AMD's 25x20 Analysed 

TechTechPotato approved.

Hmmmmm last time i was this early some people were selling 4 core cpus as flagship.

03:30 bit of trivia, Kaveri is Finnish and means Friend. Apu as well is a Finnish word meaning Help. Loosely translated APU Kaveri then means Helping Friend

AMD E350 (atom competitor) was actually amazingly power efficient. Only 18W from the wall in use, WITH 32GB of DDR3 :O

I've spent my lockdown practicing your accent whenever I answer the phone "alright guys ows it go'in".

That "alright guys how's it going?" intro lol

I think that the logarithmic plot makes more sense than linear when comparing how much % improvement there is. Any percentage is always a fixed vertical distance on the log chart, while for the linear chart the vertical distance depends on the moment in time.

HSA was pretty awesome. If it ever took off it could have really changed the playing field quite a bit. If course Intel could have worked it into theirs had it picked up support and probably won again anyway.

Absolutely fascinating as ever. Was a bulldozer owner, believe it or not under Linux it was always the chip for my workloads so I studied the ways to get the most out of it. Through that I came to research a lot about AMD architecture and as a by product came to see where they were heading. This research here is icing, you've gone much deeper than I could have. They might have banged the marketing and benchmarks, but the tech people were working their nuts off and I think we both see that. What AMD have achieved is stunning no matter how you measure or try to sell it.

The numbers are a bit fishy, but Renoir really is a massive leap forward in efficiency and they weren't totally wrong about heterogeneous compute. GPU utilisation has massively increased over the past six years, from GPU-accelerated compositing in the browser to GPU acceleration in productivity software like Photoshop and Davinci. GPU performance has become vastly more important in real-world usage, in a way that's quite hard to capture in a benchmark suite. I'm not sure it really matters whether they achieved the 25x20 plan, because the real-world achievement is arguably better than they could have hoped for. Kaveri APUs were crap chips for bargain-basement laptops, but Renoir is the one to beat at the high end. AMD were very lucky that everything came together this year, but they were also very smart - Zen delivered competitive performance on desktop, while Zen 2 built on that foundation to deliver competitive efficiency on mobile.

To be fair, if AMD intended to rely on HSA to get improved floating point performance, which was to be reflected in PCMark, then Cinebench is a reasonable way to compensate for the absence of GPU-accelerated FP performance. Where AMD went wrong is to not include a pure integer benchmark like SpecIntRate. On the other hand, the 3DMark result, which weighs the C score down, didn't have to be included because that is not how AMD intended for GPU compute to work with the HSA initiative. Overall I think AMD did close to the best they could with the performance metric after changing strategy.

I don't hold this against AMD at all. While in a perfect world, companies would thrive being completely straightforward and transparent, in the real world marketing works differently. Consumers generally don't do the DD to thoroughly research before they buy, and the competition is cutthroat. I would even go so far as saying, I'd support AMD's decision if they decided to implement Intel's misleading TDP rating system. They ought to, because i STILL see/hear tech news authors/retail sales reps stating AMD chips consume more power than intel equivalents. And the casual buyer that reads/hears this doesn't know any better. Sometimes in business, you have to match the competition's tactics.

5:37 Been always wondering about why AMD never released Steamroller/Excavator to the FX series for years, seemed like a mystery. Thanks for clearing it up.

I watched Ian's version first and it's interesting to compare the two perspectives. He suggested that it was mostly for AMD internal consumption, to give the engineers something to aim for. Oh, and you need to get a cat, Jim.

thanks Jim, I still miss my Phenom II 965 black Edition it lasted through the bulldozer episode.

You never fail to amaze me with these videos.... Im in college getting a SysAdmin degree and these video are awesome putting the industry into perspective.

It's important to remember they're also in game consoles, which do take heavy advantage of APUs and HSA.

