Intel Gaudi 3 vs Google TPU v7 Ironwood

Round by round

1. Round 1How much memory it hasGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   128 GB HBM2E

   hbm capacity

   • 128 GB HBM2E, against 192 GB HBM

   Source: cdrdv2-public.intel.com ↗

   Google TPU v7 Ironwood8/10

   192 GB HBM

   hbm capacity per chip

   Source: docs.cloud.google.com ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

2. Round 2How fast its memory isGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   3.7 TB/s

   hbm bandwidth

   • 3.7 TB/s, against 7.37 TB/s

   Source: cdrdv2-public.intel.com ↗

   Google TPU v7 Ironwood8/10

   7.37 TB/s

   hbm bandwidth per chip

   Source: blog.google ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

3. Round 3How fast it runs FP8 AI mathGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   1,678 TFLOPS

   fp8 mme (matrix) tflops

   • 1,678 TFLOPS, against 4,614 TFLOPs

   Source: cdrdv2-public.intel.com ↗

   Google TPU v7 Ironwood8/10

   4,614 TFLOPs

   peak fp8 compute per chip

   Source: blog.google ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

4. Round 4How fast it runs FP16 AI mathGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   459 TFLOPS

   FP16 (signed) matrix (MME) peak, Gaudi 3 technical paper Table 5; 1,678 is the BF16/FP8 figure

   • 1,678 TFLOPS, against 2,307 TFLOPs

   Source: cdrdv2-public.intel.com ↗

   Google TPU v7 Ironwood8/10

   2,307 TFLOPs

   BF16 peak per chip (Google publishes no FP16 figure)

   • peak bf16 compute per chip

   Source: docs.cloud.google.com ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

5. Round 5Which memory generation it usesGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   HBM2E

   memory type

   • HBM2E, against HBM3E

   Source: cdrdv2-public.intel.com ↗

   Google TPU v7 Ironwood8/10

   HBM3E

   hbm type

   Source: cloud.google.com ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

6. Round 6How new it isGoogle TPU v7 Ironwood wins
   Intel Gaudi 35/10

   September 24, 2024

   launch press release date

   • September 24, 2024, against Apr 09, 2025

   Source: intel.com ↗

   Google TPU v7 Ironwood8/10

   Apr 09, 2025

   announcement date

   Source: blog.google ↗

   Google TPU v7 Ironwood wins this round on what the company itself publishes.

Side by side

Intel Gaudi 3

Check price

Top ratedGoogle TPU v7 Ironwood

Check price

Overall score

6.2

7.4

Best

Price

See price

See price

How fast its memory is

6.2

7.3

How much memory it has

6.2

7.3

How much power it draws

5.8

—

How fast it runs FP16 AI math

7.4

7.3

How new it is

7.3

7.8

Head-to-head findings

How fast its memory is

3.7 TB/s

7.37 TB/s

How much memory it has

128 GB HBM2E

192 GB HBM

How much power it draws

900W

—

How fast it runs FP16 AI math

459 TFLOPS

2,307 TFLOPs

How new it is

September 24, 2024

Apr 09, 2025

Which memory generation it uses

HBM2E

HBM3E

How fast it runs FP8 AI math

1,678 TFLOPS

4,614 TFLOPs

How it handles scientific FP64 math

—

—

Record

Head-to-head record

2–3

3–1

Head-to-heads won

40 %

75 %

Rounds won

38 %

75 %

Pros

• bf16 mme (matrix) tflops

• peak bf16 compute per chip

Cons

• 128 GB HBM2E, against 192 GB HBM
• 3.7 TB/s, against 7.37 TB/s
• 1,678 TFLOPS, against 4,614 TFLOPs
• 1,678 TFLOPS, against 2,307 TFLOPs

• 192 GB HBM, against 288 GB HBM3E
• 7.37 TB/s, against 8 TB/s
• 4,614 TFLOPs, against 5 PFLOPs
• 2,307 TFLOPs, against 2.5 PFLOPs

More AI accelerators match-ups