segment of the market later - they artificially boost the hype train to maintain sales.

 

[Valantar]

That alone makes AMD's attitude to sales a lot more sympathetic than Nvidia's in my opinion.

• AMD tries to sell fully functional dies as much as possible, while also having cheaper versions at the same time - they play with open hands right from launch.
• Nvidia sells partially disabled dies as their high-end, and reserves fully functional ones to carve out an even higher-end segment of the market later - they artificially boost the hype train to maintain sales.

That's true - in terms of marketing that is a pretty shitty tactic. For me it also raises several questions about production though. Is this (partially) due to Nvidia's ~4x sales numbers causing ~4x more defective dice, making cut-down SKUs more viable? Is there something about Nvidia's designs (or their fabs/nodes) that makes them more error prone? Or is it purely cash-grab thinking, where at the start anything not cut down gets sold as a 2-3x priced not-Quadro or other accelerator, and then after that market is somewhat saturated they launch a consumer Ti/Super for a new PR boost like you say. There's also the question of how much of this is a remnant of thinking from older fab nodes where defect rates were much higher, back in the 2000s when you had 20 SKUs differentiated by $20 in MSRP because you wanted to sell every possible bin of a die to someone.

 

[nguyen]

That alone makes AMD's attitude to sales a lot more sympathetic than Nvidia's in my opinion.

• AMD tries to sell fully functional dies as much as possible, while also having cheaper versions at the same time - they play with open hands right from launch.
• Nvidia sells partially disabled dies as their high-end, and reserves fully functional ones to carve out an even higher-end segment of the market later - they artificially boost the hype train to maintain sales.

Yes yes, very sympathetic when AMD lock 6900XT voltage/freq in order to sell higher SKUs: 6900XT LC, 6900XTXH, 6950XT

 

[the54thvoid]

Yes yes, very sympathetic when AMD lock 6900XT voltage/freq in order to sell higher SKUs: 6900XT LC, 6900XTXH, 6950XT

Locking frequency, especially for binning purposes is different from what is being discussed. The chip remains as is, the individual differences in process node lead to different electrical tolerances, hence the binning.

Selling dies with lower componentry due to yield issues is another matter, albeit a legitimate one.

 

[nguyen]

Locking frequency, especially for binning purposes is different from what is being discussed. The chip remains as is, the individual differences in process node lead to different electrical tolerances, hence the binning.

Just scummy move IMO, had that problem with 2080Ti non-OC chip that was artificially locked to 280W TDP max, where as 2080Ti OC chip can have unlimited TDP (with XOC bios).

Nothing worse than selling highest end SKU, only to artificially limiting it in order to sell higher SKU later down the line

 

[AusWolf]

Yes yes, very sympathetic when AMD lock 6900XT voltage/freq in order to sell higher SKUs: 6900XT LC, 6900XTXH, 6950XT

Still better than selling products with only partially defective dies for a year just to mark up the good ones later.

That's true - in terms of marketing that is a pretty shitty tactic. For me it also raises several questions about production though. Is this (partially) due to Nvidia's ~4x sales numbers causing ~4x more defective dice, making cut-down SKUs more viable? Is there something about Nvidia's designs (or their fabs/nodes) that makes them more error prone? Or is it purely cash-grab thinking, where at the start anything not cut down gets sold as a 2-3x priced not-Quadro or other accelerator, and then after that market is somewhat saturated they launch a consumer Ti/Super for a new PR boost like you say. There's also the question of how much of this is a remnant of thinking from older fab nodes where defect rates were much higher, back in the 2000s when you had 20 SKUs differentiated by $20 in MSRP because you wanted to sell every possible bin of a die to someone.

I'm going with the "cash grab" option purely because you never see any product based on a fully enabled die at the launch of a series anymore. No fab works with 100% defect rate.

 

[nguyen]

Still better than selling products with only partially defective dies for a year just to mark up the good ones later.

Obviously you have never had the feeling of having the best GPU, only for it to be artificially limited .

Couldn't careless about my 3090 or 4090 if they were being 50% working die

 

[AusWolf]

Obviously you have never had the feeling of having the best GPU, only for it to be artificially limited .

Ah, so your 4090 isn't artificially limited by having 2048 of its GA102's shaders disabled?

I don't care about having the best, but I do care about having the best in the price and performance range I consider sensible for my needs.

Couldn't careless about my 3090 or 4090 if they were being 50% working die

Then why do you care about a locked voltage/frequency curve?

 

[Vayra86]

The ray tracing performance looks horrid .. the 4080 will land the knockout punch

Care level is sub-zero here for RT performance. Its tech in its infancy, who cares if you get a 4 year old or a 3 year old? They're both kids, they'll both grow over time, every advantage here is of little relevance.

IMHO the real cutting edge is chased by AMD right now, doing chiplet GPUs that are the only possible way of continuously moving GPU performance forward without creating massive monolithic dies that need upwards of 400W to speed ahead.