HSA does have huge potential even still and i hope AMD doesn't give up on it and instead pushes for MS to utilize HSA at core OS level and allow certain tasks to run on GPU where normally they wouldn't.. ay running a task that pushes a lot of float on CPU, maybe needlessly it could see this and instead direct that code to the GPU to run it.. dont hv to look far for gains HSA will give: best example of this currently is iOS and A series chips.. A12 was first apple SoC to do this where depending on the code being run it could shift the task to GPU or CPU as needed without much if any developer needing to code for it to do this...but see huge gains in both performance and efficiency cuz of this.. i imagine mac os will also implement this as well now that moving to own SoCs there as well..

GloFo's 32nm was great, 5Ghz was a thing back then... The worst part of Bulldozer was their terrible high latency cache...

HSA is still alive and well, supported from Kaveri all the way to Renoir in hardware. It's part of ROCm for one...

I have an fx 8350, and let me tell ya, it sucks with old games, but modern games run quite well on it. I can bearly run Bioshock: Infinite on it, but Hellblade: Senuas Sacrifice runs on high settings and I get 50-60 fps (under DXVK/proton). I have an RX 750 + an FX 8350.

Thank you for doing this piece, the history empowers me in my career moving forward.

It's interesting that they made what could have been an internal business initiative public, since as you point out, there's little chance someone will buy a laptop on the basis of this. But I guess congrats to AMD for setting goals and trying to achieve something, efficiency in laptop is great for battery life and thermals after all.

I just keep wishing I had listened to myself years ago and put that $10,000 I had extra, into buying AMD stock at ~$3 share....

25:45 this is relative performance, so it should be plotted logarithmically. This would additionally emphasize the jump at the start. And this is what they did. They did the correct thing. (it's the same for stock prices)

Most people in that 2014 AMD marketing department are probably long gone Jim, let them be. Renoir is great competition for Intel and that's what really matters.

Just want to say I really appreciate your work. Thanks for putting so much in context.

The first thing that poped into my mind when hearing the 25x effciency is Todd Howard's "16 times the detail".

That was interesting for someone like me who recently joined the PC master race, AMD division, Zen department.

I'm still on a 4790 since 2015. Still goin strong.

Finally a mention of HSA! Very cursory treatment, but something!

I miss my 9590. It still did work in 2019. :(

Measuring relevant performance improvements over 6 years in hardware is tricky. I mean, what did/could they start with, CinebenchR10? It was an internal goal, so like any motivational- but- attainable goal you want to get it. I don't see anything to fault about achieving a goal from a low starting point. It was a real product (Kaveri), and making something better took real work. I would be interested in how the integer performance increased, but mostly academically.

Is there anyone else still rockin' an Intel i7-2600k ? I can't believe I'm still using this chip, but it does still work quite OK. Very glad I never bought one of those Bulldozer chips, otherwise my life would have been pretty different. Planning to upgrade from i7-2600k to AMD Ryzen Threadripper 3970X, once I sell/find some gold bars.

AMD didn't get lucky. They abandoned it like you said but found that they can achieve this goal using Zen 2 and TSMC when Ravenridge came out. That's why they brought the 25x20 back from the dead.

Thank you for sharing this iinformation.Very interesting.

I think this 25x20 thing was just an appeasement for the customers ... Or share holders? I mean Intel is doing something similar, year after year they are praising their new architecture, but actually is still just a beefed up Skylake. So I guess it's up to us (or to Jim) to see through all that marketing BS 😂👌

