Kicking off a busy day of product announcements and updates for AMD’s data center business group, this morning AMD is finally announcing their long-awaited high density “Bergamo” server CPUs. Based on AMD’s density-optimized Zen 4c architecture, the new EPYC 97×4 chips offer up to 128 CPU cores, 32 more cores than AMD’s current-generation flagship EPYC 9004 “Genoa” chips. According to AMD, the new EPYC processors are shipping now, though we’re still awaiting further details about practical availability.
There is so much competition in the processor space at the moment – it’s just great. Few of us will ever get to use or even see these processors, but eventually, technologies developed for the very high end of the today will make their way down to the attainable end of tomorrow.
Thom Holwerda,
Honestly I’d like to see more architectures in the race, but I agree at least those who are in it are pushing the bounds of what’s possible.
It is impressive, but I’m not sure that this type of CPU core scaling will ever be all that beneficial for normal consumer applications. Even today most of us have more cores than needed and the software doesn’t make good use of them. Huge core counts make a lot of sense in the datacenter where you might be running hundreds or thousands of processes…but this is unusual for consumers. Even if it were “attainable”, 128 cores just seems overkill for the masses.
While many consumer applications can benefit from huge amounts of parallelism, like rendering and games in particular,, it’s just that these types of apps are typically offloaded to a GPU, which does the work much more efficiently than CPU cores. So I predict the GPU innovation will continue to be the main target for consumer parallelism.
GPU innovation has sadly become bitmining innovations. As long as people keep buying the tulips, I can’t see that changing
Adurbe,
I am more optimistic than this. Game programmers have largely focused GPU power on graphics, which honestly have become extremely marginal, but there are still new interesting applications that haven’t been fleshed out like AI, full world simulation, realtime content creation, etc. These can benefit from highly parallelized GPUs for even better character development and immersion beyond graphics.
I think there’s tons of opportunity to make AI characters with realistic intelligent behaviors.
Yeah, I’m obsessed with TH’s price per FPS column!
You might be right with AI, we have games like no man’s sky. And hopefully the future will expand that kind of thing. https://www.tomshardware.com/news/gpu-pricing-index
Adurbe,
Thanks for the link. I agree that is an interesting metric to compile the data into. Many of the older cards have much better performance per $ given the lower prices of last gen hardware. But my main concern with them is having that less memory means less future proofing.
Off Topic: Have you noticed that tomshardware’s 600px fixed width content style sheet is egregiously bad in this day and age? The tables are completely cropped off despite there being 200% more whitespace available around them.
https://i.postimg.cc/KYMYhc92/Screen-Shot.png
I sent them an email about this, they might fix it if they care about readability, but probably not, haha.
Huh? GPU innovation has kept up with AI and lots of graphics advancements, mining has left the GPU space for the most part. Miners are moving on to ASICs.
No, it just moved sideways.
There are now “tensor” cores, which are simpler arithmetic cores with lower precision, 32, 16 or even 8 bits math.
There are ray tracing / RT cores, which also has a dedicated function.
We are in the ebbs flows of the hardware design. We had generic purpose CPU + VGA frame-buffers. Replaced by fixed purpose Open GL accelerators. Those were again replaced by generic Open CL / CUDA cores. Now there is another iteration of specific purpose sub-units.
There are also innovations on the RAM / cache front, with things like “infinity cache”, HBMs, and better APUs which can actually be used for modern gaming.
Nothing wrong with this progress.
Just a minor nitpick, the OpenCL/CUDA “generic” cores were just shader cores from the OpenGL/DirectX pipelines that were exposed to the programmer. The main difference was that they added write back support for the graphics pipelines, which traditionally just dumped the data onto the frame buffer and not back onto the host bus. In the old SGI days, they used some of the graphics pipelines as Compute accelerators as well.
Similarly the RT cores are used for graphics stuff as well, like image upscaling.
I always chuckle when people, out of the loop, complain about the lack of progress in this field. Having no clue the amount of expertise and development that goes into these devices.
Also, NVIDIA had a major stockage problem, because mining demand dropped tremendously and scalpers were stuck with tons of boards nobody is willing to pay a premium for anymore. Supply chain issues are also being ironed out, and demand is dropping (now that the COVID-working-from-home boom is over). So it’s a great time to upgrade, if that is what someone needs.
javiercero1,
You are right.
