The new unobtanium™ :laugh:

Sweet, me likey.

Why not increase iGPU frequency from 2100 to 2500MHz or insert 12 Vega CU on 2100MHz? This is APU with 8 physical cores and who needs of too powerful close to 5GHz CPU part when choose APU? Normal APU are intended for small with very small power consumption and cheap PC, for office, movies, browsing and light gaming not as part of high performance desktops with discrete graphics but I see to offer many such perversions.

@AleksandarK "dektop"

TumbleGeorgeWhy not increase iGPU frequency from 2100 to 2500MHz or insert 12 Vega CU on 2100MHz? This is APU with 8 physical cores and who needs of too powerful close to 5GHz CPU part when choose APU? Normal APU are intended for small with very small power consumption and cheap PC, for office, movies, browsing and light gaming not as part of high performance desktops with discrete graphics but I see to offer many such perversions.

There's many use cases of needing a powerful CPU but not necessarily a powerful GPU. This would bring more Ryzen processors accessible to office machines for example which might do a lot of number crunching but don't need a really powerful GPU. Which without an APU usually means requiring a dedicated GPU bolted on would have made these too expensive and complicated rather than just an intel system with an iGPU.

TumbleGeorgeWhy not increase iGPU frequency from 2100 to 2500MHz or insert 12 Vega CU on 2100MHz? This is APU with 8 physical cores and who needs of too powerful close to 5GHz CPU part when choose APU? Normal APU are intended for small with very small power consumption and cheap PC, for office, movies, browsing and light gaming not as part of high performance desktops with discrete graphics but I see to offer many such perversions.

I'd gladly combine this chip with a big chunky GPU in my system. It's quick enough for basically all normal CPU tasks, and would allow the external GPU to sleep for most of the time unless needed for compute or gaming.
True, it would be better if the performance delta of the iGPU was a bit bigger, but there always has to be some trade-off.
Now they can compete on pure CPU performance again they don't need to distinguish themselves that strongly in other areas anymore and imho it makes sense that they're focusing more on the part that has the biggest impact for most of their users.

owen10578There's many use cases of needing a powerful CPU but not necessarily a powerful GPU. This would bring more Ryzen processors accessible to office machines for example which might do a lot of number crunching but don't need a really powerful GPU. Which without an APU usually means requiring a dedicated GPU bolted on would have made these too expensive and complicated rather than just an intel system with an iGPU.

In this moment from 4600G and above CPU part is more than enough. APUs with 6 and 8 cores...For what, molecular dynamic with APU?

TumbleGeorgeIn this moment from 4600G and above CPU part is more than enough. APUs with 6 and 8 cores...For what, molecular dynamic with APU?

Ah, yes, because it's entirely unrealistic to run any type of workstation workload without a dGPU, right? Yeah, sorry, that's not true. There are still plenty of workloads that can't be GPU accelerated well, and these chips should be excellent for those. And if you need the GPU, it has the PCIe lanes to add one, without the system integrator needing to validate a new CPU.

Every time I see these die layout images, I'm reminded and saddened how little area a single Vega CU takes up, and yet they removed three of them.

The reduction from Raven Ridge/Picasso to Renoir/Cezanne isn't from Vega10 to Vega8, because very few models have all CUs enabled (yields/defects?) Vega10 is 10/11 and the equavalent is Vega7 because the 4800U/5800U will be so exceptionally rare that they're either unavailable or priced outside any reasonable value for most people.

From Vega10 to Vega7 is a big oof. Bandwidth be damned, there's still plenty for a better IGP at the resolutions and framerates APUs target...

any mention of wattage?

DeathtoGnomesany mention of wattage?

Likely 65W like their Renoir predecessors.

AMD made a big thing at Zen3 launch about everything being faster despite using the exact same TDP as the previous generation.

Chrispy_Every time I see these die layout images, I'm reminded and saddened how little area a single Vega CU takes up, and yet they removed three of them.

The reduction from Raven Ridge/Picasso to Renoir/Cezanne isn't from Vega10 to Vega8, because very few models have all CUs enabled (yields/defects?) Vega10 is 10/11 and the equavalent is Vega7 because the 4800U/5800U will be so exceptionally rare that they're either unavailable or priced outside any reasonable value for most people.

From Vega10 to Vega7 is a big oof. Bandwidth be damned, there's still plenty for a better IGP at the resolutions and framerates APUs target...

While I agree in principle, and I definitely want a higher CU count APU ASAP, given that these are quite large monolithic dice made on a cutting-edge node it's kind of understandable that it hasn't happened yet. And I don't think it will for a while, sadly.

