Question AMD Phoenix/Zen 4 APU Speculation and Discussion

[uzzi38]

I can finally make this thread.

https://twitter.com/x/status/1519669375283957760

Phoenix is indeed RDNA3. My advice to everyone: treat the old APU rumours as being out of date.

 

[Shivansps]

Final price has nothing to do with costs. Thats rule #1 on new product development. If final price has to be lower than costs thats because you sc* up and you have never made that product in the first place. Costs only tells you how low you can actually go, but that does not really define the final price, that just excuses for increasing the prices.

AMD is not interested in making a 4C or a 6C APU with 12CU because they know that will be good enoght for the vast mayority of people, they could make it, sell it at a good margin even, it would be a very successfull soc. But at the end of the day they will losse money because that will mean they loss a lot of sales on big APUs and dGPUs were the margin is a lot higher.

They rather throw away dies that has the full 12CU gpu working but has a broken cpu core than selling you that.

 

[FlameTail]

You guys make good points.

Even Apple segments their M chips SKUs with both CPU and GPU tied together.

 

[qmech]

Am I the only one who wishes the CPU and GPU weren't tied together in the APU segmentation?

For example:

Ryzen 7840 : 8-core CPU, 12 CU GPU
Ryzen 7640 : 6-core CPU, 6 CU GPU

Why does the CPU and GPU have to be tied together? It's been this way in previous generations too.

Why can't we have a 6-core CPU + 12 CU GPU combo?

coercitiv absolutely nailed it. It's a question of production cost versus ASP.

AMD currently uses just 2 different die layouts to cover the entire 7040 range, Phoenix and Phoenix2/Little Phoenix. The former features 12 compute units (6 workgroups) and 8 Zen4 cores. The latter 4 compute units (2 work groups) and 2 Zen4 cores + 4 Zen4c cores.

You ask why CPU and GPU have to be tied together and the simple answer is that they *physically* are. Unless AMD switches to a chiplet approach, with the significant efficiency hit that would entail, that is just a fundamental limitation.

It is certainly possible that AMD could create an "in-between" SKU where two of the Zen4 cores in Phoenix are deactivated. For this to make sense economically, there would have to be enough dice where at least 1 Zen4 core is defective, but all 12 compute units are fully operational. There would still need to be enough 6-core/8-CU to fill the middle tier, however, and prices would get very close to each other.

There would also be the risk of this new SKU cannibalizing the 8-core/12-CU SKUs. AMD would need to price the SKU appropriately to minimize this effect, meaning that as a consumer, you would probably not save much.

If we ignore any threats to cannibalizing discrete GPU sales, the more interesting approach would have been a 2 Zen4 + 6 Zen4c die with 16+ compute units and infinity cache. Phoenix only has 2MB "L2" cache for the iGPU, which is not much when considering the low (and shared) bandwidth and high latency to main memory. Such a beast would be of similar die size to the current Phoenix chip and could realistically cover all the same SKUs, enabling AMD to phase out Phoenix.

If you really wanted to dream, you could eject the XDNA AI processor, which takes up roughly as much space as five Zen4c cores, and replace that with another 4 compute units and even more graphics cache.

Could we see some of that in Hawk Point? Maybe, but AMD seems to be doubling down on the NPU (60% increase in performance over Phoenix) and the roadmaps show RDNA 3. For the majority of its life, Hawk Point is going to be overshadowed by Zen5-based offerings, so my idle speculation would be that AMD prioritizes cost reductions. I wouldn't be at all surprised to see a smaller die with 2 Zen4 + 6 Zen5c, virtually same iGPU, and some of the die size savings used to get XDNA1 up to a reasonable performance.

 

[maddie]

Final price has nothing to do with costs. Thats rule #1 on new product development. If final price has to be lower than costs thats because you sc* up and you have never made that product in the first place. Costs only tells you how low you can actually go, but that does not really define the final price, that just excuses for increasing the prices.

AMD is not interested in making a 4C or a 6C APU with 12CU because they know that will be good enoght for the vast mayority of people, they could make it, sell it at a good margin even, it would be a very successfull soc. But at the end of the day they will losse money because that will mean they loss a lot of sales on big APUs and dGPUs were the margin is a lot higher.

