AMD's 256-core Epyc 9996 'Venice' claims up to a 3.4x jump over Intel Xeon

bhouston 19 hours ago | parent | next [-]

512 threads and 1GB of on chip cache. Love it!

I like it too, but I am more interested in the EPYC CPUs that are cheap enough to be affordable by individuals and small businesses.

For Zen 5, only a month ago AMD has finally launched the EPYC 8005 series "Sorano" CPUs, which use a smaller socket (with 384-bit memory interface and 104 PCIe lanes) and which includes some models that have a performance per dollar competitive with the 9950X champion (unlike the big EPYCs, which have a much lower performance per dollar, which is compensated only for big companies, where such big CPUs enable economies, by offering a much higher computing density, which reduces the expenses for infrastructure).

Hopefully, Sorano CPUs and motherboards for them will become available.

With Zen 6, AMD also starts with the most expensive SKUs, but perhaps the delay until launching the cheaper variants will not be so long as it was with Zen 5.

▲

bhouston 15 hours ago | parent [-]

The consumer desktop market is screwed with really high hard disk and memory prices and GPU prices. Premium desktops are a dead market.

	▲	adrian_b 14 hours ago \| parent [-]
		While this is true, premium desktops remain the cheapest computers accessible for consumers (in performance per dollar). The rise of the prices for mini-PCs and laptops is much worse, because there also the CPUs are now much more expensive, not only the DRAM and the SSDs. The mobile AMD Zen 4 CPUs had excellent prices, similar to the prices of CPUs during the last decade, but the mobile AMD Zen 5, both Strix Point and Strix Halo, had a great jump in price, going to levels never encountered before. A part of the reason is provided by the big useless NPU that increased a lot the size of the die, to enable the claims of being suitable for "AI PCs". On the other side, the mini-PCs and laptops of last year with Intel Arrow Lake H still had decent prices not much greater than the prices for such products during the last decade. But this year Panther Lake has brought a huge increase in price, between 2 times and 3 times. The great increases in price for the CPUs used in mini-PCs and laptops is less obvious for non-technical people, because there still exist cheap computers and they might even have similar part numbers with computers of the same price from one year ago (in prices for a barebone system, without memories), but that is a deliberate confusion, because comparing the technical specifications shows that at the same price you get much less, e.g. a much smaller GPU, fewer CPU cores, lower clock frequency, etc. and when you want the same as in a computer from one year ago, you must pay double, or even more. As a concrete example, I have an ASUS NUC 15 Pro with an Arrow Lake H from last year, which was around $500 as barebone, and today even a barebone NUC 16 Pro with Panther Lake of $1000 has worse specs. The same is true for AMD, where if you want to upgrade a mini-PC with Zen 4 that was around $500 you also need to pay around $1000 for a Strix Point system that is a true upgrade, i.e. which does not have worse specs. After you add 3 or 4 times more expensive DRAM modules and SSDs to a barebone twice more expensive, the result is even worse. For desktops it remains much better, because the CPUs have only become cheaper than at launch and the motherboards have stayed the same, so the price for a barebone system remains about the same as 5 years ago. Only when you add DRAM, SSDs and a GPU, you are hit by the price increases. Overall that is not as bad as for mini-PCs or laptops, yet.

▲

addaon 18 hours ago | parent | prev | next [-]

Wonder what the price will be like on the 64-core high-frequency version... if it's comparable to existing Epyc pricing, this might be the obvious answer for a CFD rig in 2027. Finally something with memory bandwidth way greater than a Mac Studio…

▲

zorgmonkey 17 hours ago | parent [-]

If you're after high clocks I think threadripper has usually been the cheaper option than epyc

▲

addaon 17 hours ago | parent [-]

CFD performance is dominated by memory bandwidth, not clock speed.

	▲	rbanffy 6 hours ago \| parent [-]
		As is inference, BTW. Just remember you’ll need all memory sticks in place to get to that bandwidth. That’s expensive.

▲

booi 19 hours ago | parent | prev | next [-]

dear lord that's a monster chip. 128 PCI-e v6 lanes. Looks like it's an TSMC N2 process and each CCD is 155mm2. That's a pretty large die, i wonder what kind of yields TSMC is getting on these.

▲

throwaway81523 18 hours ago | parent [-]

Seems to be a multi die module.

	▲	wmf 17 hours ago \| parent [-]
		Yeah, each die is 155 mm2 and the whole thing is around 2,000 mm2.

▲

petra 17 hours ago | parent | prev | next [-]

The other side of llm's is the need for rich feedback, often by simulations of various things. And memory latency is a big bottleneck there.

So this chip nicely complements the llm's.

▲

arjie 17 hours ago | parent [-]

Someone on HN pointed out the Intel Xeon Max 9480. 64 GiB of HBM2 attached directly so the thing can boot without any RAM sticks if necessary. Cool processor but it couldn't use the memory bandwidth very well apparently.

	▲	adrian_b 15 hours ago \| parent [-]
		At least in the initial benchmarks Xeon Max could not reach anything close to the theoretical memory bandwidth. I have no idea whether that was corrected later with some microcode update or a new mask stepping. Regardless, now Xeon Max is obsolete. The Chinese have revived this idea in the custom CPU with an Armv9.3-A ISA that is used in their latest supercomputer, which also uses HBM memory, but which is in a complete other league, by having a 5 times higher memory bandwidth than the theoretical memory bandwidth of Xeon Max, and the ratio is much greater when considering the actual memory bandwidth of Xeon Max, and not the theoretical value. That memory bandwidth of 8 TB/s per socket has enabled the Chinese supercomputer to exceed the throughput of the US supercomputers, which use GPUs, but even these GPUs have a lower memory bandwidth than the Chinese CPUs. However, AMD has announced a pair of new datacenter GPUs, one for HPC (Instinct MI430X) and one for AI (Instinct MI455X), which increase many times the memory bandwidth, both over their older models and over the Chinese CPUs. MI455X has a monstrous HBM interface, 24,576 bits wide, for a 23.3 TB/s memory bandwidth. MI455X is what AMD will sell to Anthropic (in the form of complete Helios racks, with EPYC servers full of GPUs) in the deal discussed in a couple of other HN threads (Reuters title: "AMD to sell Anthropic tens of billions in AI servers, invest up to $5 billion in startup"), i.e. Anthropic will not pay all in cash, but $5B will paid by giving a share to AMD.

▲

nullc 9 hours ago | parent | prev [-]

I was sad turin-x never shipped... I wonder if it got displaced by the Microsoft-only HBM integrated part.

I've been pretty happy with 9684X but was looking forward to upgraded cores. 16 channel memory though will be a difficult fit with dual socket systems... would be nice to see it paired up with HBM instead.