AMD Radeon RX 7600 Final Specs and Power Figures Leaked, Uses 6 nm

[btarunr]

Here are the final specifications of the Radeon RX 7600 RDNA3 graphics card, bound for launch later this month. The specs list springs up some surprises. To begin with, while the GPU at the heart of the RX 7600 is based on the latest RDNA3 graphics architecture, it is built on the older 6 nm (DUV) silicon fabrication process—the same one on which the previous "Navi 24" was based. The silicon has a transistor count of 13.3 billion, about 2 billion more than the 7 nm "Navi 23" Powering the RX 6600 series, but a die-size of 204 mm². The GPU has a PCI-Express 4.0 x8 host interface, and a 128-bit GDDR6 memory interface. As the RX 7600, it has a TBP (total board power) value of 165 W, which is over 30 W more than the RX 6600.

At this point, it's not known whether the RX 7600 maxes out the silicon it is based on. It gets 32 RDNA3 compute units (CU), which work out to 2,048 stream processors (with the same dual-issue instruction rate feature as the RX 7900 series); 32 Ray Accelerators, and 64 AI Accelerators. The GPU has 128 TMUs, and 64 ROPs. The GPU has 32 MB of second-generation Infinity Cache memory. The 8 GB of GDDR6 memory ticks at 18 Gbps, which over the 128-bit memory bus works out to 288 GB/s of memory bandwidth. AMD claims that when coupled with the on-die cache, the "effective bandwidth" is 476.9 GB/s. NVIDIA is putting out similar "effective" figures for its RTX 4060 series, so this could become a norm. The RX 7600 comes with a game frequency of 2250 MHz, and 2625 MHz boost. AMD is making 550 W as its PSU recommendation, compared to the 450 W it did for the RX 6600. The company considers the RX 7600 to be the logical successor of the RX 6600 (and neither the RX 6600 XT nor the RX 6650 XT).



View at TechPowerUp Main Site | Source

 

[ixi]

200euro fine for the product, 300e not so much. At least for my view.

 

[Verpal]

In a honest world this would be consider as a 6650XT replacement, considering current price level, but 6000 series will eventually sold out so I guess that's fair.

Anyway, please for the love of RTG don't price this at $300, get crushed by 4060 on day 1 review, then reduce price bit by bit in following months.
Just release it at $250 and finally compete, please?

 

[docnorth]

Only looooww pricing can save this card. Otherwise the mediocre (at best) 4060 will crush it, not to mention last generation AMD cards.

 

[Ruru]

8GB makes sense as this is a lower mid-end card with a 128-bit bus.

 

[kondamin]

Why bother releasing that?

 

[tugrul_LordOfDrinks]

Maybe higher memory not needed with that gpu power? 1080p games need 16GB VRAM?

 

[KV2DERP]

Maybe higher memory not needed with that gpu power? 1080p games need 16GB VRAM?

Seems like people are paranoid over TLOU benchmark that indicates "8GB cards are dead" despite the extreme low 1% case only appears on ultra settings (I'm talking bout 3070 ofc). On high settings, 3070 could play that game relatively well.

Yeah, more vram is better but again it will drive up the cost. Idk what the price of this card but if it cost > $300, 6700XT is a sweeter deal than this.

 

[Broken Processor]

Needing to upgrade your GPU this generation really is a war of attrition.

 

[Metroid]

6nm, charging 300 usd and while is slower than 4060, it uses a lot more power than it, meaning failed, even a 4060 has failed at 299 usd, a 199 usd would be the right price for both. I remember when an RX 580 was sold for 199 usd when it was released. Nowadays AMD is charging 600 usd for the x8xx, disgraceful. Nvidia went from 699 usd to a whopping 1299 usd for the x8xx.

 

[Chaitanya]

200euro fine for the product, 300e not so much. At least for my view.

Just give it 6 months or so, prices will come down to reasonable level.

 

[tfdsaf]

RTX 4060 is DOA with $300 and only 8GB of vram, if this launches at $250 it is going to be decent value, nothing great, but it will at least be a good alternative for the 4060 and probably be able to sell more, but knowing AMD this is going to end up 5% faster in raster than the RTX 4060 and cost the same and AMD are going to rely on a small 5% performance lead in raster to make up for all the other deficiencies and Nvidia's amazing PR propaganda that is able to brainwash people and sell them overpriced turds!

 

[pavle]

All these "up to"'s, it's getting ridiculous. Where is a proper product spec? The performance too will be "up to not worthy of buying" I'm sure.

 

[tabascosauz]

Okay, this shit needs to stop. How is this is any different than Nvidia's spec sheet using large L2 to artificially inflate the 4060's "bandwidth"?

Caches weren't invented yesterday; is the 7800X3D also supposed to use its V-cache to pump up its "effective" DDR5 bandwidth?