@AdoredTV That was a really great analysis, and I understand your narrative on the suspicion over changing the benchmark definition. I would have liked to see your final projection on the original benchmarks in Renoir, I don't recall seeing this in your analysis. For me though, as you've described it, it seems to make a lot more sense to me to pick benchmarks that are more relevant to today and the laptop energy efficiency and maximum performance potential of the platform. If they had just stuck with integer compute biased metrics, do you honestly think this is a good metric to calculate overall energy efficiency particularly when there is a good mixture of dedicate GPU, integrated IGPs to compete against in the current portable computer market? I think perhaps the better focus would have been the shortsightedness on their integer biased benchmarks and how they would not generate a full chip power load in future mobile platforms, in which case you could pick the values apart in the formula for not being relevant and "cheating" by not using the maximum power state of the chip which would have impacted the energy efficiency ratio of the equation. I am certainly no specialist when it comes to the work of chip manufacturing, but it would be difficult to make a good baseline equation predicting the future 6 years in development, when your share price had tanked to nothing, research and development budget drying up and stick to a single well defined approach for this period of time - as this video beautifully highlights, how much the AMD architecture has changed in this period. About the only thing that would annoy me about these claims and the changing definition of performance efficiency would be if AMD were not transparent about changes to the benchmark, their rationale with relation to their shift in their internal goal focus, Lisa Su took the helm around 2014 so it's unclear if these claims were made before or after the managerial changes. To me, without digging into the particulars, there seems to be good reason in terms of relevancy for changing it, and I can appreciate the general suspicions of tweaking statistics to suit AMD's own narrative, there have also been in the past two years many other cases in marketing statistic claims by both AMD and Intel which were outright ludicrous. We just need to look at the armada of Renoir chips making there way out into OEMs at the moment to see they have finally arrived with a solution truly capable of competing with good old Intel in the mobile computer segment - which at the end of the day, will mean faster efficiency gains and more price competitive hardware for consumers in the coming years. Edit: Also, isn't it true they jumped from HSA to generally calling these fusion chips APUs? As far as I can tell this was the evolution of the HSA concept into practical chips. Looking at the HSA Wiki page I could see the FPU in Renoir jumped from 128-bit interface to a 256-bit interface, which helps explain part of the huge improvement in the last iteration.

Thank god this was a "slow news" moment and you decided to make this video because I loved it. In a age where there is regulation/normalization/directives for almost anything, the marketing world seems that they can still do just anything they like, spread ill-information, rely on consumer misconception, hide/hinder important specifications information and sometimes even straight up lie (at least that's what it would be called outside the marketing world)...

Raven Ridge is quite a bit faster than Carrizo in the iGPU. No matter what 3Dmark says, the actual difference in games is MUCH bigger.

One big mistake in this. Bulldozer was a true 8 core; they * didn't lose the lawsuit.* They settled for 12 million which might sound like a lot, but is much less than the cost of a court case(heck, they would spend close to that during pre-trail discovery). The plaintiff knew he had no hope of winning from the start and was just hunting for a settlement. Also, the idea that Bulldozer has a "shared FPU" is widely misunderstood. Both cores are fully capable of doing floating point calculations at the same time and only function as a fused FPU when doing AVX. The whole idea comes from one bad slide and a reporter trying to make sense of why Bulldozer was so bad. Fun fact, Zen uses almost the exact same concept as Bulldozer, but it is a much better execution. TL;DR I am not a fan of Bulldozer, but there is no logical way to argue it isn't 8 cores. If you think they aren't full cores, you either got some bad information or don't understand the difference between multicore and multiprocessor chips.

Hello Jim Great video as always. I just have a small observation to make about the graph at 25:27. This observation goes against what you said, but is far from being important enough to change your conclusion. And the points you made on the video still stand. So, you took issue with the graph because it uses log scale instead of linear scale. And you basically said that AMD used log scale as a trick to make us think that there was a "decent progression all the way". In my opinion, log scale is the right scale to use, and there was no trickery from AMD on this part at least. Here is why: + When you look at the linear scale at 25:36. You can see that between 2014 and 2015 there was a x3.5 jump. And between 2019 and 2020, there was a x3 jump. But on the linear scale, it apears as if the x3.5 jump is much smaller than the x3 jump. Which is missleading in my opinion. + Even worse, the x10.88 jump between 2014 and 2019 appears on the linear graph to be smaller than the x3 jump between 2019 and 2020 (twice as small actually). The linear scale makes it seem as if AMD, between 2019 and 2020 (1 year), made twice as much progression as in between 2014 and 2019 (5 years). + If for example AMD consistently doubled performance every year. We would say the there was a "decent progression all the way". But the linear scale would show an exponential curve. The log scale would naturally show a straight line (slope of ln(2)) in this case. This is why I think log scale is the better choice of scale to show this kind of progression. As I said earlier, this does not make your conclusions wrong by any means.

