AMD releases a new flagship CPU for consumer market with 128M of L3 Cache
techpowerup.com
techpowerup.com
https://www.anandtech.com/show/18747/the-amd-ryzen-9-7950x3d... ("CPU Benchmark Performance: Simulation")
Worst is Factorio, which is down -47% against the 5800X3D (a previous-generation CPU that also uses stacked L3 cache), on the benchmark "10K trains". Wonder what's going on there?
- "We would have expected higher performance in our Factorio benchmark, as the Ryzen 7 5800X3D and its 3D V-Cache did yield some impressive gains. This is likely due to the AMD PPM Provisioning and 3D V-Cache driver opting for frequency over cache when running this benchmark."
"As we can see in our Factorio benchmark, we saw massive gains of over 100% when forcing the Ryzen 9 7950X3D to use the CCD with the 3D V-Cache as opposed to letting AMD's PPM Provisioning and 3D V-Cache Optimizer drivers do their thing automatically."
https://www.anandtech.com/show/18747/the-amd-ryzen-9-7950x3d... ("Conclusion")
The interesting one will be the 7800X3D.
From time to time, it can retry the test, because software frequently changes what it spends time doing.
I wouldn't be too worried about this personally, as configuring the few really sensitive apps with ProcessLasso or Aniancy is pretty easy.
Factorio players don't necessarily know anything about computers. And you don't know if they bought this CPU or if it came with their fancy prebuilt.
A properties page for the executable with a checkbox for “favour stacked cache cores” or some such.
There is currently no consumer model with 2 CCDs and X3D on both of them.
Best L3 per core outside of the X3Ds it the 7600 or or 7900 CPUs that come with the full size L3 (32mb PER ccd) but only 6 cores per CCD (instead of 8).
Similar problem to big.little but less differentiation between the cores. It's big.big. Kind of a mess outside of the primary use case that is games on windows with a static tuned scheduler.
Also linking my favorite tech reviewers benchmark they do some great Linux/tech videos also! https://www.youtube.com/watch?v=jNMcdIkef_c
https://www.phoronix.com/review/amd-ryzen9-7950x3d-linux
"AMD Ryzen 9 7950X3D Desktop CPU Review Roundup"
https://videocardz.com/150677/amd-ryzen-9-7950x3d-desktop-cp...
On the one hand side Gamers Nexus is making a point [0] (there are on/off performance charts later in the video) why it is super important to install the latest windows drivers + special windows gaming software to make use of the heterogeneous architecture. On the other hand side the linux tests are much better then I would expected since it can not know out of the box what core to pick. Maybe the 50/50 pick is good enough for how the charts look...
It's sad that performance benchmarks have made market conditions where the gain of 5% performance is somehow worth 20%-30% extra power consumption (and the noise that follows from that)
Test yourself, but I am definitely going to leave it enabled.
I could barely be stable at -15 all core on a 5900X at zero clock offset.
Worth noting the PPT is higher than the TDP by default - a "120W" AMD chip is actually allowed to boost to 168W continuously, obviously they prefer to advertise the former and have people making comparisons between "120W" AMD and Intel chips running at full boost PPT.
X3D chips are a cut above the rest of the market in terms of efficiency (although this mostly springs from the clock restrictions imposed by the v-cache CCD, which is undesirable and will be removed/mitigated once the technology permits) but in the rest of the lineup this slides them much closer to the Intel line than AMD would prefer people to think about. A 170W TDP 7950X is actually 220W, a 125W 7800X is actually 170W. People reflexively think in terms of AMD's marketing TDPs when making comparisons and don't account for the fact that the chips are allowed by the spec to boost way above that.
And it's been incredibly successful - every single AMD thread there's someone talking about their nice efficient 65W processor and lol unless you turned off the boost (which is separate from just enabling eco!) then no, it's not 65W. Still good, still usually more efficient than Intel, but AMD gave themselves an instant 30% edge on paper with the power of clever marketing and that's always been a significant amount, enough to swing purchases.
That part is not something that gaming motherboards do automatically, and I've never seen one that will enable multicore enhancement-style tuning by default. If you enable overclocking yourself (and enabling XMP/DOCP is considered overclocking since this usually increases VSOC or VCCSA to allow the memory controller to run faster than spec) then all bets are off but by default no, they shouldn't be, just the terminology difference that springs from AMD's marketing trickery around TDP vs PPT. And hey if people end up blaming partners for AMD's spec... that's just good marketing work.
