AMD 7600 reviews

[SteveGrabowski]

I'm not sure what the typical retailer/distributor cut is, but I would say it's closer to 30% than 10%. At $200 MSRP that would mean they would be selling it to the distributor for about $150.

For gaming consoles it's like 8%. Would be very surprised to see it 30% for gpus.

 

[Mopetar]

I doubt N6 is $11k/wafer. N7 was generally reported as $10k/wafer when that was bleeding edge and almost every rumor has been that TSMC was offering rebates to customers to switch over to N6 as it had fewer steps and TSMC could get much greater throughput.

Now that N5 has been out for a while and N3 is the new bleeding edge that Apple is manufacturing the next iPhone SoCs on right now, the previous nodes are going to start being discounted, particularly when the industry is hitting the post-pandemic dip and demand is a lot lower. Once TSMC recoups their initial investment costs, they can drop wafer prices considerably, which they might have to do in order to keep the fabs operating.

But even if I were to grant that you are correct and N6 actually does cost $11K per wafer, that doesn't really change anything other than AMD perhaps only making $5 instead of $25. Their margins have decreased, but they still wouldn't lose money. They have enough other divisions that are profitable enough to weather hard times here.

People can argue over margins all they want, but at the end of the day the pure profit is what matters. Would you rather have $2 billion at 75% margins, or $2.1 billion at 15% margins?

 

[jpiniero]

I doubt N6 is $11k/wafer. N7 was generally reported as $10k/wafer when that was bleeding edge and almost every rumor has been that TSMC was offering rebates to customers to switch over to N6 as it had fewer steps and TSMC could get much greater throughput.

That was before the price hikes. There's a chance that TSMC slashes prices on N6 at some point but it sounds like that hasn't happened yet. And won't until sometime in the second half , and only if demand doesn't recover.

 

[Hans Gruber]

The RX480 debuted for $199 with 8GB of vram back in June 2016. Memory prices are much lower today than they were in 2016. It's depressing to see how little the GPU market has advanced in 7 years at the entry/mid level of GPU's.

 

[PJVol]

That was before the price hikes. There's a chance that TSMC slashes prices on N6 at some point but it sounds like that hasn't happened yet. And won't until sometime in the second half , and only if demand doesn't recover.

I believe that the N6 wafer costs AMD around 7k at most, assuming N6 is not more expensive than N7.
One should bear in mind that AMD or Apple is not the same as Nvidia or Intel when it comes to business relationship with the TSMC.

 

[AtenRa]

As a reminder, N33 doesn't use the same CUs as N31. N31 enjoys 50% more vector registers (Vector GPR) than RDNA2, but N33 retains the same number of vector registers as RDNA2.

Without those registers, N33 has the same number of rays in flight as N23, which means raytracing performance should be similar on a per-CU basis. Utilization of the doubling of FP32 units takes a hit too because more optimization is needed to ensure the data is in the register.

I didnt know that, thx for the info.

Any links for that N33 CU architecture/specs ??

 

[VirtualLarry]

It wasn't all that long ago that AMD was selling cards with similarly sized Polaris dies for $229 on down. That was when 8 GB of VRAM required twice as many modules since you couldn't get a 2 GB chip. The cards also had a higher TDP so would require a stronger cooling solution and components to drive more power.

Forgot to mention, Polaris was 256-bit, and RX 6600 is 128-bit, so more expensive versus simpler P/c/b. Also, Polaris cards were usually hella over-built, whereas I find most RX 6600 cards are barely fit for purpose.

 

[Mopetar]

The entry level is basically dead. APUs are good enough to replace the $150 entry-level GPU people used to buy to play DotA or Rocket League or other games like that.

We're in this weird place right now where a lot of people are still using 1080p even though it's almost obsolete in terms of what else is available. Even 4K displays are ridiculously cheap these days if you're not trying to buy some crazy HDR ultra-high refresh rate gaming monitor.

Even the 7600 as unimpressive as it is will give you at least 60 FPS at 1080p in modern titles at the highest quality settings. Of course these days most people consider 60 FPS the bare minimum acceptable level of performance. Once upon a time that was something to aspire to, even if you weren't gaming at 1080p.

 

[Saylick]

I didnt know that, thx for the info.

Any links for that N33 CU architecture/specs ??

I'll have to dig it up, but Kepler and a few others had screenshots of code that showed the number of registers, etc.

 

[PJVol]

Any links for that N33 CU architecture/specs ??

We've been discussing this in the rdna3 spec thread. And no, AMD hasn't published N33 block scheme yet.

 

[Hans Gruber]

Is there any data on the cost of the bus. Eg. 128bit, 192bit, 256bit. Are they trying to save money using 128bit bus on the 7600 or are they intentionally gimping performance?

 

[PJVol]

Is there any data on the cost of the bus. Eg. 128bit, 192bit, 256bit. Are they trying to save money using 128bit bus on the 7600 or are they intentionally gimping performance?

If you mean cost in terms of area or number of layers, i'd guess it's proportional to the bus width, apart from mask cost and other design related extras.
And there might be other reasons I think, which ultimately aims at effective utilizing current node's PPAC

 

[Hans Gruber]

If you mean cost in terms of area or number of layers, i'd guess close to proportional to the bus width, apart from mask cost and other design related extras.