Meh. Marketing is marketing. Never take it at face value, in particular when it makes claims so far into the future. As far as HSA, AMD has never been large enough or had enough clout to force a paradigm shift like that. HSA was dead the moment Intel stated they wouldn't support it. The irony is, now that AMD has dropped it, Intel appears to have fully embraced it with their Xe GPU.

i think it makes sense. perf efficiency is misslabeled but its a good metric. its like big gas engine which has a lot of power but takes a lot of fuel at idle too, vs small gas engine which idles very lighly on fuel but can't output much power at max. the metric is about having the potential of maximum power for rare tasks like compiling or rendering, while still having great efficiency at idle and low load which is what the vast majority of the time is spent doing

I know we are all die hard gamers with GPU's. APU's are the future of volume sales. Consoles, laptops, TV's, Fridges :P

Thanks again Jim for another excellent analysis, really enjoyed watching this. Of course, we know Intel are fanatics at re- imaging/presenting their data using changing metrics and unfortunately for us, their marketing department influence too frequently damages any chance that the data presented is honest and true. Some might argue this is fair since they have a product to sell, and, blatant and deliberate dishonesty is rampant in many companies marketing. Even if the data is good and speaks for itself many will still want to 'embellish' it. As for the comparison methods and tools used, I don't think it's unreasonable to continually examine the best ways to reflect progress made. This isn't uncommon - data sets in the scientific community are frequently re-analysed using new or different tools and in fact, it's critical that folk are open to this. Thinking about how best to do this should never be a one-off task and if it was, that would be unfair on the data, the tester and the reviewer since important conclusions might be missed. Equally important is thorough peer review to spot incorrect or dishonest presentation, which brings us back to here - AdoredTV - and is the primary reason I love this channel. Keep up the fantastic work Jim.

Sorry for jumping around in recent post. I was really responding to the last 3 vids that I couldn't quite understand the point of while being stuck in that should I or shouldn't I moment. I really shouldn't have. Thanks, apologies

The thing is, the vast majority of compute isnt integer anymore. javascript numbers, what 90% of the web uses are doubles, plenty of high level code just runs things as doubles to prevent bugs from casting. I do think its an insane waste of compute power to do that. but thats the reality.

21:58 'The vast majority of the time will be spent at this short idle' * looks at visual studio 2019 and its RoslynCodeAnalysisService that bogs down my entire machine * Yeah... nah...

To be fair to AMD: they didn't chose Kaveri as baseline to play the numbers. They just didn't have anything better at the time.

Wasn't expecting to see a dedicated video on this after hearing it be talked about during the podcast. I don't mind, though! Regarding the change in benchmarks used, though, I don't necessarily think a change is inherently wrong; AMD set the initial goals with the expectation that emphasis on INT operations would be the future, but when this didn't turn out to be the case, it stood to reason that the original set of requirements was no longer realistic for 'real-world' scenarios, and thus the benchmark should evolve to reflect that. We were just lucky that Zen 2 was such a success that AMD afford to represent their APU capabilities somewhat honestly.

That intro though... :D Thought my headphones were off

I still always wondered what a 28nm fx-cpu would have been like relative to fx-8350. We would never truly know, just speculate and estimate based on other hardware. Still using an FX-8350 to this very day. Its just enough paired with an rx580 and 16gb of ram in most title. Some stutters here and there, but still a "good enough" experience. With ever Zen Generation I always found myself saying wait until what the next gen had to offer, because I wanted to see a CPU that would at least match existing intel in single threaded performance, not caring as much about multi threaded performance. zen3 might finally be that point. Now its a matter of performance per dollar. I believe Intel will not be able to market themselves out of Zen 3 and will likely go AMD this time as well, but Intel is still on the table if AMD decides to get greedy and price higher.

Haha, yeah, even though I think a logarithmic straight-line graph is the correct one for this, the jump from 11x to 31x is silly, and only because Renoir has 8C in the APU (was that a requirement to meet this in the design stage perhaps?). OTOH it seems Zen 2 was a little late, hence Zen+, so that's why 2019's APU was only 11x - maybe they should have designed a 6C APU for 2019 as well (but it's only now they can do APU-gasm with Renoir, Van Gogh, etc, they didn't have the resources before). TBH I don't think there's anything stopping someone using their own mix of benchmarks and seeing if they get 25x or not.