PBO is the setting where you'd see power limits increased and I've never seen that on by default, it's always hidden behind a "THIS VOIDS YOUR WARRANTY AND KILLS YOUR DOG" disclaimer in a separate panel.
Yeah, what about games on Linux? I'm pretty sure it will be scheduled all over the place.
Also I doubt that scheduler is smart enough to detect it dynamically. Do they just use a list of known applications for that?
Code compilation tests are tied with 7950X.
Apps : 79w vs 169w
Games : 56w vs 143w
No, because the best silicon is reserved for the highest performance parts.
https://www.techpowerup.com/review/ryzen-7800x3d-performance...
Wow. I don't needed such CPU now, but if I needed a low-noise/silent but performant PC I would definitely look at this one.
> A 700 MHz frequency difference may sound like lot first, but as the data in our full Ryzen 9 7950X3D review shows today, the frequencies significantly above 5 GHz are only reachable on the second CCD, the one that does not have the cache, so it can run higher voltages, which lets it achieve those clock speeds.
I'd like them just add a toggle to let me select who use the gpu first or evenly instead. And their guesses never went right.
System > Display > Graphics > Default Graphics Settings
It would be interesting to see what kind of performance could be achieved with PERFECT scheduling. That would probably require the CPU to do process scheduling via specialized circuitry though as opposed to (or in conjunction with) the OS.
CPU's usually have a cache-as-ram mode that can enable this.
Does this mean that systems where you can hot swap all the ram at once will be possible?
I used to see SQL Server 7 databases running on NT4 servers with 64 MB of memory! This cache is double that.
Of course in some objective sense I have to recognize that the workloads were sized to fit…
I can also foresee an entire industry popping up to do this on purpose in the name of “compatibility”. I’m sure the marketing will talk about the efficiency of being able to run “true” IE6 on top of Chrome in 2030.
It’s also just a matter of time until we see Docker images for running Windows 95 apps in the browser via WINE on top of a user-mode Linux kernel emulator.
Check back here in 7 years to see which of the above prophecies materialised.
https://git.furworks.de/coreboot-mirror/coreboot/commit/a245...
It still seems a bit far off now, but come on, in 5 years or less we'll see 1GB, that gigabyte level is a huge implicit marketing number.
Might be a harbinger of a new movement in computing: all-cache computing, and reworking OSs and slimming down libraries and other things to fit into cache. It might be the only shot at getting a slimdown revolution going before moore's law really stalls in a couple decades.
- Jeremy Clarkson
Some do make them bigger and bigger, like the whole-die cerebras AI chip.
SRAM doesn't need to be refreshed constantly like DRAM, so SRAM will use LESS power when not switching. We think it's different because something like L1 cache is CONSTANTLY switching. It should be noted that the power cost per GB still isn't worse than DRAM, but there's way more data being transferred.
Once you reach L3/L4 cache, it's not switching nearly as often and the actual total power becomes way less than sending data over from RAM. This is one reason why AMD (and now Nvidia) saved so much power with Infinity Cache on RDNA 2.
The big issue with SRAM is that it requires 6 transistors per bit while DRAM requires 1 bit plus 1 capacitor, so capacity per mm2 of die area is much worse. That capacitor keeps thing cheap, but is also the reason why individual DRAM cells can only operate at 300-500MHz while SRAM can easily operate into the many GHz range.
I would have thought reading/writing SRAM is much more expensive than idling it not because of the cost of toggling the flip-flop, but all the cache coherency stuff associated with making a change in multicore.
If you do a full reticle chip and use most of it for memory you can get about 1GB of SRAM. But that would also be extremely expensive.
TSMC N5 is 0.021um per SRAM cell in high-density configuration (still going to be 5-10x faster switching than DRAM). That's about 47.62 bits/um. With 1,000,000 um2 per mm2. EUV reticle limit is 858mm2.
That amounts to 40,857,142,857 bits or about 4.756GB. That's almost exactly 250 dies per wafer. TSMC N5 wafers cost $17,000 or roughly $68 per chip (you can get very close to 100% yields on this kind of very basic chip). If we use part of the chip for internal controllers and interconnects, it would be about $70 per 2-4GB.
The highest-clocked DDR5 RAM I could find with minimal searching was $360 for 32GB.