Renoir is reasonably large at ~149mm2 (~11*13.6mm). Now, the aspect ratios from the Videocardz die shot above and AnandTech's measurements don't quite line up (the VC shot is a bit less than 0.75:1 while 11*13.6 is 0.81:1), so these numbers will be a bit off. Still, Cezanne has a more square (but not quite) aspect ratio - assuming a similar error as for Renoir and that the die width is the same (which it looks like) let's say it's 13,6mm square for the sake of simplicity. That's 185mm2. That relatively small increase in die size means a drop in dice per wafer from 381 to 304 (20%) and likely defect-free dice from 333 to 258 (22,5%), assuming the same process and yields (with TSMC's published 0.09/cm2 error rate). So just by going from Zen 2 to Zen 3 we're increasing per-die costs by >20%. Using some slightly oversimplified math (defective dice are still utilized as much as possible after all), assuming a ~$20 000 cost per wafer that's an increase in per-die cost from ~$60 to ~$77.5 Adding even a couple more CUs would further increase costs, possibly bringing per-die costs close to the $100 range.

Now, this would be reasonable for a premium, high-performance part. The issue is that these are mass-market parts used in the entire spectrum of APUs from AMD. This also at least partially explains the continued use of the Lucienne Zen 2 SKUs in the mobile APU lineup, btw, as they will still perform well (and have power efficiency updates since Renoir) but are significantly cheaper and easier to produce in higher volumes. But that's besides the point.

The issue here is that AMD is using a single monolithic die to create everything from high-end mobile APUs (35-45W Ryzen 7/9 5xxxH) to desktop APUs (65W Ryzen 7/9 5xxxG) to mid-range thin-and-light chips like the 15W 5600U. They've excluded these more expensive chips from the very low end by using Lucienne there (and I guess also the rumored Van Gogh for lower-power implementations, though it's so far nowhere to be seen, and rumored to have been cancelled due to fab constraints). That's a pretty large range, and it inevitably means a cost squeeze for the lower end parts. Raising per-die production costs makes that squeeze harder - and lower end SKUs sell in much higher volumes than higher end ones, after all. the Ryzen 5 5600U is likely to outsell the Ryzen 7 5800H by 10:1 or more, so it's imperative that they make some money even on the low end SKUs.

This dilemma can be solved one of two ways: by making more monolithic dice (which AMD is sort of sliding into, though not through concurrently designing several variants of the same silicon like Intel, but rather by iterating on older designs while introducing newer ones), or by implementing a chiplet approach for APUs too. The former, if continued to the point that they need to concurrently design 2-3 APU dice each generation, is really expensive. The latter is, so far, seemingly too complex. This might be down to substrate-based IF links being too power hungry, issues packaging and cooling direct-die MCM laptop chips, package size constraints (MCM needs more area than monolithic dice), a combination thereof, or anything else, really. I still think we'll see MCM APUs in the future, but I think until then, we won't see a proper high-performance iGPU APU like us enthusiasts seem to want - it's simply too expensive, and doesn't make sense economically in the grand scheme of chip production.

I'm hoping that AMD is planning to get MCM APUs into the market alongside mainstream (LP)DDR5 adoption in laptops. That would be pretty much perfect. But IMO that's a pretty optimistic vision.

ValantarWhile I agree in principle, and I definitely want a higher CU count APU ASAP, given that these are quite large monolithic dice made on a cutting-edge node it's kind of understandable that it hasn't happened yet. And I don't think it will for a while, sadly.

Renoir is reasonably large at ~149mm2 (~11*13.6mm). Now, the aspect ratios from the Videocardz die shot above and AnandTech's measurements don't quite line up (the VC shot is a bit less than 0.75:1 while 11*13.6 is 0.81:1), so these numbers will be a bit off. Still, Cezanne has a more square (but not quite) aspect ratio - assuming a similar error as for Renoir and that the die width is the same (which it looks like) let's say it's 13,6mm square for the sake of simplicity. That's 185mm2. That relatively small increase in die size means a drop in dice per wafer from 381 to 304 (20%) and likely defect-free dice from 333 to 258 (22,5%), assuming the same process and yields (with TSMC's published 0.09/cm2 error rate). So just by going from Zen 2 to Zen 3 we're increasing per-die costs by >20%. Using some slightly oversimplified math (defective dice are still utilized as much as possible after all), assuming a ~$20 000 cost per wafer that's an increase in per-die cost from ~$60 to ~$77.5 Adding even a couple more CUs would further increase costs, possibly bringing per-die costs close to the $100 range.