I remember the original GPGPU scene where they used shaders with off-screen buffers, which were later manually transferred back to the main CPU’s memory.
I did not know about the previous usages, but I am sure even since Atari days, people would have “misused” (or rather “appropriated”) graphics processors for other work.
Well, graphics has always been compute dense algorithms, and most graphics kernels are basically lots and lots of matrix multiplications. Most early high end graphics boards used dedicated DSP or RISC cpus for the geometry and shaders. Both raster and raytracing graphics pipelines tend to be very data parallel, which again has a lot of overlap with lots of scientific codes, specially linear algebra, which also have similar parallelism. So people have been repurposing graphics HW for general scientific computing for ages (and vice versa).
To be fair with the G80, NVIDIA made the pipelines far more visible to the programmers than previous generations, and with CUDA they struck gold since there had always been a pent up demand to access all those FLOPs sitting on the graphics pipelines.
I have to agree, CPUs passed “good enough” quite a long time ago, heck even my old AMD FX based gaming rigs I passed down to the grandkids are having no issue playing games like Overwatch and War Thunder with nothing more than upgrading from the old R9 380x that was in them to RX580s.
And I seriously doubt I’ll be upgrading for several years as thanks to GPU acceleration my 5700x spends most of the time with idle threads even when rendering video so I’ll be more likely to just throw another 32GB of RAM in the box as that would do more good than adding more cores at this point.
For 30 years I was the “gotta have the new PC hotness” guy but once I got past 4c/4t? I honestly can’t feel much of a difference anymore, I just can’t seem to justify going for even more cores when I can’t find enough work to keep the 8 I have now fully fed.
Every so often I look at prices for the previous generation and wonder if I could self build a Linux workstation. It could either be a beast, or just a big fan haha.
While much of the “industry” has moved, GPU are still suffering from shortages and inflated prices as a result of the rigs churning away. Manufacturers even have official mining specialist firmware to encourage it. As far as I’m aware this hasn’t been the case with AI, but please point me if I’ve missed it. As I see it right now, the market is still distorted by mining. A bitcoin miner is buying many many cards, I’m buying 1. I can’t blame AMD and others prioritising them over me.
You’re way out of the loop. There is literally a surplus of GPUs right now, both AMD and NVIDIA reported huge inventories the past quarters. And most miners have moved over to ASICs.
Actually GPUs have never been cheaper, hell my grandson just threw an RX6600 in my old AMD FX box I gave him because he could get one used for just $140 USD which was only $40 more than the RX580s. Sure if you want Nvidia you’ll pay out the ying yang but Nvidia has made it clear they have zero cares about the consumer market as they went all in on mining and now AI so you’d frankly have to be nuts to buy Nvidia at this point, especially with all the reports of broken traces and burning power connectors.
BTW as for your Linux workstation? Go buy an aliexpress Xeon special, you can get a dual socket 28c/56t with 64Gb of RAM and board for just $178 USD, that would make one hell of a workstation.
I did this, actually, My workstation is a dual-socket with two Xeons, parts all from AliExpress. My favourite thing is the BIOS, which is so loaded with features I’m pretty much afraid to even load it up. It’s pure insanity, all undocumented.
Can you provide a link, or some search terms? I’m having trouble finding anything like that with some quick AliExpress searches.
atrocia,
I’ve never bought from aliexpress. I searched briefly and found a lot of a la carte xeon parts but not complete builds. I figured bassbeast was talking about a DIY build. I haven’t bought anything lately, but in the past I’ve got some really great bargains for old servers. The problem is they’re EOL with no support, but with x86 servers compatibility is usually pretty good. Server equipment is also built to really high standards so I don’t worry too about buying used.
I understand – I meant that I priced some components, and I couldn’t find anything close to the specs he gave for that price.
I currently use an HP Z440 with an E5-1620 and 32GB of RAM, bought used for a good price off eBay . It’s not the most powerful machine around, but it’s very well built, very quiet, and it makes a very satisfying workstation for my higher performance computing needs. The only downside is that some things are proprietary, and when setting it up I had to wrestle with some HP annoyances.
However there’s a problem if one does 3d rendering with the Daz Studio software, as that requires Nvidia to render in Iray. If you use that, you are stuck with needing an Nvidia GPU at present.