AMD 7600 reviews

[Aapje]

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

Indeed. Both the stores and AIBs are not getting big margins. It's Nvidia and AMD that gobble up most of the profit.

 

[Joe NYC]

AMD isn't losing any money even at that price. It's just that no one else is going to be making (much) money either.

The Navi 23 die itself costs about $25 to manufacture, add another $25 for the memory and it's about $50 in total for the silicon.

Even if AMD passes that along for $75 that leaves an additional $125 for the PCB and cooling solution, neither of which are significant for Navi 23. Suppose the AIB manufacturers spend $75 on that and sell the product to a retailer for $175 (again taking $25 for themselves over cost) That means the retailer can offer it at $200 to make their own $25 profit on the sale.

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.

These low end cards are not expensive to manufacture. The higher prices are everyone trying to cling to the insane prices they could charge during the mining boom. Yeah it sucks to go back to earning only $25 for something you could easily get $50 if not $100 for only a few years ago, but now that the miners have stopped buying, there's no one left who's willing to pay double MSRP so it's back to thin margins for everyone involved.

I think you forgot the cost of the entire design team, hardware and software.

 

[Joe NYC]

Can you source that claim?

The gaming division of AMD was profitable only due to very healthy console sales last couple of quarters. dGPU portion was most certainly under water.

 

[Aapje]

I think you forgot the cost of the entire design team, hardware and software.

It's a lot easier to pay for that if they actually sell a decent number of cards...

 

[Joe NYC]

Because they barely compete with their older stuff and is embarrassing and tarnishing their entire brand? I have to imagine they could do better by offering cheaper prices for OEMs to put them in laptops or All-in-Ones or something.

Yes, we get they aren't going to just abandon a market segment entirely and would put out a new product just for the features set spec sheet, but at this price its bad. There's also a huge disparity between $200 and the $270 MSRP. $225 or $230 would make a big difference in how this is viewed.

$269 is a good price. It allows AMD (maybe) recoup its investment and (maybe) break even, or even earn a few dollars and stay in this segment.

Without this card, NVidia would not even bother offering 4060, and the lowest price for budget gaming would be $399. Which is nearly 50% higher price than 7600 for 15% extra performance.

 

[Joe NYC]

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.

BTW, $229 in 2017 = $286 today, after inflation.

 

[Joe NYC]

It's a lot easier to pay for that if they actually sell a decent number of cards...

The 6600 XT is surely raking in the big bucks for AMD, selling bucketloads at $200.

Oh wait! It is not.

 

[Aapje]

The 6600 XT is surely raking in the big bucks for AMD, selling bucketloads at $200.

Oh wait! It is not.

Of course it is not selling bucketloads at $200, because it is not actually sold for that price.

A 7600 at $230 would be quite attractive, at $250 it is OK, but at $270 it becomes a very hard sell over the much more expensive to produce 6700 (XT).

 

[Aapje]

BTW, $229 in 2017 = $286 today, after inflation.

Except that inflation is much lower in Asia, where these cards are actually made. Also, inflation is a measurement of the average price increase of consumer products. It doesn't mean that every product becomes this much more expensive. Some products have a bigger price increase than inflation and some less.

There are very good reasons to assume that the actual costs of making these cards is relatively inflation-insensitive, especially for the 7600, which is made on an 'old' process node. The costs associated with buying the machines were made before the inflation increase. The costs to manufacture on a node gradually go down as the process gets older. So even if there are some costs increases due to inflation, you'd expect at least part of this to get compensated by the costs going down of N6 over time. Especially since N6 that is used to make the 7600, is a cost-optimized variant of the more expensive N7, that is used for the 6600 cards. And with occupancy of N6 being problematic, you'd expect TSMC to offer relatively good prices on this node. So everything points to the 6700-chip being cheaper to make than the 6600-chip. So even if the assembly of the card is a bit more expensive due to inflation, you'd expect this to be cancelled out in full or in part by cheaper component prices.

You write a lot of things that might seem sensible to you, but that are based on incorrect facts (like the price of the 6600 XT*) or based on a fundamental misunderstanding of how things like inflation actually work and what the manufacturing costs actually are based on (which is very much not the average price an American pays for consumer goods).

* Ironically, if this were actually true, it would actually hurt your argument that the 7600 is well priced.

 

[Joe NYC]

6600 is not selling by bucketloads at $179
6600 XT is not selling bybucketloads at $214
6650 XT is not selling bybucketloads at $219
(my MicroCenter prices)

But somehow, 7600 would sell by bucketloads, if only price was $30 less.