Now, this would be reasonable for a premium, high-performance part. The issue is that these are mass-market parts used in the entire spectrum of APUs from AMD. This also at least partially explains the continued use of the Lucienne Zen 2 SKUs in the mobile APU lineup, btw, as they will still perform well (and have power efficiency updates since Renoir) but are significantly cheaper and easier to produce in higher volumes. But that's besides the point.

The issue here is that AMD is using a single monolithic die to create everything from high-end mobile APUs (35-45W Ryzen 7/9 5xxxH) to desktop APUs (65W Ryzen 7/9 5xxxG) to mid-range thin-and-light chips like the 15W 5600U. They've excluded these more expensive chips from the very low end by using Lucienne there (and I guess also the rumored Van Gogh for lower-power implementations, though it's so far nowhere to be seen, and rumored to have been cancelled due to fab constraints). That's a pretty large range, and it inevitably means a cost squeeze for the lower end parts. Raising per-die production costs makes that squeeze harder - and lower end SKUs sell in much higher volumes than higher end ones, after all. the Ryzen 5 5600U is likely to outsell the Ryzen 7 5800H by 10:1 or more, so it's imperative that they make some money even on the low end SKUs.

This dilemma can be solved one of two ways: by making more monolithic dice (which AMD is sort of sliding into, though not through concurrently designing several variants of the same silicon like Intel, but rather by iterating on older designs while introducing newer ones), or by implementing a chiplet approach for APUs too. The former, if continued to the point that they need to concurrently design 2-3 APU dice each generation, is really expensive. The latter is, so far, seemingly too complex. This might be down to substrate-based IF links being too power hungry, issues packaging and cooling direct-die MCM laptop chips, package size constraints (MCM needs more area than monolithic dice), a combination thereof, or anything else, really. I still think we'll see MCM APUs in the future, but I think until then, we won't see a proper high-performance iGPU APU like us enthusiasts seem to want - it's simply too expensive, and doesn't make sense economically in the grand scheme of chip production.

I'm hoping that AMD is planning to get MCM APUs into the market alongside mainstream (LP)DDR5 adoption in laptops. That would be pretty much perfect. But IMO that's a pretty optimistic vision.

Oh yeah, I understand entirely why the balance is what it is. A classic example of "one size does NOT fit all"

These APUs in various states of yield and config have to serve every market segment from budget laptops for non-gamers all the way up to professionals needing certified 3D API support.

I am sad because the choices that make sense for the many do not align with the needs of the few, especially when I am one of those few. MCM's may well fix the problem by adding some flexibility into the array rather than the "one size fits all" that we're stuck with at the moment.

Ryzen Pro is actually the real threat to Intel. iGPU-less chips like 5600X etc are a small part of the market requiring a dGPU, and even if they captured 100% of their part it wouldn't have much impact to unit sales.

These target the corporate desktop as well as the huge (vast majority) of the retail desktop market that neither needs nor wants to pay for a dGPU. Of course, ability to deliver product in volume is critical here or the big OEMs won't bite as they won't deal with supply chain disruption. Ryzen Pro 4XXX didn't make much impact, but these might if AMD can get past their supply issues.

Most likely for OEM only, just like previous Ryzen 7 PRO 4750G.

So the naming is one gen ahead of reality, cool. Next gen that means they need to bring a 5.75 Ghz APU for the 6750G.

Right? :toast:

Vayra86So the naming is one gen ahead of reality, cool. Next gen that means they need to bring a 5.75 Ghz APU for the 6750G.

Right? :toast:

Here's hoping Intel responds in kind with 11.9 GHz versions of their 12th gen CPUs :D

TumbleGeorgeWhy not increase iGPU frequency from 2100 to 2500MHz or insert 12 Vega CU on 2100MHz? This is APU with 8 physical cores and who needs of too powerful close to 5GHz CPU part when choose APU? Normal APU are intended for small with very small power consumption and cheap PC, for office, movies, browsing and light gaming not as part of high performance desktops with discrete graphics but I see to offer many such perversions.

There is no point in either of those things. The GPU is so severely memory bandwidth limited that the CU count increase would benefit only fringe use cases. Expect more when DDR5 is out of the door.

MathraghI'd gladly combine this chip with a big chunky GPU in my system. It's quick enough for basically all normal CPU tasks, and would allow the external GPU to sleep for most of the time unless needed for compute or gaming.
True, it would be better if the performance delta of the iGPU was a bit bigger, but there always has to be some trade-off.
Now they can compete on pure CPU performance again they don't need to distinguish themselves that strongly in other areas anymore and imho it makes sense that they're focusing more on the part that has the biggest impact for most of their users.