0:00 HA! Love the new graphic.

AMD after a string of major wins that finally got it to be at parity with Intel started making some bold calculations and gmables that nearly bankrupted the company... HSA as far as the approach they were thinking is dead as a doornail, however it did live on somewhat but via proprietary and/or specialized means. Nvidia's cuda measure is basically a form of HSA on a limited scale app Acceleration)

I was shocked I actually received the claim $$ from the FX class action. Crazy it made it that far.

Finally a proper video explaining the difference between AMDs high end desktop FX line and APU line up. Couldn’t for the life of me make sense of it all before. Too many codenames and architectures. Recently purchased an Athlon x4 845 which was the last desktop based cpu based on Excavator that replaced an A4-6300. Big difference in single core performance, machine feels a lot more snappy.

I never want my CPU stuck on the same chip as my GPU, since I usually upgrade the GPU's graphics card two or three times more frequently than my CPU. Even now, I run all the diverse games I play (at 1920x1080) very well at "high" or "ultra" graphics settings on my 3.3GHZ i7-5820k (no overclocking) with an 8GB RX-480. Not "high end" anymore, but still cruises along playing most games at near 60FPS (average) while only using about 30% CPU. My eyes aren't fast enough to even see anything close to that frame rate, so I'm good. As long as I can continue to get performance like that, I really don't need to upgrade much of anything... I only upgrade hardware when I need to, and usually that's the graphics card and GPU much more often than the CPU! I remember the first time I saw the specs for Bulldozer, and even though I was impressed with the core count I was very puzzled over the odd organization of the floating point units. The following expert analysis that quickly came out showed clearly that this was an architecture to be avoided... The last times I replaced my motherboard and CPU was in 2013 and 2016 (after a liquid cooler leak wrecked the 2013 system, which is now air cooled even better than the new replacement liquid cooler could do). Still happy with what I've got so far, but expecting to upgrade to a reasonably priced Zen latest generation CPU the next time I need to!

so in summation; AMD realized they were on the wrong path, changed their ways which resulted in a better APU for all. i dont see the problem other than AMD coming out publicly and saying "we were stupid and we were wrong in the past" what do you want them to do? hoping AMD doesn't go back to their old ways but atm i am happy with their offerings, Renoir is pretty awesome compared to the competition, especially when comparing price to performance.

They were kinda right with the fact that the usual usage of a laptop processor is slowly getting closer to only run at idle and they made the math favor that in a major way, even in the beginning. I'd loved to see a graph with the old graph overlayed over the new one and how they compare plus an estimate on how the Zen parts would've fared in the old benchmarks, but that would've been a lot of work

That HSA actually sounds like a pretty rad idea. Yet another example where AMD's lake of market power on the software side limits the potential of the hardware. I understand Adobe finally implemented GPU-acceleration for AMD gpu's recently, some maybe time to get your Kaveri APU laptop out of storage? 😏

AL RIGHT SO, i haven't even started the video, I just wanna say I'm jumping from my chair with excitement!!!!!! THANK YOU JIM!!! AÑKSDJBQBE

I got my FX 8150 for a price intel couldn't compete at in 2014 for the perfs it gave me, I'm still using it, it still does everything I ask it to, which is what I ask of a computer. I build it to last me ~10 years, and with my needs, it'll do just that. I upgraded it to 16gb of memory and it's got a bit of a boost since I'm still on spinning drives (will pass to ssds when 1tb sata drives gets in my money range). 8gb just doesn't does it anymore for my needs. Thanks for this video ^^

My thuban held me over to Ryzen fortunately. :D

They changed the rules, but achieved it at the end Awesome video I would've liked seeing gpus executing CPU operations in conjunction when parallelization was possible, it's a very very good idea Unfortunately, the original purpose of Fusion and HSA are dead

The lesson for reviewers and media: As soon as such claims are made, demand the formula and list of benchmarks, demand they reveal the metric.

Good views, thanks. Ah looks like the marketing covered up some previous marketing and tech crap. But has it improved anything for the customer? Does it do what they claim it can do?