They rather throw away dies that has the full 12CU gpu working but has a broken cpu core than selling you that.

I think you're overestimating the production cost reduction of a lower-end product using the latest tech.

What appears apparent, at least to me, is the small area savings attained on modern SOCs by limiting either/both the CPU and GPU segments. Phoenix 1 & 2 are two good examples. A 23% area saving and you lose 67% of the GPU with the CPU going from 8 full Zen4 to 2 full zen 4 & (4) Zen4C. In Intel lingo, the "uncore" is almost the chip now.

 

[Shivansps]

I think you're overestimating the production cost reduction of a lower-end product using the latest tech.

What appears apparent, at least to me, is the small area savings attained on modern SOCs by limiting either/both the CPU and GPU segments. Phoenix 1 & 2 are two good examples. A 23% area saving and you lose 67% of the GPU with the CPU going from 8 full Zen4 to 2 full zen 4 & (4) Zen4C. In Intel lingo, the "uncore" is almost the chip now.

Phoenix 2 is nothing new, this is what they did on Picasso era, it was called Raven2 or Dali, i dont remember if both were the same or what.

That die was used to make the the refreshed Athlon lineup for desktop 3000g and similar products(to reeplace the Athlon 200 series that were full cut down Picassos), and the Athlon 3000 series for mobile, and some of the Ryzen 3 for mobile, like the 3250U.

This is the same, the only difference is that they now belive they can launch souch a product as a Ryzen 5 on desktop and get away with it. The price will have to be reallly good.

 

[moinmoin]

Phoenix 2 is closer to Raven Ridge and Picasso. The successor to Raven2 is Mendocino.

 

[LightningZ71]

If we believe the reports and rumors of very good N5/N4 yields, it is likely that the most common defective SOCs have a single defective element, either a single core, a single WGP, the NPU, some part of the uncore, or some sort of interconnect flaw. There's no reason to believe that their 6 core Phoenix 1 parts don't typically have a fully functional iGPU with just a portion fused off for intentional segmentation, or, a flawed iGPU with a fully functional CCX that has good cores fused off.

AMD could choose to bin 6 big core APUs with 12 CU iGPUs and market them as a Z1R or something like that for a target market. Its unlikely that they release an AM5 part like that though as they prefer to push you up market.

 

[jpiniero]

Most users would prefer an 8 core CPU / 6 CU GPU, IMO.

 

[FlameTail]

Most users would prefer an 8 core CPU / 6 CU GPU, IMO.

Wahut

 

[jpiniero]

Wahut

Course you would be better off with Raphael.

5700G's original MSRP was $359, 5600G's was $259. I imagine the price would be a tad higher for the Big Phoenix models? The Little Phoenix would then be about $200 and $150-175.

 

[SteinFG]

Course you would be better off with Raphael.

5700G's original MSRP was $359, 5600G's was $259. I imagine the price would be a tad higher for the Big Phoenix models? The Little Phoenix would then be about $200 and $150-175.

That was when 5600X was selling for 300. Once chip shortage was over, prices quickly fell. Even 7600X doesn't cost 300 now.

 

[Shivansps]

Phoenix 2 is closer to Raven Ridge and Picasso. The successor to Raven2 is Mendocino.

Mendocino is more a product for embedded applications that compite with Intel ADL-N than anything else, it is so cut down that they can never launch it on AM5. The closer thing to Raven2 right now is Phoenix2. I do agree that they were using Raven2 to cover that market but it was no adequate for it, you need to cut down more stuff and make something that is more in line with what Intel does with their small cores, that what Mendocino is.

Course you would be better off with Raphael.

5700G's original MSRP was $359, 5600G's was $259. I imagine the price would be a tad higher for the Big Phoenix models? The Little Phoenix would then be about $200 and $150-175.

The original prices should be around 3400G MSRP (150-180), everything after that got artificially bumped due to the chip shortage.