Agreed.

I love APUs for the fact that you don't need a graphics card for diagnostics. I, for example have a 5700 XT in my PC. If it decides to stop working for some reason, or I just sell it to get something new, I'll have to wipe its driver and install the nvidia one for my "diagnostic" 1050 Ti. And if my new GPU is from AMD as well, then I'll have to wipe the nvidia driver, and install the AMD one again. I'm not lazy, but it just clutters my OS with pointless registry entries and stuff. With an APU, I could just take my graphics card out, and use the PC as normal.

All I wonder is: will these APUs be any more available for DIYers than the 4000 series ones?

the 3400G just came in the mail yesterday. I guess in 5 years when the 3400G is ready for an upgrade this might be available?

I am not trolling, all you guys complaining for lack of GPU performance, what do you use your iGPU for? My questions comes from ignorance probably, but I just don't see the point. Clearly, if you chose to go APU you have prioritized CPU performance in a small package and, for me at least, no need for 3D performance. The way I look at it, if you need proper GPU performance, you would not choose and APU due to lack of PCI-e lanes and other restrains compared to dedicated CPU.

DrediThere is no point in either of those things. The GPU is so severely memory bandwidth limited that the CU count increase would benefit only fringe use cases. Expect more when DDR5 is out of the door.

That's not quite true. I saw significant performance increases on my 4650G moving not only from DDR4-3200 to 3800 but also 1900MHz to 2100MHz. With faster DDR4 (which these should handle easily) or LPDDR4X on mobile, these could probably benefit significantly from higher clocks still.

jonupI am not trolling, all you guys complaining for lack of GPU performance, what do you use your iGPU for? My questions comes from ignorance probably, but I just don't see the point. Clearly, if you chose to go APU you have prioritized CPU performance in a small package and, for me at least, no need for 3D performance. The way I look at it, if you need proper GPU performance, you would not choose and APU due to lack of PCI-e lanes and other restrains compared to dedicated CPU.

Some people want to build small, compact, low-power PCs yet still want performance. I have a 4650G in my HTPC, and I'm very happy I didn't have to stick a dGPU in there, as it would have increased noise and power draw even if it would also obviously have increased performance. The 4650G can handle Rocket League and other non-graphically intensive games at good frame rates, runs dead quiet, sips power, and I only have a single fan in the system, so it's pretty much perfect for me. My case can fit an LP GPU, so I could have stuck a 1650S in there, but that would have added two ~60mm fans running constantly, which was a no-go for HTPC use. And a full height GPU would have necessitated a larger case and/or worse CPU cooling. It's all a question of balance.

And obviously in the future, if I can run more stuff natively on the HTPC without moving away from its tiny, semi-passively cooled, single-fan build, that's something I'd want.

ValantarThat's not quite true. I saw significant performance increases on my 4650G moving not only from DDR4-3200 to 3800 but also 1900MHz to 2100MHz. With faster DDR4 (which these should handle easily) or LPDDR4X on mobile, these could probably benefit significantly from higher clocks still.


Some people want to build small, compact, low-power PCs yet still want performance. I have a 4650G in my HTPC, and I'm very happy I didn't have to stick a dGPU in there, as it would have increased noise and power draw even if it would also obviously have increased performance. The 4650G can handle Rocket League and other non-graphically intensive games at good frame rates, runs dead quiet, sips power, and I only have a single fan in the system, so it's pretty much perfect for me. My case can fit an LP GPU, so I could have stuck a 1650S in there, but that would have added two ~60mm fans running constantly, which was a no-go for HTPC use. And a full height GPU would have necessitated a larger case and/or worse CPU cooling. It's all a question of balance.

And obviously in the future, if I can run more stuff natively on the HTPC without moving away from its tiny, semi-passively cooled, single-fan build, that's something I'd want.

This. Not to mention, when ones GPU breaks and is busy with an RMA process, you still have something to work/play with.

Fast APUs with Vega iGPU can be an excellent choice for gamers playing a mix of old and new games. You can run your new games on your shiny newest super powerful dGPU (and the CPU part will be able to handle it) and when the nostalgia kicks in, you can just use the fast enough and rock solid Vega and forget about any visual glitches and frame drops.

krukFast APUs with Vega iGPU can be an excellent choice for gamers playing a mix of old and new games. You can run your new games on your shiny newest super powerful dGPU (and the CPU part will be able to handle it) and when the nostalgia kicks in, you can just use the fast enough and rock solid Vega and forget about any visual glitches and frame drops.

This only applies to mesa, the Windows drivers are terrible with old games.