The SRAM version would cost $560 at the cheapest for just the 32GB RAM not counting the RAM controller chips and all the packaging costs. With profits and everything included, your final price tag would likely be around $750 if you could produce in bulk and $1000 if you could not.
On the plus side, worrying about stuff like CAS latency would basically be a thing of the past with the bottleneck once again becoming the speed over the wire. Power consumption would also be lower without the need to constantly refresh. I doubt that anyone wants to pay that much money for disproportionally marginal increase in performance, but looking at all the people forking over thousands to scalpers for GPUs, maybe I'm wrong.
You might want to check your maths on that. It's more like 30 maximum reticle size dies per wafer IIRC. Just the area based limit is pi * 150^2 / 858 = 82, but you obviously lose a load because dies are rectangular and wafers are circular.
Cache uses 6 transistors per bit while normal RAM uses 1 transistor and 1 capacitor. You can buy a LOT more RAM per dollar than you could buy cache. This has been further exacerbated by poor SRAM scaling with each new fabrication node. It's usually 20-30% instead of 60-80% for other transistors and is actually 0% for TSMC's upcoming N3 node. AMD chose to fab their RDNA 3 GPU die on 5nm, but their SRAM dies on 6nm because the cost savings were much bigger than the effect of a slightly larger die.
DRAM typically has a maximum cycle speed of around 300-500MHz due to the capacitors having more restrictive charge/discharge rates (when you see something like 6400MT/s, that's not random access speed, but is instead the aggregate speed of slowly reading thousands of cells in parallel at one time into a cache then quickly sending the results over the wire).
Once you move away from bulk storage into high-performance caches, you have to discuss speed, latency, and associativity. As cache gets bigger, keeping lookup times low requires more and more transistors to control the cache and keep it coherent with both the CPU/threads and main memory. If you want to make that controller smaller, it will mean lowering clockspeeds, altering latencies, changing associativity, etc. At a physical level, faster caches will require higher-performance transistors which require more die space and power per transistor too.
Sometimes, the tradeoffs aren't what you might expect. A great example is AMD moving from 64kb of I-cache in Zen1 to 32kb I-cache in Zen 2, 3, and 4 because they found the tighter latencies were more important in their uarch than the higher hit rate.
There's a further constraint for AMD's 3D chip where the cache reduces heat transfer forcing lower CPU clockspeeds. Unfortunately there's not much room on the substrate to add the cache beside the CPU. Infinity Cache on RDNA3 also uses a significant portion of it's die space for the super-fast interconnect with the GPU die. I suspect a similar thing happens here too.
For example, I'd quite like my disk encryption key to never end up in RAM - it's much harder to steal from inside a running CPU...
The fact that most people leave their computers in sleep mode means that to a determined attacker, disk encryption is bypass able by reading the key from RAM.
P.S: the table in TFA lists the 7700 as having the same 105W as the 7700X. I think that's a slight mistake and the 7700 is 65W max listed TDP instead. FWIW I'm considering a 7700 for I'm not ultra happy with my 3700X (one of the reason being it has no iGPU and I could find a passive GPU with a NVidia chipset, forcing me to use proprietary drivers in Linux and it's a major PITA for if I hack on ther kernel I then need to apply the NVidia patches every time)
[1]: https://www.amd.com/en/products/apu/amd-ryzen-9-7950x3d [2]: https://chipsandcheese.com/2023/01/05/amds-zen-4-part-3-syst...
The 3 die chips like this are probably not the best choice for passive cooling, as they burn a lot of power on interchip communication.
Initial linux support is buggy, and you basically have to run a bleeding edge kernel.
Assuming your motherboard has video outputs it'll be much cheaper to stay on AM4 and throw in a 5700G
I'll probably build something around an Asus B650-Plus mobo: it's got USB-C stock, 2.5 Gbit/s ethernet, DDR5 (I know it's pricier and boots slower, but it's faster too), etc.
I'll reuse one of the NVMe .2 SSD I've got laying around so the whole build should set me back 1 K EUR (with 32 GB of DDR5), which is quite okay.
A desktop or IPMI? Sure, it has one.
An actual game? It may as well not. It won't compete with the iGPUs on Intel CPUs or even AMD's APUs.