Didnt want to sleep anyway

At 17:45 ish, those are all quad core units. If you dealt with dual core units the numbers would be much, much worse. The lawsuit doesn't actually have any effect on the definition of a core. Compare the Phenom II to the FX and do the math with an FX having a 3 core or a 4 core vs the Phenom II and the FX looks respectable. That's simply not the case. FX was based on the same concept as Netburst, high, fast integer with smaller FPU and SIMD taking up the slack. The industry moved away from that as it didn't work as Intel had planned but AMD was committed at that point. That being said the original designs for the entire FX line were abandoned, tho some of the plans found themselves in the Zen architecture. Looking at the software used for energy efficiency, AMD used integer heavy apps to make the Bulldozer arch look as good as it could in performance, which also was the highest power draw they could use. When they changed over to Cinebench later it used a ton of FPU, which again, showed the strengths of Zen but again used a ton of power.

The HSA APU stuff could have been really cool, if they had something simulair to the CUDA API. This product could have been used for ML / Deeplearning. Where all the fp operations could have been pushed to the GPU

(having had an fx8320 from 2012-2015 before going to lga2011) I bought an fx8350 in 2018 for an upgrade to my fx backup system. It came with a wraith prism RGB and cost £50 new from amazon....I was very happy! until a lighting storm killed it and I had to return it and its beautiful cooler both of which probably just got binned :(

I have a question. You mentioned HSA was mostly retracted or tamped down. Didn't AMD just release a full hUMA system better known as the Playstation 5? I'm not sure how much better integrated it is compared to the PS4 but it is better, as I understand it. Sorry I'm unfamiliar with this particular flavor of acronym soup.

So can we all sue Intel for not providing even a single core CPU until the 486 DX? Seeing as none of their first three generation chips had a floating point unit at all.

Intel 4th generation was in 28nm process and from 5th generation to 11th generation in 14nm, Zen (1) Global Foundry 16nn, man that feels ancient!

Thank you great stuff 👍

The point tho was performance vs efficiency, they did show a steady gain in both the entire time. Look at the Cinebench scores. The energy efficiency curve is pretty stable, no argument, but the performance curve was growing the entire time, particularly after they jumped from the Bulldozer arch. That massive jump at the end was going from a quad core for the entire comparison to an eight core Zen2. Had they not used an eight core for the last CPU the performance wouldn't have increased as much, it would have been right below the 25x numbers, close enough to fudge. In all fairness, the core count used is precisely what they were selling for that power envelope in laptops. It would be simple to find out by measuring a quad core Renoir APU and doing the math ourselves.

It's funny that HSA and "The Future is Fusion" is discussed here. It's been on my mind since WWDC. Apple has likely solved the heterogeneous compute problem with Apple Silicon and its software frameworks. AMDs effort was futile in retrospect - they were never going to get Microsoft and developer backing when the products were so poor. In contrast, Apple is starting with mature, fast chips and will make the frameworks standard across the entire Apple ecosystem. Targeting the GPU, Neural Engine, AMX etc. is the default option, not a special case. I think heterogeneous compute is what will give Apple a significant lead over PC for many years. The fact they have the fastest CPU architecture (IPC) is just a bonus compared to this. PC-land could have been well ahead on this, but I guess such failures are expected from an industry that idly goes along with whatever Microsoft and Intel wants! (no fanboyism here, btw - just my observation!)

subbed for the accent stayed for the great content

Isn't that how logarithmic graph work? They make exponential growth look like linear.

did i miss something? they calculate energy efficiency by looking at max power workload scores ignore the actual energy used and take the idle energy instead? :'D

Kaveri was my first cpu (Apu), you shall not mock it

The Opteron series was perhaps one of the biggest casualties of the Bulldozer flop in my view. Because much like FX, those among the last "high performance" AMD CPUs to not have the PSP built in with support for modern instruction sets. Xeon platforms by contrast has had the ME built in since 2008. They could have had a much longer usable lifespan or updated I/O, but ultimately it didn't happen.