 

[tvshacker]

Weird times...
New arch + (slightly) smaller node = ~30% higher board power
One can only hope perf/W scales linearly as well

 

[BoboOOZ]

In a honest world this would be consider as a 6650XT replacement, considering current price level, but 6000 series will eventually sold out so I guess that's fair.

Anyway, please for the love of RTG don't price this at $300, get crushed by 4060 on day 1 review, then reduce price bit by bit in following months.
Just release it at $250 and finally compete, please?

It's not a 6600XT replacement, it has the same number of CUs. CU's typically go up a bit of each generation, due to the die shrinks. 7600 is an honest name and it's the first honest name for AMD this generation. Let's hope it will also be a honest price.

 

[Icon Charlie]

Effective bandwidth = market speak for blowing pixie dust up People's @$$'es. I might pull the trigger in buying a used 6800XT as they are getting down to the $380.00 range. But to be honest about even doing that feels like a luxury to me as my RX 5700 is breezing along rather cool like too

 

[TumbleGeorge]

Weird times...
New arch + (slightly) smaller node = ~30% higher board power
One can only hope perf/W scales linearly as well

If you read declared teraflops - yes with a lot. If you make comparison of FPS... will see when reviewed, but I also lose hope.

 

[Daven]

It's not a 6600XT replacement, it has the same number of CUs. CU's typically go up a bit of each generation, due to the die shrinks. 7600 is an honest name and it's the first honest name for AMD this generation. Let's hope it will also be a honest price.

Nvidia went down in compute cores on the 4060 and 4060TI compared to the 3060 and 3060TI. Compute cores stayed the same between the 4070 and 3070.

 

[tvshacker]

If you read declared teraflops - yes with a lot. If you make comparison of FPS... will see when reviewed, but I also lose hope.

By that metric we're on 6800XT levels of performance.
I doubt it will happen on the fps charts but I'm very willing to be wrong.

 

[BoboOOZ]

Nvidia went down in compute cores on the 4060 and 4060TI compared to the 3060 and 3060TI. Compute cores stayed the same between the 4070 and 3070.

Yes, and those are fake names also. The 4060 is just a 4050 in fact, and the 4070 should be called a 4060TI. But It looks like the fake naming has backfired on both Nvidia and AMD.

By that metric we're on 6800XT levels of performance.
I doubt it will happen on the fps charts but I'm very willing to be wrong.

The TFlops are theoretically huge because former integer processing units can now do double purpose, hence double the floating point performance numbers. But in practice, the gain is somewhere between 15 to 25% at most, as seen on Ampere and on the 7900s.

 

[Colddecked]

It's not a 6600XT replacement, it has the same number of CUs. CU's typically go up a bit of each generation, due to the die shrinks. 7600 is an honest name and it's the first honest name for AMD this generation. Let's hope it will also be a honest price.

I've lost all hope AMD will price this correctly. They are more than happy letting stock of 7000series build and empty 6000 stock. Itll be 300 until Oct I feel.

 

[BoboOOZ]

I've lost all hope AMD will price this correctly. They are more than happy letting stock of 7000series build and empty 6000 stock. Itll be 300 until Oct I feel.

They still have time to price this right, but it's true that if you judge based on recent history there's not much reason for optimism. Still, AMD wants to sell these GPUs, they can't just allocate more capacity for the professional market as Nvidia does, so there's still a sliver of hope.

 

[john_]

I think after getting their first RX 7000 cards in their labs and running some first tests on them, AMD's engineers where probably asking themselves
"Is this what we where building for 2 years? Is there any real difference over RDNA2?"

 

[ToTTenTranz]

So from those bandwidth numbers it looks like Navi 33 is using the same Infinity Cache arrangement as RDNA2 cards.

Navi 31 reduced the amount of L3 cache while doubling the bandwidth per cell, which is why they increased effective bandwidth contribution from the L3 despite lowering the cache from 128 to 96MB.

In the case of Navi 33, the only bandwidth increase seems to be from chancing GDDR6 14Gbps to 18Gbps. A difference of 4Gbps does 64GB/s at 128bit, which is the bandwidth difference between Navi 33 and Navi 23 according to that table.

200euro fine for the product, 300e not so much. At least for my view.

This card is dead in the water if AMD tries to charge anywhere near 300€. Even if they try to charge something like $280 people will just prefer to pay more for the 4060 as no reviewer or influencer is ever going to advise its purchase.

This needed to be 230€ to get major recommendations and it would set the world on fire at 200€. Which is probably where it will end up being priced at anyway, by the end of the year.
It all depends on whether AMD is willing to price it right or they're willing to risk massive pushbacks and disappointment from consumers.

The TFlops are theoretically huge because former integer processing units can now do double purpose, hence double the floating point performance numbers. But in practice, the gain is somewhere between 15 to 25% at most, as seen on Ampere and on the 7900s.

IIRC the 7900s have zero gain from the dual-pumped FP32 units because the compiler doesn't really support the functionality yet.