Well, here is a forecast: At some point in its life, 7600 will get a $30 discount, 7600 will still not sell by bucketloads, and your argument will go up in smoke.

 

[Aapje]

I never said that it would sell by bucketloads at $30 less. You are putting words in my mouth. I said that it won't sell at this price and will sell much better at $250 and way better still at $230. If they really want to sell a lot, the price should go to $200 or below.

And I predict that it will get a discount soon enough, exactly because it won't sell at this price.

 

[jpiniero]

I never said that it would sell by bucketloads at $30 less. I said that it won't sell at this price and will sell much better at $250 and way better still at $230. If they really want to sell a lot, the price should go to $200 or below.

And lose money on each sale. But, hey, they will make it up in volume, right?

 

[Aapje]

And lose money on each sale. But, hey, they will make it up in volume, right?

If they are not losing money by selling the 6600 for around $200, then they are certainly not going to lose money by selling the 7600 at $230 or $200.

You and Joe can keep repeating this nonsense that the profit margins are zero at this price, but it makes zero sense.

 

[jpiniero]

If they are not losing money by selling the 6600 for around $200, then they are certainly not going to lose money at $230 or $200.

Who says they aren't losing money selling the 6600 at $200? As I've said multiple times though, it's unclear as to who is taking the hit.

 

[Joe NYC]

Except that inflation is much lower in Asia, where these cards are actually made. Also, inflation is a measurement of the average price increase of consumer products. It doesn't mean that every product becomes this much more expensive. Some products have a bigger price increase than inflation and some less.

The people who designed Navi 33 chips, people who write drivers and other software for it are based in the US (mostly).

AMD operating expenses (mostly US based) were higher last quarter than the cost to fab the products.

There are very good reasons to assume that the actual costs of making these cards is relatively inflation-insensitive, especially for the 7600, which is made on an 'old' process node. The costs associated with buying the machines were made before the inflation increase. The costs to manufacture on a node gradually go down as the process gets older. So even if there are some costs increases due to inflation, you'd expect at least part of this to get compensated by the costs going down of N6 over time. Especially since N6 that is used to make the 7600, is a cost-optimized variant of the more expensive N7, that is used for the 6600 cards. And with occupancy of N6 being problematic, you'd expect TSMC to offer relatively good prices on this node. So everything points to the 6700-chip being cheaper to make than the 6600-chip. So even if the assembly of the card is a bit more expensive due to inflation, you'd expect this to be cancelled out in full or in part by cheaper component prices.

Yes, definitely agree with that. This is the one of the reasons why Navi 33 is selling at $269 while Navi 23 sold for $379.

A lot of people, including me thought $379 for 6600 XT was a decent deal for a decent card.

 

[In2Photos]

You guys do realize that the engineering costs for the drivers and even the hardware are spread across the entire line of cards and not just a single one right? And it's not like they start over with drivers. They use the same drivers as they did for previous generations and just make changes to them. The majority of engineering was already paid for when the 7900 XTX & XT were launched.

 

[Aapje]

Who says they aren't losing money selling the 6600 at $200? As I've said multiple times though, it's unclear as to who is taking the hit.

Let's do the calculation for the 7600 instead, because we are actually talking about what that one should cost. During the mining boom, AMD introduced the 6400 for an MSRP of $159. That must have been a profitable price, or it would have made no sense to choose that MSRP, certainly at that time.

The 7600 has twice the VRAM. So how much does that cost?

The MSRP for the 4 GB 6500 is $199 and for the 8 GB variant it is $219. So we can logically assume that the extra VRAM costs at most $20.

So then the only remaining difference is a 204 mm2 die instead of 107 mm2. Based on leaks about wafer prices, which a chip-expert who knows the actual prices (under NDA) recently said are way too high, we can calculate that a 204 mm2 chip on N6 won't cost more than $30-35. So if a 107 mm2 then costs about half, that, you can assume at most a $20 difference for the chip.

$159 + $20 for the extra VRAM + $20 for the bigger chip = $199. So given that the price of the 6400 logically must have been profitable, that $199 must be profitable.

The 6600 XT should actually be a bit more expensive to make than the 7600, given that it's 237 mm2 on N7. So in theory it can be unprofitable, but the 6600 XT is not all that much more expensive, with just the difference in die mattering, so at most it's going to be a difference of a few dollars when chips around this size cost at most $30-$35. It's just not very credible that the profit margin at MSRP is set so low that a few dollars less profit suddenly results in making a loss.