Regarding Bulldozer/Piledriver, I don't think the FPU is bottleneck in games, because 4 threads is easily enough as we've seen with i5 up until 2016 or so. I forced games to use each other core, so threads didnt have to fight for shared FPU, but that didn't help (because Windows already scheduled each other CPU thread, like it does with SMT) Can one thread on Bulldozer fully utilize FPU?

Context is important and Jim, let's face it you aren't crazy about marketing departments. It's the breakdown between salespeople and tech's... age old.

I just like the video before it's even started. *shrug* Jim needs more exposure as 99% his content is just stupendous.

It would turn out to be a good idea to state the typical use case blah-blah as "such as" for those benchmarks. They strictly speaking never said they'd use those specifically mentioned. Just something like it :P That being said, no matter how contrived or "polished" this is, it's the kind of stuff that makes investors feel all warm and fuzzy. Bean counters and corporate fat cats love this kind of BS. Let AMD give it to them, if that's what it takes to keep their stocks climbing. AMD is still only at a tenth of Intel's revenue, so we're absolutely not out of the woods yet.

AMD never made a proper toolchain for HSA was the problem. They never worked with like, Microsoft, and yes with Intel to get the shit done. You basically got language feature support through OpenCL and DXCompute, and that's about it. Everything was a hack, and you had to constantly be thinking about timing and futures and so on. What they should have done is added CPU instructions for tagging the GPU as a vector unit and loading that in at least L3 (preferably a optomized option for L1 or L2 or register) Cache, Called it 3dNOW2 or HSA+ or some shit so people could just make an #ifindef "3dnow2" and flip a compile switch, devs could then use templates to create various codepaths in their DLLs for APUs and CPUs, a kind of "virtual" AVX-512 or 1024 or whatever. At least then the Visual Studio, CLANG, and GCC guys could do some very generic optimizations. It could have started as a header and then made it's way into visual studio, gcc, and lower levels of OSes and hardware. Even if they were just virtual instructions in the microcode written in assembly that they would license to intel, would have been fine for the start, the actual hardware could have came later once the science and methodology was more mature. Even if they had the skill which is quite rare even in software engineering, who has the time to write custom microcode really other than console devs getting tens of millions for it. Relying on OpenCL and DXCompute runtime environments and Kernels killed it, high overhead and a pain to write for. Hell they could have licensed whatever the PS4 and Xbox1 guys were doing in this regard to get shit running on those consoles at decent framerates (i heard GPU got used like a PS3 SMU or SPE or whatever those CELL things were called very often) and gave us that, at least then maybe we'd get some ports that ran decent on APUs, do a joint venture with Sony or some shit, cause they did a lot of that stuff for PS4. Lets be honest, not many people are able to write their own assembly paths and do it well, it's also not obvious what the benefits and consequences are without access to the closed circuit design, and lastly very few people who can do it do it well, and millions in console dev gets pumped into this where as on PC everything is just done in C++ and C# especially but not only regular application development, AVX with or even without SMT "just works" even if you are using SSE or regular Floats with the right compiler flags, even without them or with "default" flags honestly because of hardware and microcode optimizations, the extra hardware still benefits older code, so that's always going to be popular. You can be sure Intel is going to be doing something like this with Xe processors, and AMD is going to have to license shit from them even though they will had the actual tech something like 10 years earlier it's going to be Intel that makes it a programming standard. Same reason Geforce Game Ready shit is everywhere and AMD's GPU Open shit is largely ignored, Intel and Nvidia have support hotlines where you can buy time from their devs, telling people to just "read the code and the hardware manual" is not enough, even Sony provided a bunch of crap in their toolchain for methodologies and styles people were used to from past eras. Even if I had 10 lifetimes I wouldn't want to waste one or two of them sifting through fucking white papers and assembly micro code just to make a product usable, especially as it would go pretty much ignored without corporate clout and end up in some vault or dead github repo, especially for a measly low end of 6 figures like most programmers make, that's AMD's job to make people want to buy and use their gadgets not mine.