 

[dr1337]

There's no reason to believe that their 6 core Phoenix 1 parts don't typically have a fully functional iGPU with just a portion fused off for intentional segmentation, or, a flawed iGPU with a fully functional CCX that has good cores fused off.

I think their main goals in binning isn't so much market segmentation but hitting advertised TDPs and clocks. Assuming good yields I agree its more likely for defects to be minor. But if defective chips can't hit target clocks or consume 10w more than they should,. it really simplifies things to just fuse off parts to spec instead of having a SKU for every possible die and power variation. Availability problems comes into play if a certain bin is too rare.

This is why phoenix 2 exists at all, start off with a 30-15w part and you don't have to cut off nearly as much silicon as you do with a chip that is designed for up to 90w. Meaning faster lead times and more availability for OEMs as well as less time spent binning and less good silicon wasted.

This is also why exactly AMD is so slow to get APUs to desktop; It takes a year of binning the worst and hottest, yet fully functional, dies they have to get slapped onto a socket-able package and sold to feeble DIY builders.

 

[SteinFG]

Its unlikely that they release an AM5 part like that though as they prefer to push you up market.

Groundbreaking discovery: All companies want you to buy a higher-priced product. heh.

Upping CU count with each tier is one way of incentivizing it. It's all about a product portfolio, different cpu for different price points. I mean, i3/i5/i7 was made to achieve exactly that - have low-end offers for anyone who can't spend much, and each tier up offer better perf/$ to incentivize people to spend more if they are willing to.

 

[eek2121]

Mendocino is more a product for embedded applications that compite with Intel ADL-N than anything else, it is so cut down that they can never launch it on AM5. The closer thing to Raven2 right now is Phoenix2. I do agree that they were using Raven2 to cover that market but it was no adequate for it, you need to cut down more stuff and make something that is more in line with what Intel does with their small cores, that what Mendocino is.


The original prices should be around 3400G MSRP (150-180), everything after that got artificially bumped due to the chip shortage.

Nah, Ryzen embedded is for embedded applications. Mendocino is for $200-$400 “budget” laptops.

 

[jpiniero]

The original prices should be around 3400G MSRP (150-180), everything after that got artificially bumped due to the chip shortage.

Given that it costs more than Raphael, I see no reason to think either Big Phoenix model will be less than the 7600X, which currently is $249 at Newegg. I think the only reason it's getting a release is because of the end of the Vega drivers.

 

[Abwx]

Given that it costs more than Raphael, I see no reason to think either Big Phoenix model will be less than the 7600X, which currently is $249 at Newegg. I think the only reason it's getting a release is because of the end of the Vega drivers.

The lower parts wont cost much, the curves published by AMD for the 2C + 4c point to barely more than 45W for the 8500G and something like 35W for the 8300G, that s 3x their targeted TDPs for mobile variants and well above their efficiency range.

These should be low cost parts for everyday PCs, i would expect 100-110$ for the 8300G and 150$ for the 8500G.

The 8700G will likely be in the 300-320$ segment and the 8600G at 200-220$, these are high performing parts comparatively, FI the 6C 8600G@45W should perform better than the 2C + 4c 8500G@45W.

 

[qmech]

Is anything known about Hawk Point? (Other than its existence on roadmaps showing 1Q2024 availability and a bump in XDNA TOPS.)

There's a Geekbench entry from yesterday that's making the rounds with what appears to be 2024 version of the 7940HS. What's notable is that it scores quite a bit higher in Geekbench 6.2.1 than the 7940HS (+100 points in ST, +1400 MT) despite apparently having same base/boost frequency and same power consumption.

My initial thought was that AMD simply bumped the TDP or that the processor was a nicely binned selection like the Lenovo S models, but the AuthenticAMD CPU identifier is different, so there *might* be something more in it. The 7940HS (and other 7040 family "big" Phoenix) features family 25, model 116, stepping 1. This new chip is model 117, stepping 2.

Standard Standard - Geekbench

Benchmark results for a Standard Standard with an AMD Eng Sample: 100-000001319-50_Y processor.

browser.geekbench.com

 

[tamz_msc]

Is anything known about Hawk Point? (Other than its existence on roadmaps showing 1Q2024 availability and a bump in XDNA TOPS.)