 

[jpiniero]

So then the only remaining difference is a 204 mm2 die instead of 107 mm2. Based on leaks about wafer prices, which a chip-expert who knows the actual prices (under NDA) recently said are way too high, we can calculate that a 204 mm2 chip on N6 won't cost more than $30-35. So if a 107 mm2 then costs about half, that, you can assume at most a $20 difference for the chip.

If you're talking about the Semiwiki guy, I got bad news for ya. He's basically the TSMC witeken.

 

[Aapje]

If you're talking about the Semiwiki guy, I got bad news for ya. He's basically the TSMC witeken.

Don't know Witeken, according to his Twitter he got banned for being under 13. Somehow I don't think that he has the credibility of a retired chip designer.

But I still just used the commonly cited prices and didn't go below them, so what are we even arguing about?

 

[insertcarehere]

You guys do realize that the engineering costs for the drivers and even the hardware are spread across the entire line of cards and not just a single one right? And it's not like they start over with drivers. They use the same drivers as they did for previous generations and just make changes to them. The majority of engineering was already paid for when the 7900 XTX & XT were launched.

If you're trying to argue that the R&D costs for AMD wouldn't change much even without N33 in the mix... I am not sure that's a valid argument as without N33, AMD wouldn't need to backport RDNA3 to N6, nor would they have to make monolithic RDNa3. All of which are costs borne solely from having N33 in the lineup.

 

[jpiniero]

Don't know Witeken, according to his Twitter he got banned for being under 13.

That got me to look and it's still there.

 

[USER8000]

This doesn't sound sensible at all.
What you want for N32 is the same thing AMD did with N33, and we know from reviews It's a flop. Basically, only a port of N22 to 6nm using RDNA3 architecture.
There is no point for such GPU when we already have N22.
.
5nm N32 with 60CU, 64MB IC, 256bit 16GB GDDR6 sounds much better to me than 6nm version with 40CU, 96MB IC, 192bit 12GB GDDR6.

Though with 5nm being in demand it could be that the rx7600 will be easier for amd to make in bulk.

 

[Joe NYC]

You guys do realize that the engineering costs for the drivers and even the hardware are spread across the entire line of cards and not just a single one right? And it's not like they start over with drivers. They use the same drivers as they did for previous generations and just make changes to them. The majority of engineering was already paid for when the 7900 XTX & XT were launched.

No, it wasn't already paid for. Each RDNA card, including every 7600 gets allocated portion the driver development, portion of the common RDNA3 development particular Navi 33 die development and cost of its masks, cost of the corporate overhead.

People who are figuring out the cost of 7600 by using the wafer / die cost calculator and think they know the cost of the card are smoking crack.

 

[Mopetar]

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 physical bus isn't expensive, but you need to add more physical interface for a wider bus and that takes up a fixed amount of area on die, so doesn't benefit from node shrinks. It just makes each die more expensive. Hence AMD using the older N6 node for their MCDs.

It also means that there are more memory chips that you need to put on the card. A 256-but bus will need twice the memory modules as a 128-bit bus. Historically a lot of the reason for having massive busses was because the chips themselves didn't have a lot of capacity and so you just needed more of them to hit a total of 4 GB VRAM. Now we're getting to the point where 4 GB will be available on a single chip, so if you're just targeting 8 GB of VRAM you could do it with a smaller bus. Memory speeds have also increased which does offset reduced bus size to some extent, but that's only linear growth and wouldn't offset the halving of bus-width in terms of bandwidth.

But it general, smaller busses are less expensive because the die will be a lot smaller and the cards need fewer memory chips.

I think you forgot the cost of the entire design team, hardware and software.

That gets too far into the weeds as AMD isn't selling only Navi 23 cards. Some of the design costs are shared with CDNA where margins are higher and their partnership with console manufacturers also pays for some of these expenses.

Obviously if margins are lower you need a lot of volume to cover high fixed costs, but there are many industries that work this way because they're largely dealing in commodity parts or products.

 

[In2Photos]

No, it wasn't already paid for. Each RDNA card, including every 7600 gets allocated portion the driver development, portion of the common RDNA3 development particular Navi 33 die development and cost of its masks, cost of the corporate overhead.

People who are figuring out the cost of 7600 by using the wafer / die cost calculator and think they know the cost of the card are smoking crack.

I literally said the cost of engineering and drivers was spread across the entire line of cards. But the majority of that cost is recouped from the sale of the top tier products as they are higher margin items.