Ryzen 3rd Generation Threadripper CPU appears online

[WYP]

The death of Threadripper has been exaggerated. Zen 2 is coming.



Read more about AMD's Ryzen 3rd Generation Threadripper processors.

 

[Kei]

I'm looking forward to this. It's been a little bit depressing seeing 3900X's coming so close to 1950x performance and it makes my 1920x look a bit ropy. Thankfully, I went for threadripper for the pcie lanes more than the cores. That said, it still hurts a little to think that only 2 years down the road it's been spanked by mainstream offerings.

 

[grec]

I wonder what the layout would be with these now that the number of dies no longer impacts IO, 8x(1+1) dies for the maximum L3 cache & heat distribution? 2x(4+4) to maintain minimum 4/8c latency? Or 4x(2+2) as a compromise? Or maybe even asymmetric 8x(2+0) or 4x(4+0) cores if there's a lot of spares going. If I had to bet I'd probably guess 4x(2+2) myself, more cache than mainstream, wider distribution for higher clocks, but without sacrificing too much inter-cache latency.

 

[demonking]

With AMD offering a 16-core Ryzen 9 3950X on AM4, it seems unlikely that the company would ship another 16-core model on their TR4 socket

more than possible though. Intel did the same thing a few years back if I remember rightly on their X79 platform (i7 3820???). You may not need 32+ threads but you may need the PCI-e Lanes and im sure the quad channel memory is going to be a bonus too?

 

[grec]

Yeah that's what I was thinking, it's why first gen thread ripper had an 8-core model too, it's particularly useful for all the professional software that has per-core licensing.

 

[NeverBackDown]

I can also see a 16 core. Though what it will cost is the question. At that point you have to weigh all the chipset benefit cost over normal 3950x and would it actually be worth more especially if the boards are super expensive?

Just needs to have a clear and defined gap if they release a 16C TR.

 

[grec]

Yeah for a few years now Windows Server for example has been licensed to a minimum of 16 cores per server so if you have any more than that you need to buy extra license packs. It's £200 per 4 extra cores, so for a 24-core model you'd be looking at about £400 more just in your OS costs if you had to use Windows Server (due to certain software or service requirements). If you needed all that IO or memory bandwidth for your tasks but not so much the cores then you'd be stuck between a rock and a hard place here, it'd be a major oversight imo.

 

[fadingfool]

I'm looking forward to this. It's been a little bit depressing seeing 3900X's coming so close to 1950x performance and it makes my 1920x look a bit ropy. Thankfully, I went for threadripper for the pcie lanes more than the cores. That said, it still hurts a little to think that only 2 years down the road it's been spanked by mainstream offerings.

Spare a thought for us lot running i7-7820X - 28 PCIE lanes isn't much of advantage and 8 cores is old hat.....

 

[WYP]

Yeah that's what I was thinking, it's why first gen thread ripper had an 8-core model too, it's particularly useful for all the professional software that has per-core licensing.

Yeah, but that model was discontinued pretty quickly. For Threadripper first gen the best seller was the 16 core model. More people want more threads than people who want more PCIe, and a lot of the more PCIe crowd also want more cores.

There was a reason why AMD didn't release a Ryzen 2nd Gen Threadripper 8-core.

I can also see a 16 core. Though what it will cost is the question. At that point you have to weigh all the chipset benefit cost over normal 3950x and would it actually be worth more especially if the boards are super expensive?

Just needs to have a clear and defined gap if they release a 16C TR.

Yeah, the Threadripper might benefit from more memory bandwidth in some use cases and will probably come with a higher power limit. TBH it's hard to know what AMD would do. It competing with itself.

I'm looking forward to this. It's been a little bit depressing seeing 3900X's coming so close to 1950x performance and it makes my 1920x look a bit ropy. Thankfully, I went for threadripper for the pcie lanes more than the cores. That said, it still hurts a little to think that only 2 years down the road it's been spanked by mainstream offerings.

That's the cost of fast progress. At least your current TR4 board should be able to support 3rd Gen Threadripper, that will be a big upgrade if you choose that path. None of the uneven memory/IO latency etc and NUMA issues.

 

[grec]

Yeah, but that model was discontinued pretty quickly. For Threadripper first gen the best seller was the 16 core model. More people want more threads than people who want more PCIe, and a lot of the more PCIe crowd also want more cores.

There was a reason why AMD didn't release a Ryzen 2nd Gen Threadripper 8-core.

But also because most licensing agreements for software since 2017 start at 16-cores anyway. Going below that rarely has a consequence on costs now, but going above that now can quickly cost thousands of pounds extra in software licensing even for just 8 more cores(Which for anything but a 90%> parallel workload wouldn't be very worthwhile). It'd essentially be blocking people with Zen1 16-core models from upgrading at all without spending in some cases £2k+ total extra. The financial sector for instance has many of these ~80-90% parallel workloads that do great with 10-18 core models and benefit hugely from IPC gains, but would see little value in a move to 24 or more cores, and would be completely nonviable on a mainstream platform due to the lack of proper ECC support, and the fact these spreadsheet crunching like workloads are super data/memory & IO intensive.

While Intel are making countless dedicated SKU's for these sectors and workloads due to how much of the HEDT market they account for, can AMD really afford to just ignore them?

 

[WYP]

But also because most licensing agreements for software since 2017 start at 16-cores anyway. Going below that rarely has a consequence on costs now, but going above that now can quickly cost thousands of pounds extra in software licensing even for just 8 more cores(Which for anything but a 90%> parallel workload wouldn't be very worthwhile). It'd essentially be blocking people with Zen1 16-core models from upgrading at all without spending in some cases £2k+ total extra. The financial sector for instance has many of these ~80-90% parallel workloads that do great with 10-18 core models and benefit hugely from IPC gains, but would see little value in a move to 24 or more cores, and would be completely nonviable on a mainstream platform due to the lack of proper ECC support, and the fact these spreadsheet crunching like workloads are super data/memory & IO intensive.

While Intel are making countless dedicated SKU's for these sectors and workloads due to how much of the HEDT market they account for, can AMD really afford to just ignore them?

Don't get me wrong, I'm not saying that there is no point in a 16-core Threadripper. All I'm saying is that the 8-core Threadripper didn't work out well for AMD when compared to their higher-end SKUs.

All I am saying is that AMD's early Threadripper sales data was skewed to their higher-end models, but TBH I don't think the same applied to Threadripper 2 and the 24/32 core WX series.

AMD doesn't have the same binning power as Intel. They don't produce anywhere near as many processors. AMD needs good general-purpose SKUs, not hyper-specific SKUs to compete in every possible market.