We should contact Gamers Nexus to do independent testing of this. Steve looks like he enjoys this kind of testing. EDIT: Updating benchmarks as time goes on by itself is not bad. But i do agree more wider look might've been better. By the looks of it they did compromise on the age of the benchmarks, i bet even newer benchmarks would have skewed the results even further. After all, what matters is relative power performance *today*, not 5 years ago, neither 2-3 years from now (as software updates) for typical use case. For that CB R15 + 3DMark'11 might be a decentish compromise. SPEC Int missing tho is a bit ... ? but then again, what that measures is more likely to be used on EPyC chips rather than laptop chips. In any case, Gamers nexus or someone else could take more thorough look using old and new benchmarks how close AMD came to this. Zen 2 is brilliant never the less, does not change that fact, only question is did they achieve their target or not. 25x in 6 years tho is a TALL TALL order for ultimate performance, so measuring typical usage, idle etc. makes total sense. Measuring the average use case in other words. who uses their laptop CPU at 100% 24/7? No one. Not even servers tend to do that! Our servers typically hover around 90% CPU Idle as load is I/O.

They made HSA sound really good, yet nothing came of it. I wonder if it'll be picked up again and actually implemented now that Intel is getting better GPUs.

Usually I can agree, with even critics in this channel, but this seems just like nitpicking. Did they have to be using still the benchmarks like this: ru-vid.com/video/%D0%B2%D0%B8%D0%B4%D0%B5%D0%BE-wKMEA_byY78.html for their metrics after Intel bribes? NO, I don´t think so. All they did was to focus on real use cases which CB15 scores reflect really well(except some "Intel favored games"), CB20 even better. Why did they have to focus on INT based benchmarks after they realised the world is still relying on FP and they fixed their acrhitecture also to remedy the situation. Not like they did only manipulate the resulsts(like some other companies do) and still use same Bulldozer arch which performed bad, was powerhungry and hot. Now they came to the market with products which actually perform superb on every possible scenario, they are efficient, they blast competition away in Heterogenous workloads- remember they REDUCED CU-s on iGPU part BUT still outperform their last products and competition. And to reflect all those real world(really hate to use the term after some company ruined it, khmIntelkhm) gains they decided to chose the benchmark tools which actually reflect their incredible achievments instead of gimping all the effort on paper(benches from Intel parent company), some are still not happy with it? They set a goal and they achieved it, how does it matter if it was at the start of the run or they had a great end spurt? Doesn´t make any difference to me. I´m not an AMD fanboy btw, only now switched to Ryzen after years of using my old good Ivy Bridge. This kind of aimless lashing out, I would expect it from other "journalists". At least could have added some counter arguments which I pointed out.

I don't think a change in calculating their "performance efficiency" is automatically wrong. They made a bet on a unified compute architecture, built their equation around it, and compute simply didn't go that way. Changing that equation to reflect where CPU compute is now does makes sense. There might be some cherry picking going on here though. Either way, great break down. It makes me wonder how all performance and efficiency claims, generation to generation of CPUs/GPUs, stack up given how fluid companies calculations could be.

So what is Renoir's scores using SpecIntRate/Passmark/PCmark, and how does that hold up in their claim?

I was happy with FX-8370, Athlon X4 860K & 880K the latter was just bought so I get a new cooler! :P (FX-8370 & Ryzen 5 3600 is my current systems)

Was wondering when a new video would pop up. Unfortunately I just downloaded the 1962 Mutiny On The Bounty and a new Mel Gibson film ... Forces Of Nature? Something like that. Don't worry I'll be watching this video at around 3:30AM tomorrow morning! XD

long ago we discussed at amdzone forum future stuff. i got idea about sea of gpu cores, managed by fp system, so the cpu does not have typical fp units, but this thing instead. quite hard to implement, if possible. maybe future epyc will have..

lol. lipstick on a pig. Not heard that for so long.

"I don't really feel strongly about this subject at all" Yet without this video, AMD would probably have gotten away with no one noticing.

INB4 "Intel fanboy" accusations are thrown at Jim :D

so, they failed at HSA, but why specifically? it sounds actually like a brilliant idea to save on fpus and offload this to a way more suitable and much more performant component. did they leave the utilization of this to the devs like with primitive shaders on vega, instead of bunching the cpu's fpu and the igp behind an intransparent api, or why didn't it take root in the common programs?