That's not what I am talking about. My 2012 AMD 7950 with 3GB of vram had a 384-bit bus. Others have pointed the RX480 had a 256bit bus. That is the pipe, aka. width of a bus to transmit data. The 4060ti has a bus width of 128bit which is pathetic for a midgrade Ti card. The RX 7600 has a 128bit bus. The 3060 had a 192bit bus.

So are both AMD intentionally gimping cards by using a 128bit bus with the RX7600 and 4060ti cards? I should add the 6600 an 6600xt were 128bit bus cards. Nvidia has dropped down from 192bit to 128bit like AMD.

I remember a few years ago people worried about cards dropping from a 256bit bus to a 192bit bus. Now they are going all the way down to 128bit.

 

[Aapje]

Is there any data on the cost of the bus. Eg. 128bit, 192bit, 256bit. Are they trying to save money using 128bit bus on the 7600 or are they intentionally gimping performance?

It's 128 bit, of course. Just like N23.

It's a 1080p card, don't cha know.

 

[blckgrffn]

That's not what I am talking about. My 2012 AMD 7950 with 3GB of vram had a 384-bit bus. Others have pointed the RX480 had a 256bit bus. That is the pipe, aka. width of a bus to transmit data. The 4060ti has a bus width of 128bit which is pathetic for a midgrade Ti card. The RX 7600 has a 128bit bus. The 3060 had a 192bit bus.

So are both AMD intentionally gimping cards by using a 128bit bus with the RX7600 and 4060ti cards? I should add the 6600 an 6600xt were 128bit bus cards. Nvidia has dropped down from 192bit to 128bit like AMD.

I remember a few years ago people worried about cards dropping from a 256bit bus to a 192bit bus. Now they are going all the way down to 128bit.

The trade off being the larger onboard cache that is more power efficient and yeah, I am sure part of the reason they add it is to bring the cost of the card down. It doesn't seem like a step they would take in order to spend more.

That and the scaling of bus widths was hitting some natural walls.

 

[GodisanAtheist]

As a reminder, N33 doesn't use the same CUs as N31. N31 enjoys 50% more vector registers (Vector GPR) than RDNA2, but N33 retains the same number of vector registers as RDNA2.

Without those registers, N33 has the same number of rays in flight as N23, which means raytracing performance should be similar on a per-CU basis. Utilization of the doubling of FP32 units takes a hit too because more optimization is needed to ensure the data is in the register.

- Weird that AMD would even waste time and effort on this die at all. Could have saved the 2 billion extra transistors and done a straight shrink of N23 on 6nm, saved a bit more space and sold it for less.

Frankly AMD should have shrunk N22 down to 6nm as well and starting pumping those out on the cheap as well instead of playing the old "where the hell is N32" game with us.

 

[Saylick]

- Weird that AMD would even waste time and effort on this die at all. Could have saved the 2 billion extra transistors and done a straight shrink of N23 on 6nm, saved a bit more space and sold it for less.

Frankly AMD should have shrunk N22 down to 6nm as well and starting pumping those out on the cheap as well instead of playing the old "where the hell is N32" game with us.

Yeah, it does seem like a waste of time and effort in hindsight.

It does highlight the difference in approach between AMD and Nvidia though. I get the impression that Nvidia comes up with the software first, then works backwards to figure out what kind of hardware would accelerate that workload. AMD seems to do the opposite: come up with the hardware then try to figure out how to leverage it using software after it's launched.

 

[Ranulf]

Sales don't seem too bad at NE for the first day (or is it day 2 now?). 4 of 12 models sold out. Most staying at MSRP or up to $290 for custom cards. 3fan Asrock cards are $289 too. Gigabyte has a 3fan card at msrp.

 

[VirtualLarry]

I got the impression that those hadn't had the stock come in yet. You really think that both Asus cards would sell out, what with their reputation problems lately?'

I mean, do you think that there is all of this pent-up demand for the RX 7600?

The $180 triple-fan Gigabyte RX 6600 isn't even sold out yet.

OK, I did buy a triple-fan XFX model, for MSRP.

Can't wait for the rebates and promo codes making these $250 two weeks from now.

 

[Ranulf]

Hmm, good point...maybe not in stock yet for the first time.

I see most of the 6600/x50 XT cards are down to $240 or so. A $15-20 haircut from just a few days ago. That MSI 6750xt at $329 is probably the best bet now.

$199 6600 cards aren't too bad but that isn't much of a price cut either.

 

[PJVol]

Sales don't seem too bad at NE for the first day

Hmm... who'd have thought, so many YT banners with distorted faces and outraged headlines...

 

[GodisanAtheist]

It's time for our regular game of "Is it selling well or is there low supply" AMD edition.

 

[Hans Gruber]

At some point forum members here need to realize GPU's are not selling. I am not just talking about the 4060ti and the RX7600. GPU's from top to bottom are gathering dust on store and warehouse shelves.

This idea about high margins is pie in the sky for both Nvidia and AMD.

 

[coercitiv]

Looks like some people will need an adapter:

Some 6+2pin power cables may not fully fit into Radeon RX 7600 reference card - VideoCardz.com

AMD Radeon RX 7600’s trouble with some power cables Reviewers report that some cables may not fit AMD’s reference card. Radeon RX 7600 6+2pin power cable issue, Source: TechPowerUP The issue affects only some power cables, and it is not hard to notice. As it is reported, some power cables may...

videocardz.com

 

[mmaenpaa]

For gaming consoles it's like 8%. Would be very surprised to see it 30% for gpus.

Around 10% is "normal" in this business, I agree.