I hope that HEDT isn't dead and it might be at least barely affordable (e.g. 32 cores for less than $2K as it was with 39xx lineup)
https://pcpartpicker.com/products/memory/#E=1,11
Filter for the memory amount and prices that work for you.
Not that it won't sell, but still these price points for something still called consumer start to be high. You could get a whole portable PC for same money.
The first Core i7 extremes (which fill roughly the same product niche) were ~$1000 like over 10 years ago (https://www.tomshardware.com/reviews/core-i7-990x-extreme-ed...)
This is a flagship CPU part. Consumer is a broad segment. Even the peasants like shiny things every now and again.
You could get something that would be imperceptibly-slower for most workloads and fits into the same socket for less than $300.
You can have absolutely fine 8c16t CPU which will handle anything you can possibly throw at it for 230$ from Amazon (https://www.amazon.com/AMD-Ryzen-5800X-16-Thread-Processor/d...).
GPU pricing seems a bit more insane, given a lot of the demand has actually died down at this point, and the mfgs are definitely restricting production to keep prices (and margins) up. It's kind of sad there really isn't anything comparatively good on the lower pricing end ($100-300) where historically you'd see around 60-70% the performance well under half the price.
> With the cuts, the wholesale cost of a 733-MHz Pentium III drops 3 percent, from $776 to $754, while the 700-MHz version drops 3 percent from $754 to $733 in volume.
https://www.cnet.com/culture/intel-cuts-pentium-iii-prices-1...
$754 in today's dollars is over $1250.
Cutting edge processors are cheaper than they were 20 years ago.
Today 750$ for a flagship CPU that buys you the best possible compute in the consumer space is not a bad deal at all. It's never been a better time for X86 CPU shopping. You most likely don't need to buy the most expensive flagship CPU. There's great options for every budget.
The GPU market on the other hand is completely fucked. In 2007 you could get a 8800 GTS for 360$ (520$ today) giving you near flagship performance.
I feel like AMD and Intel are competing in the CPU space while AMD and Nvidia entered in a gentlemen's agreement to control the GPU market.
https://www.redhill.net.au/iu.php is a good read down memory lane about those things, even as far back as the 286 you had it. Sometimes very rarely and almost as a mistake the "best bang for the buck" processor AND the "fastest in the world" would overlap.
Not with the same performance you cant...
Not even close
>>The inflation has really hit the consumer market... 700$ for CPU and over thousand for GPU...
$700 for the Top of the line CPU is actually low, at the highest of intel monopoly period they were charging well in excess of $1000 (closer to $1700 if I remember) for their I7 K line enthusiast PC's.. Even today $700 is cheap compared to the Threadripper line AMD seems to have given up on.
GPU prices have gone insane though with mid-range having somehow moved up to $400-$500 where high-end was a decade ago.
I also think another factor in the rise of GPU prices is how quickly things have been moving on the display front. Not only did we go more or less directly from 1080p to 4K, thereby quadrupling resolution, now we're not shooting for 30 or even 60 fps minimums anymore. We want 1% lows to be above 120, 144, or even 240fps so that our high refresh rate monitors are properly fed. Oh, and we don't want to turn down image quality while doing it. Now throw in VR, where high-end headsets can be 4K or more per eye, and you've got a lot of pixels to push. So top-end GPUs have to be these behemoths with several kg of cooling hanging off of them, because they are now the biggest power consumer in the system. Yet they still have to fit into the real estate allotted to just a few peripheral slots, usually in an orientation that's not optimal for airflow. It's quite the engineering challenge, and I kind of see why GPUs at every level have gotten so expensive, since they're all priced relative to the top range.
i5 12400f 6c/12t is 155 eur.
You can get 8c/16c for a bit more, 5700x for 190eur.
Value DDR4 16/32gb kit and a reasonable basic mobo will give you an amazing system.
The only thing which still is kind of pricey is GPUs.
This 700eur cpu is the very upper end of consumer market.
5800x3D 8c/16c for 300eur is the most I would spend personally.
This is the problem of late stage capitalism, where both companies are owned by the same large investment funds and they use their power to ensure neither company could eat too much profits of another.
When AMD was falling behind, they seem to have been "allowed" to do the Ryzen and it's back again to "elephant race".
Currently there is a paradox, that if you want to buy high quality and performant laptop and affordable laptop, Apple is very much the only option. No competition can come close. Sort of they got caught with their pants down as they stopped innovating just like AMD and Intel.