There's a Geekbench entry from yesterday that's making the rounds with what appears to be 2024 version of the 7940HS. What's notable is that it scores quite a bit higher in Geekbench 6.2.1 than the 7940HS (+100 points in ST, +1400 MT) despite apparently having same base/boost frequency and same power consumption.

My initial thought was that AMD simply bumped the TDP or that the processor was a nicely binned selection like the Lenovo S models, but the AuthenticAMD CPU identifier is different, so there *might* be something more in it. The 7940HS (and other 7040 family "big" Phoenix) features family 25, model 116, stepping 1. This new chip is model 117, stepping 2.

Standard Standard - Geekbench

Benchmark results for a Standard Standard with an AMD Eng Sample: 100-000001319-50_Y processor.

browser.geekbench.com

The margins of the Geekbench scores you indicate are within run-to-run variance and data point variance of the benchmark for a given CPU model.

 

[igor_kavinski]

The 7940HS (and other 7040 family "big" Phoenix) features family 25, model 116, stepping 1. This new chip is model 117, stepping 2.

Standard Standard - Geekbench

Benchmark results for a Standard Standard with an AMD Eng Sample: 100-000001319-50_Y processor.

browser.geekbench.com

Standard Standard vs ASUSTeK COMPUTER INC. ROG Zephyrus G14 GA402XI_GA402XI - Geekbench

browser.geekbench.com

Performance uplift in some ST benchmarks is too high (22% to 40%) so probably better frequency boost coupled with faster RAM?

 

[qmech]

The margins of the Geekbench scores you indicate are within run-to-run variance and data point variance of the benchmark for a given CPU model.

Yes, they certainly are - and given these are laptops with wildly varying cooling solutions, there is ample opportunity to fit in a handful of Occam's razors. Hence the bolded *might* in my post above.

What caught my eye in all of this was really the difference in model#.

 

[qmech]

Standard Standard vs ASUSTeK COMPUTER INC. ROG Zephyrus G14 GA402XI_GA402XI - Geekbench

browser.geekbench.com

Performance uplift in some ST benchmarks is too high (22% to 40%) so probably better frequency boost coupled with faster RAM?

How accurate is Geekbench's boost and memory bandwidth detection? Both seem very much in line with 7940HS numbers. But even minor changes to the amount of time the processor is allowed to boost could explain the difference. Again, the model# change is what piqued my interest.

 

[SteinFG]

Is anything known about Hawk Point? (Other than its existence on roadmaps showing 1Q2024 availability and a bump in XDNA TOPS.)

Rumors point to it being a respin of the same phoenix + overclocked NPU

 

[Shivansps]

Nah, Ryzen embedded is for embedded applications. Mendocino is for $200-$400 “budget” laptops.

Yes, as i said, Mendocino is tailored for diferent market, the same as ADL-N, and as a result they cant launch it on AM5 due to the cutdowns. Raven2 was on AM4, budget mobile devices and the embedded market.

The lower parts wont cost much, the curves published by AMD for the 2C + 4c point to barely more than 45W for the 8500G and something like 35W for the 8300G, that s 3x their targeted TDPs for mobile variants and well above their efficiency range.

These should be low cost parts for everyday PCs, i would expect 100-110$ for the 8300G and 150$ for the 8500G.

The 8700G will likely be in the 300-320$ segment and the 8600G at 200-220$, these are high performing parts comparatively, FI the 6C 8600G@45W should perform better than the 2C + 4c 8500G@45W.

I wonder if AMD is going to allow reviewers to compare it with older apu from 2200G to a 5700G. I really want to see the 8500G compared to a 2400G/3400G to see what the 5/6-years worth of improvements looks like.

 

[moinmoin]

I wonder if AMD is going to allow reviewers to compare it with older apu from 2200G to a 5700G.

How would AMD prevent reviewers from doing so? There are plenty reviewers doing reviews of stuff they bought themselves and as such are not bound to any inane rules manufacturers might come up with.